design

This skill should be used when a task's criteria are grounded and it is time to decide how to build them, for example "design this task", "write work orders", "architect this feature", "plan the build

SKILL.md

GitHub

Design

Design decides how, and writes work orders. The work orders and the order they run in are the architecture: there is no separate architecture document, no overview file, and no file per component. A work order is a small unit of build that four roles can each act on without asking what was meant: one writes the tests, one writes the code, one criticises the result, one applies fixes.

Design does not write code and does not run tests. It does not decide whether a criterion is right; that was scope's conversation. It reads what research found, decides fit, and writes the orders that build it.

Design's own records go through design-actions.sh, named in this skill's own grant, so they run without asking. One write below goes elsewhere. A lookup is recorded through the research skill's research-actions.sh, because that script is the single producer of a research finding, and it is named in the grant too. So is the project skill's recipe-point lookup. Any other Bash command still asks for approval. Dispatching an agent needs no approval either; it is also named in this skill's own grant.

design-actions.sh and research-actions.sh take a text value from a file. Write the text with the Write tool in the session's scratchpad folder. With no scratchpad, write it in the task folder, and delete it after the call. Then pass @/<absolute path> as the value. Use this for a text that names a command, such as git. Claude Code refuses a command whose quoted text names git when the session works in a worktree.

The dispatch message is the role, the run mode and the paths. Name the role on the Agent call, and set the model where the step says. The message itself is one line per item: the run mode, interactive or autonomous, then each path the step hands over. One word a step names, a lens or a stage, is a line too. Nothing else goes in. The role's rules and its return shape live in its agent definition, which reaches it on every dispatch. So the message restates neither, and two runs of one step hand the role the same words. Each dispatch below names this shape and lists its own paths.

Determine the run mode

Look for a stated run mode on the task active in this conversation. Found, and it says autonomous or light: act autonomously through this whole invocation. Anything else, including no active task: act interactively, the safe default. Decide this once, at the start. A mode that names stages in brackets, such as autonomous (implement), covers this stage only when the list names design; otherwise this stage is interactive.

Find the task

Resolve the active project's own folder first, from the project-folder: line of the ## AIDA block. Then resolve the task, <taskId> when given or whichever task is already active in this conversation. Neither known: say so in one line and name the task skill. Stop; there is nowhere to write.

An invocation line whose first word is close is the person's yes on the design. The task is the word after it, or the active one. Go straight to "Close the design" below. When no design-critique-*.md exists yet, under records/ or under design/, run "Critique the design" first. An earlier close moved the files into design/, so look in both. A name that begins unfinished- is a critique that stopped, and it counts as no critique here. A light task runs no critique, so go straight to the close.

Once found, the task's own folder is <projectPath>/tasks/<task-id>. Every call below takes that folder.

Read what is already there

Run:

"${CLAUDE_PLUGIN_ROOT}"/skills/design/scripts/design-actions.sh read "<task_folder>"

This prints summary lines. contract: says present or absent and contract-file: names the file. criteria: and non-goals: list ids only. work-orders: is a count, and one work-order: line names each file already on disk. Read the criteria's text from the contract file when drafting.

read is the one call here that runs from anywhere. Every other call refuses at exit 79 when the task builds in its worktree and this window is elsewhere. The refusal names the tree and the route that works from here, either the EnterWorktree tool or the call started with cd <worktree> &&. Take the route it names, the way /aida:next does, and run the call again.

contract: absent: say so in one line and name the scope skill. Stop; a work order with nothing to serve is nothing this stage can check.

work-orders: above zero: this is a resumed or repeated run. Read each named file before drafting anything new, rather than starting over.

A reopen that changes only owned files, done-when rows or accounted findings on an existing order may skip the research and guide reading below. Those calls are add-owned-file, remove-owned-file, add-done-when, remove-done-when and account. The reading informs an order's shape, not its file list. The route is the change, then check, close with the verdict the last design-closed.json records, and distill. A reopen that creates or merges an order, removes one with remove --id --reason, or changes an order's interface, criteria or dependencies, reads as a first run does. Recording an old merge with remove --merged-into changes no order, so it takes the cheap route above. Once implementation has a snapshot, an order create adds enters the build at the next start, after this close. A defect that a review found is fixed on the task branch, and implementation's finish runs again.

Four changes do not stop an order that implementation already started: an added owned file, an account call, a reason added with update --append-reasoning, and a row marked with update --absence-reviewed. Once design closes again, the next start takes that order's live copy in place, even when the order closed. Any other change to a started order stops it for design drift at the next start. Two routes lead back. Restore the design and start clears the stop. Or take the redo in the implement skill's references/finish.md, which rebuilds the order against the new design. The cheap reopen stays cheap for an order that has not started.

Then start design:

"${CLAUDE_PLUGIN_ROOT}"/skills/design/scripts/design-actions.sh start "<task_folder>"

Research is optional. When it has not closed, start goes on and writes a warning to the task record; say so to the person. A NOTE: line names a stated mechanism edited after research grounded it; read that claim as ungrounded.

Then account for each research finding. Each <task_folder>/research/<search>.md holds the findings of one search, numbered from 1. A finding is named <search>#<n>. Each finding is used by an order or set aside with a reason. A finding that names a collaborator, a service, a convention or a constraint is used by the order it affects. A finding that says nothing was found is accounted for too. A need no finding covers is looked up first, as "Look up what you decided to use and research did not" says. Once the orders a finding affects exist, record the finding on every one of them:

"${CLAUDE_PLUGIN_ROOT}"/skills/design/scripts/design-actions.sh account "<task_folder>" \
  --id <woId> --finding <search>#<n> [--set-aside "<why this order does not use it>"]

Without --set-aside, the order uses the finding, and the build brief carries its text. A second call for the same finding on the same order replaces the entry. --remove in place of --set-aside deletes it. check prints an unaccounted: line for each of three cases. A finding is in no order's list. Research changed a finding's text after it was recorded. An entry names a finding research no longer holds. A person's close writes each one to the task record as a warning. An unattended close refuses, exit 7. Both modes account the same way.

Then read <task_folder>/records/playbooks.md. It holds the plays research loaded, the rules this project and this person want followed. Name a play that decides an order's shape by its id in that order's reasoning.

Read the guides and recipes research found

Research named these without opening them, so design is the first read. Research recorded an address for each; open it through the navigator the same way, and read a project's own guide file directly. Open a tooling recipe with the navigator's tooling --name mode, by the name research recorded. One agentic recipe covering the work means the decision is already made: follow it. Its ## Verifier becomes the proof of each order it covers, as "Carry the proof from its source" says. Two: read both, pick the one that fits, say why, and build from that one alone. None: architect from the findings and from this project's own conventions; this is where design quality shows.

Record each body as it is read, once the navigator has given its path on disk:

"${CLAUDE_PLUGIN_ROOT}"/skills/design/scripts/design-actions.sh read-guide "<task_folder>" \
  --path "<the body's path on disk>" [--name "<the name research gave it>"]

It writes one entry per path, with the body's sha256 and the date, to <task_folder>/design-guides-read.json. Give --name when research named the guide, so the entry joins its finding. A second read of the same path replaces the entry and keeps the name when none is passed. A path naming no file is refused: record only what was read. Nothing else records which bodies design opened, and a later session cannot ask this one.

read and start print guidesRead:, the count. On a resumed run, when work-orders: was above zero, they also print one guide: line per entry: changed, unchanged or missing. Read the changed ones, and the ones research named that no entry records. A missing body is gone from its recorded path; resolve it through the navigator again and read it. Skip the unchanged ones; a body the first run read, and that has not moved since, is not read again. No entry at all means the first run recorded nothing, and every body is read.

Read the process recipe for this project's framework

Run this, with one --framework flag per framework:

"${CLAUDE_PLUGIN_ROOT}"/skills/project/scripts/project-actions.sh recipe-point "<projectPath>" --point design --framework <framework>

It asks the project's own sources, then the navigator, and prints one POINT: line per framework. answer= is a path to the body on disk, or a word. A folder that holds nothing is not an answer, so a folder miss never skips the navigator. When searched= names folders, record them beside the answer. A later reader then tells a folder miss from a project with no folder. The no-recipe path below is reached only when answer= reads no-recipe. Read the body from the path. Never fetch a catalog address yourself and never read a cached copy behind the navigator's back. Verdict words and a missing heading follow ${CLAUDE_PLUGIN_ROOT}/skills/tool/references/reading-a-recipe.md.

Three answers, not one. No recipe for this framework, a listing that could not be reached, and a failed network are three different things, and only the first says anything about the framework. Record which one happened, in those words.

Judge the fit once, after the body is read and before the first order. Does this method, by its description, Goal and Preconditions, describe the work the criteria and non-goals name? Keep the verdict, the path and one sentence of reason for close. On false, interactive: say so with the reason, then ask one question with four answers. Continue with the recipe, continue without it (the fallback below), write the recipe first, or stop. Autonomous: continue with the recipe and record false.

This is what makes the work orders right, and no check below can replace it. Read it for what AIDA cannot know on its own:

  • What kinds of thing a work order can be about here. In Drupal a module, a service, a plugin, a theme, a component, a configuration entity. This is what an order is sized around.
  • What is built with configuration rather than code. A view or a content type is a work order with no code in it. It states no test. It is created with --proof gate. Its proof is its own verify lines, then the implement recipe's ## Configuration gate lines. The recipe's sizing rule decides what it owns.
  • What is a document rather than code or configuration: a dependency review, a report, a note. Such an order owns files under the project folder, in a folder the project commits, the task folder's deliverables/ by default. A report may land beside earlier reports elsewhere in the project folder. Never records/, which the project ignores. It states no test. Its proof is its done-when rows, so it is created with --proof record. add-owned-file sets that value itself once every owned file lies under the project folder, on an order created with no --proof. Such a file must live in a folder the project commits: the design check refuses a record order that owns an ignored path.
  • What is proved by what a page shows rather than by a test. A layout, a rendered block, a page as it renders. Such an order is created with --proof observe. It names at least one --surface and declares no test. Each done-when row is the sentence a model judges. After the build, the orchestrator opens the page the order builds at one size in a browser. It judges the row against what renders, with a screenshot as the evidence. Write each row as one thing the page must show. The judge is a model. At review's close, the person answers each such criterion from the acceptance verifier's evidence.
  • What has to exist beside a class for it to work: a services entry, a route, a permission, a schema. Name these in the order, or whoever builds it invents them.
  • What one unit exposes to another, which is what the interface field holds.
  • Where business logic belongs, and what a thin layer may contain. An order that puts logic in the thin layer is drafted wrong.
  • The entry point every feature has that is not a screen. Name it in the order that builds the feature, so the feature is reachable without its UI.
  • What order the framework forces, where it forces one.

The proof follows what the order produces. Ask what the order leaves behind when it is done, and pick the kind from that answer:

  • Code that a test can pin: tests.
  • Tools run that change state, such as a dependency update, a database update or a configuration export: gate. Nothing new exists for a test to pin, and a test written for it only restates a file.
  • A document or an analysis, such as a report or a review: record.
  • A document, a recipe, a skill or agent file, or hand-written configuration in the code repository: confirm. It holds no logic for a test to prove. The reviewer reads it, and the person confirms each done-when row at review. add-owned-file sets this value itself when every owned file is a document or configuration. A gate order is never this case. Own a skill by its files: a folder reads as code, and add-owned-file prints proof-note: when it does. When text has a behaviour of its own, put it in its own order, so that its lines run first.
  • Something only a person or a browser can see: observe.
  • Code, on a task whose contract says it has no automated tests: confirm. Nothing runs a test. The implementer builds it, one reviewer reads the diff, and the person confirms each done-when row at review. add-owned-file sets this value itself on such a task, on an order created with no --proof whose first file is code. The person may pick another kind with update --proof.
  • Changes to tests alone, from an approved test cleanup. A merge, a shrink or a rewrite is a tests order that owns only test files, and it reuses the code its tests call. Whole files to delete are one tests-only order: confirm, owning only files of the cleanup run the goal names. Name each file in a done-when row, and make the order depend on every tests order.

A machine-verified criterion does not mean tests. Tests prove one, and so do the gate lines and run lines. A text-only order proves none, so the check refuses it. A look or a record order's reading proves none either, so the check refuses one an observe or record order owns, unless a run line proves it. Code with logic gets tests even when it owns no machine-verified criterion. create and update print impliedProof: beside the proof the order declares. It says record when every owned file lies under the project folder. It says confirm on a task with no automated tests, and when every owned file is a document or configuration. Otherwise it says any kind for a machine-verified criterion. What the order produces decides. When the product is truly unclear, tests stays the default, on a task that has tests. A wrongly tested configuration order wastes one build, and a wrongly untested code order ships unproven.

No recipe covers this framework: say so, and write written without framework input into the reasoning of every order in this pass. Do not invent a kind of unit and do not guess at a framework convention. Interactive, when the navigator found no recipe: ask whether to write the recipe first, before drafting anything. Reached from a false fit, that answer was already put with the fit question, so do not ask it again. The shape is in ${CLAUDE_PLUGIN_ROOT}/templates/process-recipe-design.md, which carries the sections the catalog requires and the five things design needs from a framework. Skip the ask for a framework that task.json's recipesDeclined names; research recorded that answer.

The stated approach

A task may carry a stated approach in task.json's mechanismHints, recorded by scope. Read it as a claim, never as a specification, whatever its status. Weigh it against what research found the ordinary way; a required status means it must be followed once judged sound, not that it is exempt from judgment. start compared each claim against the hash research recorded at its close. A claim start printed a NOTE: for was edited after research looked at it, so its grounding no longer covers it. Judge it as ungrounded, and send it back to research when it matters.

The reuse decision

For every prior-art candidate research handed over, ranked by closeness, give it an answer: reuse, extend, supersede, or decline, in that order. An unanswered candidate is a proposal nobody acted on. A candidate you find yourself gets an answer too, not only research's list: an exported configuration entity of the same kind as the unit, say. Decline is the answer for a candidate you weighed and set aside; recorded, it is told apart later from one nobody weighed.

Decide fit by the candidate's own distance (same name, same directory, same layer) and by version 5's own cost model: build cost is paid once, carry, agent and risk cost are paid forever. A candidate is not always code; an existing view or content type is a candidate too, and extending it may produce no code at all.

Record the disposition through the script that will own the work order it lands on, never only in your own reasoning. Once that work order exists, give the script the candidate, its closeness as research stated it, the cost dimensions compared, the verdict, and why:

"${CLAUDE_PLUGIN_ROOT}"/skills/design/scripts/design-actions.sh --run-mode <interactive|autonomous> \
  dispose "<task_folder>" --id <woId> --candidate "<what research found>" \
  --distance <same-name|same-directory|same-layer> --cost <build|carry|agent|risk[,...]> \
  --verdict <reuse|extend|supersede|decline> --why "<why this verdict>" \
  [--path "<where it lives>" --interface "<what it exposes>"]

Give --path and --interface whenever the order's build or tests will call the candidate. --path is where the reused thing lives, relative to the code repository: a file, or a configuration path when the candidate is not code. When the return names several paths, give the file the order calls as --path. The other paths stay in the interface text. --interface is the text internal-searcher returns, word for word. Do not read the candidate's source here, and do not reword the return. The test author may not read the source either, and the tests brief carries this text in place of it. A dispose that omits both records no reuse.

Before that dispose, dispatch internal-searcher once per candidate the order's build or tests will call, with the message this file names. Its lines are the role, the run mode and the word interface. Then come the candidate as research named it, the order file's path, the code path, and the project folder. The code path is the task's worktree, from the worktree: line. A worktree: none means the project's code path. Light: dispatch once for all orders. Repeat the candidate line and the order line, one pair per candidate, before the code path. Pass each returned block as that candidate's --interface.

A decline takes no --cost and no --path: nothing is compared, and nothing is reused. Its --why names what was weighed. It stands in both modes, because a decline with a reason is a recorded decision, not a downgrade.

The script applies a fixed table and appends the outcome to the order's reasoning, one paragraph per candidate. Re-disposing a candidate adds a paragraph; the last one stands. It prints disposition:, which is what stands. A supersede citing only build cost or no cost dimension, or a candidate sharing only a layer, comes back as extend. Autonomous, a supersede comes back as extend. Each downgrade is in the reasoning and is a warning in the task record. A reuse or extend citing no cost dimension stands, with the thin reasoning recorded, because there is nothing to downgrade it to.

Interactive, ask the person once before a supersede. Name the candidate, how close it is, and that a supersede widens this task and owes a migration. A yes is --verdict supersede. A no is the verdict the person chose.

A rejection that lives only in the conversation is not a rejection anyone can check later.

Autonomous: once every disposition of an order is recorded, dispatch disposition-confirmer once for that order, with the message this file names. One confirmer judges every candidate of the order. Its lines are the role and the run mode. Then the order file's path, from the DISPOSED: line. Then the verdict file's path, from the confirmFile: line. Then the code path, which is the task's worktree, from the worktree: line that dispose prints beside it. A worktree: none means the project's code path. A dispatch that names no role runs as the general agent with write tools and this session's model. This one must write nothing but its verdict file to mean anything.

Never this conversation's own account: being denied that is the entire reason the role exists, and handing it over turns the check into the decision reading itself. Record what it found with update --append-reasoning, which adds a paragraph after the text dispose wrote.

close reads each disposition made unattended against its order's verdict file. A missing value, or a value the confirmer answered differently, is a warning in the task record, one line per disposition. The builder and the architecture reviewer both read the reused interface, and both stop on a misfit. To take the confirmer's value, dispose that candidate again with it.

Interactive runs do not dispatch it. A person read the reasoning, and the role has nothing to add.

Look up what you decided to use and research did not

Design names things research had no reason to search for: a particular module, a framework API, a pattern. Those were not decisions yet when research ran, so no search covered them. An order's interface can also need a collaborator, a service, an input source or a convention that no finding covers. How a class in this framework gets the current time is one such question. A question a quick lookup can answer is looked up, never guessed, before the close.

First read what research already searched for. Every research/<search>.json carries searchedFor, the words that search used. A name inside those words was searched, and the answer is already in that file. Do not pay for it twice.

For a name that was not searched, look in this order, and stop at the first answer:

  1. This project's own code. Dispatch internal-searcher with the words to search.
  2. The catalog. Dispatch catalog-identifier with the words to search. It identifies only. It returns names and never resolves a body, so one name costs one lookup. Read a body only when a match is worth reading.
  3. When no guide or recipe covers it, what reputable sources recommend. Dispatch outward-searcher, one subject.

Each dispatch carries the message lines that ${CLAUDE_PLUGIN_ROOT}/skills/research/SKILL.md names for that role, under "Dispatch one agent per search".

Record the answer through research's own record action. The research store keeps one producer that way, and the finding is checked the same way as every other:

"${CLAUDE_PLUGIN_ROOT}"/skills/research/scripts/research-actions.sh record "<task_folder>" \
  --search design-lookup-<name> --searched-for "<the words this lookup searched for>" \
  --text "<what was found, or that nothing was>" --source "<where it came from>" \
  --criteria-served <id[,id...]>

One lookup is one search, with a name of its own. <name> is the thing you looked up, in lowercase letters, digits and single hyphens: design-lookup-responsive-image. Never put two lookups in one search. A search records one set of words, and a second set is refused.

Always give the criteria. You are looking this up for a work order, and that order serves criteria, so give their ids. A finding attached to no criterion is reported as work nobody asked for, and it turns the research check red.

Record a nothing too. A name looked up with no guide behind it is a fact the implementation stage needs, and it stops the same lookup running again there.

Do this again while drafting, whenever an order names something new. A guide found late still costs less than a guide found after the code is written.

Design never silently changes the scope

Drafting work orders turns up criteria nobody wrote, and criteria that cannot be built as stated. Neither is design's to fix alone.

Interactive: say what was found and ask the person before proceeding, then use the scope skill's own update path if the contract needs to change.

Autonomous: record what was found in the work order's own reasoning field and continue. A recorded note reaches a person; a silent change or a silent skip does not.

Size a work order

An order holds a handful of behaviours, not a feature. The design check warns on an order with more than eight done-when rows, or more than three for each criterion it owns. Three to seven build steps, ten at the most. Split when the steps mix independent concerns, or mix phases. Merge two orders when each has fewer than three steps, they address the same component, and they cannot run in parallel anyway. The script folds one into the other:

"${CLAUDE_PLUGIN_ROOT}"/skills/design/scripts/design-actions.sh merge "<task_folder>" \
  --into <woId> --from <woId> --reason "<why the two are one order>"

Every list on the folded order joins the survivor's, without duplicates. The folded order's interface and reasoning are appended to the survivor's, each under a line From <woId>:. The title and the diff budget stay the survivor's; the output says carried: and dropped: so nothing goes unseen. Retitle with update when the survivor's title no longer covers what it owns. The folded order's file is removed, and every dependsOn that named it now names the survivor. The folded id, the survivor and the reason go into design-removed.json, the same record remove writes. The two proofs must agree; set one order's --proof first when they do not. Never remove or edit an order file by any other means. A write outside the script prints nothing, so nothing records that it happened.

An order that no longer earns its place, for example after the contract changed, leaves through the script with its reason:

"${CLAUDE_PLUGIN_ROOT}"/skills/design/scripts/design-actions.sh remove "<task_folder>" \
  --id <woId> --reason "<why the order is gone>"

The script deletes the order's files and records the id and the reason in design-removed.json. The close copies that list into design-closed.json. A removed id is never minted again. Move a dependency or a criterion with update first. The design check refuses a dependsOn that names nothing. It also names a criterion left with no serving order or owner. For an order implementation started, fold it into the order that takes its work with merge. Implementation's start stops it for design drift, and drop sets its records aside. A finding that the removed order accounted for shows in check. Account for it again.

An order folded before merge wrote that record left no entry, so the design shows a gap in its numbers. Record it with the survivor and a reason:

"${CLAUDE_PLUGIN_ROOT}"/skills/design/scripts/design-actions.sh remove "<task_folder>" \
  --id <woId> --merged-into <woId> --reason "<why the order was folded>"

The script refuses an id that is still an order, and an id the record already holds. check lists each such id under unrecordedIds:. Recording changes no order, so it stops nothing. Then run check, close with the verdict the last design-closed.json records, and distill, as for an owned-file reopen. The close copies the entry into its record.

A test that no longer belongs on an order leaves through the script too. The merge may have doubled it, or the order became a gate:

"${CLAUDE_PLUGIN_ROOT}"/skills/design/scripts/design-actions.sh remove-test "<task_folder>" \
  --id <woId> --description "<the test's description, exactly as declared>"

The design check refuses a tests order that owns a machine-verified criterion and declares no test. Add the replacement first.

A done-when row or an owned file leaves the same way. Moving a file to the order the sizing rule names leaves its rows on the old order. The checkpoint would then judge that order on a file it may not write:

"${CLAUDE_PLUGIN_ROOT}"/skills/design/scripts/design-actions.sh remove-done-when "<task_folder>" \
  --id <woId> --text "<the row's text, exactly as declared>"
"${CLAUDE_PLUGIN_ROOT}"/skills/design/scripts/design-actions.sh remove-owned-file "<task_folder>" \
  --id <woId> --path "<the path, exactly as declared>"

Each prints what it removed. The design check refuses an order with no owned file, because it gives the builder no boundary. Add the replacement first, or fold the order with merge. A record order left with no file under the project folder loses its proof, and the output says so. Set --proof again if that was wrong.

A shared decision, like one base class serving two later orders, lives in the order that builds the shared thing, in its own reasoning. The orders that depend on it point at it through dependsOn; nothing else needs to hold that reasoning.

Draft each work order

For each unit of build: agree its title, which criteria it serves, and, when it produces the observable outcome a criterion describes, which one it owns. Most orders own none. Exactly one order owns each criterion; if two orders both seem to produce the same outcome, that is a sign the work is split wrong, not a sign both should claim it.

Light: the first order, wo1, is the walking skeleton. It is a tiny version that links the input, the logic and the output end to end along the demo path. Every other order depends on it, directly or through its chain, and check refuses one that does not. A project keeps at most one script that walks the demo path. start prints the path-script: line, and its first word says the case. create: wo1 also owns that script. For a page, it sits where the harness reads its tests. For a command surface, it is the file the command runs. A person registers it with /aida:surfaces. registered: the project has the script, and no order writes or extends it. check refuses an order that owns one of its files, wo1 included. declined or off: no order writes one. A change to the script goes through /aida:surfaces with a person.

Create it:

"${CLAUDE_PLUGIN_ROOT}"/skills/design/scripts/design-actions.sh create "<task_folder>" \
  --title "<what this builds>" \
  [--criteria-served <id[,id...]>] [--criteria-owned <id[,id...]>] \
  [--non-goals <id[,id...]>] [--depends-on <id[,id...]>] \
  [--interface "<what it exposes to what depends on it>"] \
  [--reasoning "<why, if this is a shared decision>"] \
  --diff-budget "<a plain-words signal, e.g. small: one class and its test>" \
  [--proof <tests|gate|record|observe|confirm>] [--surface <id>]...

This mints the next id and writes the file, and prints the id and the fields set. It never prints the record; read the file at the printed path when a field is needed. dependsOn may name a work order not yet created in this conversation; the id space is shared and minted in order, so naming it ahead of its own create call is safe as long as it is created before design finishes.

Write --interface so that it names each exposed element in backticks. That is the service id, the class, the interface, the method, the route, the hook and the config key. The build brief lists only the backtick-quoted names, and the builder names each one in its record, which the reviewer compares. So write a choice the builder makes, such as two calls it may pick from, as plain text. An order that exposes nothing takes no --interface.

Start --diff-budget with small, medium or large. The build reads the first word. An order whose budget starts with large gives its builder a second resume at the turn limit. Number the parts of such an order in its reasoning, as (1), (2) and so on. The builder commits after each part, so a stop loses one part and not the order.

Name a --surface when the order changes a page or a screen a person sees, by its id in the surface registry. Most orders name none.

Before naming one: this order changes a page, and <projectPath>/project.json has surfaces null or a kind that is off and not declined. Skip a kind that read names under surfaces-deferred:. Before you ask, run "${CLAUDE_PLUGIN_ROOT}"/skills/surfaces/scripts/surfaces-actions.sh read. When its surface-branch: names a branch, offer nothing. Name that branch, and say the surface file reaches this task once that branch lands on the trunk. Interactive only, offer the setup once per task, naming every such kind. Say so in one line. Ask whether to set the surfaces up now, per kind, with a recommended answer per kind. Give three answers per kind: yes, not this task, or no. Say in the ask that "no" is project-wide and "not this task" is not. A yes on a kind invokes the surfaces skill through the Skill tool, naming that kind. Do this once per kind said yes to. This order then names the id the setup registered. "Not this task" on a kind runs "${CLAUDE_PLUGIN_ROOT}"/skills/task/scripts/task-actions.sh defer-surface --project "<projectPath>" <task-id> <kind>, and no later stage of this task asks again. A no on a kind runs "${CLAUDE_PLUGIN_ROOT}"/skills/surfaces/scripts/surfaces-actions.sh decline <kind>, project-wide, and that kind is never asked again. Autonomous: nothing is offered. A page scope did not see is often first named here.

Then, one call per item, add what the order still needs:

"${CLAUDE_PLUGIN_ROOT}"/skills/design/scripts/design-actions.sh add-owned-file "<task_folder>" \
  --id <woId> --path "<a file or a directory this order may write, never a glob>"

A configuration unit is sized around the operation, by the recipe's rule, and owns every file that operation rewrites. Deleting a field owns each display that lists it. An order that owns the field files alone and leaves the displays to other orders cannot import on its own. The same rule holds for code. An order that changes a class owns every file the recipe couples to that class. A constructor change, for one, rewrites the service definition. A file added here to an order the build already started is taken in place at the next start, once design has closed again on the live files. Nothing stops when nothing else on the order changed.

A file that several orders only add to may have more than one owner. A list of enabled modules and a dependency manifest are examples. Each order that adds to it passes --shared with that path. An order can then enable the module it builds, and does not wait for another order. Two orders that own one path, and do not both mark it shared, still fail the check. A file that one order rewrites belongs to that order alone. A lock file that a tool generates is such a file. Nothing checks that a shared change only adds, so the reviewer of each order reads that order's own changes to the file.

"${CLAUDE_PLUGIN_ROOT}"/skills/design/scripts/design-actions.sh add-done-when "<task_folder>" \
  --id <woId> --text "<what must be true for this order to be finished>"
"${CLAUDE_PLUGIN_ROOT}"/skills/design/scripts/design-actions.sh add-test "<task_folder>" \
  --id <woId> --description "<the one behaviour this test proves>"

Write an absence as a done-when row of its own. An absence says the change added nothing of a named kind, such as no new dependency or no static call to the container. No test can watch it fail, so the build routes it to review by its exact words. A row that joins it to a behaviour, such as "contains no static call and saves no entity", cannot be routed without a reopen.

A criterion whose verifiedBy is machine, on the tests order that owns it, needs a test here. The check below refuses an order that skips this. The other kinds declare no test. The configuration check judges a gate order's machine criterion at build time. A record order has no checkpoint: its reviewer reads the document against its done-when rows. A model judges an observe order's done-when rows against its surfaces after the build. The person confirms a confirm order's done-when rows at review. A criterion whose verifiedBy is person needs no test.

A test is what this page defines: ${CLAUDE_PLUGIN_ROOT}/skills/implement/references/what-a-test-is.md. Add its file to the order with add-owned-file. The test author writes into it, and the freeze refuses a test in a file the order does not own. Each description names the one behaviour its test proves, and that page says what a test is not.

To change a scalar or an id list on an order already created, update takes the same flags as create, replacing whichever are given:

"${CLAUDE_PLUGIN_ROOT}"/skills/design/scripts/design-actions.sh update "<task_folder>" \
  --id <woId> [--title <text>] [--criteria-served <id[,id...]>] \
  [--criteria-owned <id[,id...]>] [--non-goals <id[,id...]>] [--depends-on <id[,id...]>] \
  [--interface <text>] [--reasoning <text>] [--strike-reasoning <n>]... \
  [--append-reasoning <text>]... [--diff-budget <text>] \
  [--proof <tests|gate|record|observe|confirm>] [--surface <id>]...

--reasoning replaces the whole field, the paragraphs dispose wrote included. To keep them, pass --append-reasoning: it adds the text as a new paragraph after a blank line. The two flags are refused together. When a critique or a person changes a rule that a reasoning paragraph states, strike that paragraph. Pass --strike-reasoning <n> with the new rule in --append-reasoning. Count n from 1 in the order's .md. The old paragraph stays, marked [struck], and no brief carries it. Do not leave both rules live: the builder reads every live paragraph. To strike several paragraphs, repeat --strike-reasoning in one call. Count each n before the call. One n that names no live paragraph refuses the whole call, and nothing is written. update prints struck: with the paragraphs it struck. A repeated --append-reasoning adds each text as its own paragraph, in the order given. When --proof becomes gate, record, observe or confirm, update prints stillNamesATest: naming each of interface, reasoning and diffBudget that still names a test file.

Carry the proof from its source

The knowledge that covers an order says how to verify it. Carry that onto the order, in the form its kind takes. A run entry is one command a script runs. A check entry is one sentence a model judges. Every entry cites its source and says whether it is binding.

An agentic recipe covers the order. Copy its ## Verifier:

"${CLAUDE_PLUGIN_ROOT}"/skills/design/scripts/design-actions.sh verify "<task_folder>" \
  --id <woId> --recipe "<the recipe body's path on disk>" [--not-binding]

The script copies each entry of the verifier: block as a run entry, with its pass and its kind. It copies each numbered item of the prose as a check entry, verbatim. It reads nothing else, and it never turns a sentence into a command. A recipe may carry a verifier: block, a prose list, or both. Pass --not-binding when research said the source is not one this project accepted.

No recipe covers it, and research found how to verify it. Add one entry per finding:

"${CLAUDE_PLUGIN_ROOT}"/skills/design/scripts/design-actions.sh verify "<task_folder>" \
  --id <woId> --run "<one command>" [--pass "<exit 0 | stdout empty | stdout contains <text>>"] \
  [--kind <config-assert|live-site|self-fixture>] [--criterion <cN>] --cite "<the source research recorded>"
"${CLAUDE_PLUGIN_ROOT}"/skills/design/scripts/design-actions.sh verify "<task_folder>" \
  --id <woId> --check "<one sentence to judge>" [--kind <config-assert|live-site|self-fixture>] \
  --cite "<the source research recorded>"

Write a --run only when the source gives the command. Prefer a command whenever it does. A source that describes a result in words gives a --check. Such an entry is never binding. Give --kind live-site when the check needs a served site; the source says so or it does not. A research --run never runs until a person approves it at the close. Without that, the reviewer judges it as a check. Give --criterion when the command's output alone proves a machine criterion the order owns, such as a linter run. That criterion then needs no test. When run lines prove every machine criterion an order owns, give the order --proof confirm. The lines run in its check, and the person confirms its done-when rows at review. A text order rests no machine criterion on a line. Its line is evidence for the person. On a tests order, the freeze runs the line over the test author's files, and the row-checker judges whether the line proves the whole clause. A clause that names two tools needs a line for each. A line runs unattended when it equals, word for word, an argv in the project's check recipe check_commands. Any other research line needs the person's approval at the close.

A skill or prompt with an eval takes two orders. The first owns the case files under evals/, each by name. Its proof is confirm, and its done-when rows name the scenario and the graders. It holds no scaffold_script, because a .sh file makes the order tests. The second owns the skill, dependsOn the first, and runs the eval as its run line. So the skill's author does not write the case, and the owned-files check stops that author from changing it. A claude plugin eval line names --case, --runs 1, --ablation none, --trust-plugin, --no-publish and an --output-dir outside the tree. Its graders call no judge. What fails first for each kind of work is in ${CLAUDE_PLUGIN_ROOT}/skills/implement/references/what-a-test-is.md. --clear empties the list, and --recipe replaces it.

A run line is argv, never a shell, so --run refuses a shell character. A recipe entry with one is skipped, with a warning. It also refuses a pass outside the three forms. Read the refusal and write the entry again, or leave it out and say why in the reasoning.

What each kind does with the entries:

  • gate: the run entries run first, then the ## Configuration gate. The worse verdict stands.
  • tests, record and observe: the run entries run in the order's first deciding check, after its own answer, from the code worktree. The check is met only when both are.
  • confirm: the run entries run the same way. A failing entry stops the build. A passing one leaves the check waiting for the person.
  • Every kind: the reviewer judges each check entry. On an observe order the look also judges each live-site check as a row of its own, because only that kind shows on a page.

Serving a criterion is not completing it

The decidable half is what the check below counts: an owner exists, is exactly one, and declares a test when the criterion needs one. Whether the named order actually produces the outcome, and whether its tests actually observe it, is judgment, and it needs a person.

Interactive: for each criterion, show it beside the one order claiming to own it, and ask whether that order is really going to produce this outcome. One order to look at, not a list of every order that mentions the criterion. A no means the work is drafted wrong; revise the orders involved before moving on.

Autonomous: do not skip this and do not mark it passed. State plainly, once, at the end of this run, that per-criterion ownership was judged by shape only and not by a person, and name which criteria that leaves unconfirmed. This is not a failure; it is the honest state of an unattended run, and it is what the end-of-task review will actually test.

Run the design check

Once every criterion has a drafted owner, run:

"${CLAUDE_PLUGIN_ROOT}"/skills/design/scripts/design-actions.sh check "<task_folder>"

This reads every work order's JSON and writes its report to <task_folder>/records/design-check.json. It prints status:, the report's line count and report: with the path. When the status is not zero it adds one open: line naming what stops the design. It prints one unattended: line for each finding that stops only an unattended close, and one warning: line for each finding that never stops. The report holds:

  • a work order file with a missing, empty or malformed required field;
  • every criterion with no work order serving it, and every criterion owned by zero or by more than one work order;
  • every work order serving no criterion;
  • every work order that owns a machine-verified criterion and declares no test, unless its proof is gate, record, observe or confirm;
  • every work order whose proof is gate and that declares a test;
  • every work order whose proof is record and that declares a test, owns a file outside the project folder or under a path the project ignores, or has no done-when row;
  • every work order whose proof is observe and that declares a test, names no surface, or has no done-when row;
  • every work order whose proof is confirm and that declares a test, or has no done-when row;
  • every work order whose proof is confirm and that owns a file a test glob matches. The globs are the ones this project's earlier freezes recorded. A tests-only order is exempt;
  • every work order whose proof is confirm on a task with tests, unless it is text only, run lines prove every machine criterion it owns, or it is a tests-only order;
  • every tests-only order with an owned file no done-when row names, or a row that names no owned file, and every one that does not depend on each tests order of its task;
  • every work order that owns a test another task froze, while that task is not complete;
  • every work order whose proof is tests and whose every owned file is a document or configuration;
  • every machine-verified criterion owned by an order whose every owned file is a document or configuration, a gate order excepted;
  • every machine-verified criterion owned by an observe or record order, unless a run line proves it;
  • every machine-verified criterion that a confirm, observe or record order rests on a research run line that never runs: not binding, not approved, and not a check command. This stops only an unattended close;
  • every work order that owns nothing and that no owning order depends on, directly or through the chain, and every dependency cycle;
  • on a light task, every work order that owns a file of the registered script that walks the demo path;
  • two work orders sharing a declared owned file that one of them does not mark --shared;
  • any criterion, non-goal, or work order id named anywhere that resolves to nothing real.

check also prints verifyNotBinding:, the orders holding an entry that is not binding. Interactive: before the close, show each such order's entries with their sources. Say that the source is not one this project accepted. The person keeps each entry, drops it with --clear, or asks for another source. A kept run entry runs only when the person also approves it, in every run mode. The approval is the person's answer, recorded with the task skill's person-answer.sh, step design close, with the lines as the words. Then pass --approve-runs to the close, and the close stamps each such entry. Without the answer, --approve-runs refuses with exit 70 and prints the command. Autonomous: the line and the rendered order carry it to a person later. A confirm, observe or record order can rest a machine-verified criterion on such a line. Then the autonomous close stops instead, because nothing would prove that criterion.

check also prints impliedProofDisagrees:, at every exit code. It names every work order whose proof is tests that owns criteria of which none is machine-verified. The design still closes with those orders open, because which of the other proofs fits is a judgment. Either the order owns a machine-verified criterion after all, which update --criteria-owned sets. Or its proof is one of the others, which update --proof sets. Read the verification clause of each criterion the order owns, and ask what would settle it.

check also prints absenceJoined:, at every exit code. It names each done-when row that holds a negation word and the word and, by order and row number. This is best effort, because a script cannot parse a clause, and it never blocks the close. Read each row it names. When a row joins an absence to a behaviour, remove it with remove-done-when. Then add each clause with its own add-done-when call. When a row does not join the two, mark it reviewed:

"${CLAUDE_PLUGIN_ROOT}"/skills/design/scripts/design-actions.sh update "<task_folder>" \
  --id <woId> --absence-reviewed <row>...

The order keeps the row's text, and check stops naming that row. A row whose text changes is named again. Repeat --absence-reviewed to mark several rows of one order in one call.

check also prints interfaceUnquoted:, at every exit code. It names each order whose interface holds no name in backticks. A script cannot tell whether prose names a code element, so the line never blocks the close. Rewrite each such interface with update --interface, and quote each element it exposes.

check also prints interfaceUnowned:, at every exit code. It names each path an order's interface holds in backticks that the order does not own, no reuse of it declares, and no order in its dependsOn owns. A ./ prefix and a :line or ::member suffix are ignored. The build brief lists every backticked name for the builder's record. A builder leaves out a path it did not touch, and the reviewer finds the two texts apart. It is best effort, and it never blocks the close. Answer each one: write the path as plain text with update --interface, or add the reuse with dispose --path --interface.

check also prints interfaceShaped:, at every exit code. It names each name in backticks in an order's interface that holds < or {. Such a name is often a shape, such as type.<id>. The builder's record can name a shape only by repeating its text, which the reviewer reads. A generic type holds < too, so the line never blocks the close. Name each real element instead, or write the shape as plain text with update --interface.

check also prints interfaceChoice:, at every exit code. It names each pair of names in backticks that an order's interface joins with or. Such a pair is often a choice the builder makes, and a builder that takes one option leaves the other out of its record. The line never blocks the close. Write a choice as plain text with update --interface. Leave a pair whose two names must each exist, such as two exit codes.

check also prints callsUndeclared:, at every exit code. It names each name() call in an order's done-when rows and tests that no interface the order declares holds. It reads the order's own interface, its reuses and its dependsOn orders. The test author may not open source, so a missing signature stops the tests. It is best effort, and a call a clause denies is named too. It never blocks the close. Answer each one: add the reuse with dispose --path --interface, or say why the call needs no signature.

Exit 0: nothing stops the design, subject to the judgment step above. An unaccounted: line names a research finding in no order's list, with its id and the first line of its text. Account for each one, as "Then account for each research finding" says. A person's close writes it to the task record as a warning, and an unattended close stops on it.

Exit 5: a finding stops the design. The open: line names each one. Read the report file when the lines are not enough, and fix the specific order each line names. The same repairs answer the unattended: and warning: lines:

  • a work order file that is broken, not valid JSON, not an object, or with a missing or malformed required field, needs another update call;
  • a criterion with no serving order needs a work order that names it in --criteria-served;
  • a criterion with no owner, or with more than one, needs its --criteria-owned reconciled so exactly one order claims it;
  • a work order serving nothing needs a real --criteria-served, or it should not exist;
  • a work order missing a required test needs an add-test call;
  • a gate order declaring a test needs a remove-test call for it, or --proof tests if it builds code after all;
  • a record order declaring a test needs a remove-test call for it. One owning a file outside the project folder needs --proof tests if it builds code after all. One owning a file under an ignored path needs remove-owned-file, then a path the project commits. One with no done-when row needs an add-done-when call;
  • an observe order declaring a test needs a remove-test call for it. One naming no surface needs update --surface <id>. One with no done-when row needs an add-done-when call;
  • a confirm order declaring a test needs a remove-test call for it, or --proof tests if the task has tests after all. One with no done-when row needs an add-done-when call. Each row is a sentence the person confirms at review, so write it as one thing they can check. A confirm order on a task with tests that owns code needs update --proof tests and its tests, unless run lines prove every machine criterion it owns. When its other files are text, a folder it owns needs remove-owned-file, and add-owned-file for each file the folder holds;
  • a tests order that owns only documents or configuration needs update --proof confirm, or record when its files lie under the project folder;
  • a machine-verified criterion on an order that owns only documents or configuration is a person's. Change it to person with scope's update --verified-by person;
  • a machine-verified criterion on an observe or record order is a person's too. Change it the same way, add a run line that proves it, or give it to an order that builds code;
  • a confirm order that owns a test file needs remove-owned-file for it. Nothing runs that file, and the freeze never sees it. When the task does have tests, a person answers yes in scope instead, and the order takes update --proof tests and its tests;
  • an order that owns nothing is reached only when an owning order depends on it. Add it to that owner's --depends-on. The edge points from the owner to the order it needs, never the other way. An order no owner needs is dead work, unless it owns a criterion of its own. The entry point that is not a screen belongs in the order that builds the feature, as above, not in an order of its own; an order for it alone needs a criterion that names it;
  • a cycle needs one of the dependsOn edges in it removed;
  • a light order that owns the registered demo-path script needs remove-owned-file for it. A change to that script goes through /aida:surfaces with a person;
  • a criterion that rests on a research run line nobody approved needs the person's yes at an attended close, with --approve-runs, or another proof;
  • overlapping owned files need one order's ownedFiles narrowed so the paths do not repeat, or --shared on both orders when each only adds to the file;
  • a wildcard in an owned file needs the directory named instead, or each file added, because implementation derives what the test author may not read from this list and reads it as paths;
  • an unknown id needs correcting to one that actually exists.

Then check again. Exit 3: the script could not run the check at all. Read its stderr and fix the named problem, then check again.

Design is done when this check finds nothing that stops it. Implementation refuses to start on an open: finding, so leaving one only moves the stop to a later and more expensive place. A warning: line never stops the close. Show each one to the person. The close writes it to the task record, and review lists it. An unattended: line stops an unattended close, because nobody reads it before the build.

Critique the design

check printed critique: skipped, light run: dispatch no critic, and go to the close. The check logged the skip.

Once the check comes back clean, and before closing, have three readers who were not in this conversation read the orders. The check counted ids; it read no sentence. Dispatch design-critic three times, in parallel, with the message this file names. Its lines are the role, the run mode, the task folder, and one lens of contract, reuse and buildability. Add the design recipe's path when one was read. Add the ids check printed on closedOrders: when that line is not none. A dispatch that names no role runs as the general agent with write tools. Never give it a summary of this conversation: being denied that account is why the role exists.

Each critic writes <task_folder>/records/design-critique-<lens>.md, a findings table and a findings: N last line. Wait for all three files. A file that never arrives, or arrives without its findings: line, means that lens was not read. Dispatch it again, once. If it fails twice, say so and go on: the close leaves that file out and names it. Then run check before you read the files, also when an earlier session wrote them. check removes each row that names only closed orders, and prints critiqueDropped:. Read the three files, never the dispatch replies.

The close moves each finished file into <task_folder>/design/ and commits it with the record. A critic goes on writing into records/, which the project ignores: a critique is a working file until the close decides it is evidence. The close moves an unfinished file there too, under the name unfinished-design-critique-<lens>.md, and leaves it out of the count. It is the only copy of what that critic wrote before it stopped.

A person reads the findings, because a critic that can block trains the builder to write for the critic. The critic decides nothing about closing, and neither does the count.

Interactive: show the findings grouped by work order, each with its severity, what it found, why it matters and what would fix it. The person decides what to change. Nobody answers each finding one by one: a finding nobody acted on stays in its file, which the close commits. A change is one of the update, add-owned-file, add-done-when or add-test calls above, then check again. Closed orders are not critiqued. A finding on contract is a scope question: ask, and use the scope skill's own update path when the contract has to change.

Answer each change with the lines the call printed, and end the turn. Do not show the orders again, and do not ask whether the design is ready to close. The close below is an action the person takes, never a question this skill asks. Scope learned this from a run that asked for a yes on the whole contract after each of five corrections.

Autonomous: ask nothing and change nothing. The findings stay in the three files, and the close below commits them and records their paths and the count, so a person sees them later. Say once, at the end of this run, that the critique's findings were recorded and not judged.

Close the design

The close is the approval, and there is no second approve action. An interactive close records that a person was present, and that record is the yes. The person closes by running /aida:design close <task-id>, or by saying in their own words that the design is right. Map those words to the call below. Autonomous, nobody says so: run it once the check is clean. Run:

"${CLAUDE_PLUGIN_ROOT}"/skills/design/scripts/design-actions.sh --run-mode <interactive|autonomous> \
  close "<task_folder>" --recipe-fit <true|false|unsure> --recipe-path <path> --recipe-reason "<one sentence>" \
  [--critique-outcome "<one line>"]... [--approve-runs]

Pass the fit verdict judged above. Pass --no-recipe instead only when no recipe body was read. close refuses with neither, and a later close restates the verdict rather than carrying it over. Unattended, close stops on an unattended: line, exit 5, and on an unaccounted finding, exit 7. A person's close writes those lines, each warning: line and each disposition no confirmer agreed with to the task record, and closes.

On a first close, interactive, when critique files exist: ask the person for one line. It names how many findings changed an order and how many were left with a reason. Pass it as --critique-outcome. The record holds it as critique.outcome beside the files and the count, and close prints it as critiqueOutcome:. Autonomous: pass one line that says the findings were recorded and not judged. A count alone said nothing about what changed. A re-close after a reopen, with no new critique file, needs no flag in either mode. close keeps the last record's answer and prints critiqueCarriedFrom:, naming the close that gave it. When an order was added, removed or merged since that close, the answer is still carried. The close writes a warning to the task record that names both sets of orders.

This runs the design check again. It writes design-closed.json only when that check finds nothing that stops the design. The record is committed when the stage closes: close commits the task folder, and the work order edits above commit nothing. Closing records what design closed on: a hash over the contract and every work order, the run mode, and who was present. Pass the run mode you settled at the start. An interactive close records person, an autonomous one records nobody, and implementation reads which. It also records the critique files it moved into design/, their finding count and the outcome line. So a person sees later what was read and answered before closing, and can open the files the record names.

A design left open at exit 5, with a reason recorded in an order's own reasoning, is not closed. Resolve the open item first, or record why it cannot close yet, and tell the person before you stop.

Implementation reads this record before it freezes anything. It refuses to start on a contract or a work order that does not match the recorded hash. This is what catches a work order edited after design closed but before implementation started.

Did a work order change after closing? Close again. A second close is allowed and expected. It replaces the old hash with the new one. Closing again is the supported way to change a design that already closed.

Once design-closed.json is written, dispatch distiller once, with the message this file names. Its lines are the role, the run mode, the task folder, the stage design, and the paths of design/*.json and design-closed.json. Never a summary of this conversation. It writes records/design-distill.json. Light: scope and research dispatched no distiller, so this one dispatch names all three stages. Write scope and the path of alignment.json, then research and its paths, then design and its paths. distill then prints a stage: line before each stage's lines, and a stale on any stage means this same dispatch again. Then run:

"${CLAUDE_PLUGIN_ROOT}"/skills/design/scripts/design-actions.sh distill "<task_folder>"

It prints standsAlone: and one gap: line per gap, and exits 0 on either value. Show each gap: line; acting on one is an update and a second close. standsAlone: stale means the orders or the close changed after the distiller read them: dispatch it again, then run the same call again. standsAlone: none means the sidecar was missing or unusable, and the task record holds a warning. The scope skill states the one retry rule under "The distill check".

Then hand off to implement, as ${CLAUDE_PLUGIN_ROOT}/skills/next/references/handoff.md says: interactive, name /aida:implement <task-id> and stop; autonomous, invoke it when the run mode covers it.

What this skill never does

It never writes a single overview file or a file per component. The work orders are the whole deliverable.

It never mints a second id for a criterion or a non-goal. It only ever names the ids scope already minted in alignment.json.

It never treats a criterion nobody could confirm as owned as if it were owned. An autonomous run says plainly what it did not judge.

It never reads design/<id>.md, the rendered file, back into anything. It is for the four roles who build from a work order to read; nothing here parses it.

It never accepts an opinion from a source research marked as not accepted by this project as settling anything on its own. It informs the design; it does not decide it.