forge
Take raw idea notes and adversarially decide whether each should become a skill — fold into an existing one, add as new, or drop — with an in-principle rating. Trigger on:- forge, assess skill ideas,
SKILL.mdskills/forge/SKILL.md
GitHubForge
Turns raw ideas into skill decisions. For each idea it models the strongest version of
the skill implied, runs an internal Proposer/Skeptic debate, and renders a verdict plus
a rating. This is notice's "Reviewing" step, done properly instead of skipped.
When to use this
docs/skill-ideas/has accumulated candidates and you want them triaged.- Pointed at someone else's skill directory to assess that set.
- Revision mode — a prose argument (not a scan dir) naming a change to what an
existing skill does: its procedure, triggers, tools, or a mis-calibration surfaced
by
/notice. No new idea file needed — skill management is the point. - Skip only for a pure wording fix to an existing skill — just make it.
Procedure
Read
.specify/memory/constitution.mdandskills/README.md(the authoring standard, the registry, and the Overlap check).Resolve the input — if the argument is prose describing a change to an existing skill rather than a path, this is revision mode: skip the scan. Model the target skill as it stands and the proposed change, run steps 3–5 on the change (does it earn its place in that skill, should it fold into a different one, does it bloat an always-loaded
description), write the report section (step 7), and on confirm apply theEditdirectly — no idea file to archive. Otherwise, scan: the argument as a dir, elsedocs/skill-ideas/.Glob*.md(non-recursive; the_dirs are skipped). Note whether this is local mode (docs/skill-ideas/) or external mode (any other path). Process each file in order:Proposer pass — model the strongest skill the idea implies: name, one-line purpose, trigger phrases, the procedure it would run, the artifact or outcome it produces, and the recurring need it serves. Steelman it. In external mode the input may already be a
SKILL.mddraft — model from that, don't reinvent it. Write this case into the report before starting the Skeptic pass.Skeptic pass — assume the skill should not exist; make the idea earn its place. Attack on four axes:
- Overlap — apply the Overlap check against every
skills/*/SKILL.md. Name any skill it duplicates or heavily shares discipline with. - Realness — is the need recurring or a one-off? Cite evidence:
docs/notice.md,git log, the idea file itself. - Skill-shaped — is this a reusable procedure, or a preference / fact /
one-time task that belongs in memory or
CLAUDE.md? - Context cost — does an always-loaded
descriptionfor this pay for itself?
- Overlap — apply the Overlap check against every
Synthesis — reconcile the two passes, citing named Proposer and Skeptic points (a named overlap must be addressed, not waved past). Apply the rule:
- Distinctness ≤ 2 → fold into
<existing>— unless Recurrence ≥ 4 and folding would push the host skill past ~120 lines, then add. - Skill-fit ≤ 2 or Recurrence ≤ 2 → drop (say which axis, and where the content should go instead if anywhere).
- Otherwise → add as new.
- Distinctness ≤ 2 → fold into
Rating, in principle — score independent of the verdict (a strong idea can still be
fold). Five axes, 1–5, one line of justification each:Axis 1 5 Recurrence seen once arises weekly Distinctness ~duplicate no neighbour Skill-fit belongs in memory clearly a repeatable procedure Scope tightness vague / broad single crisp trigger→outcome Context cost heavy for niche value cheap or high-value Write the report to
docs/skill-ideas/_assessments/<YYYY-MM-DD>.md— one file per run, each idea a section: modeled skill, Proposer case, Skeptic case, verdict, rating.Act on confirm — local mode only, per item:
- add — write
skills/<name>/SKILL.mdto the authoring standard, append a registry row toskills/README.md. - fold — show the exact Edit (target skill, section, added text); apply on yes.
- drop — record the reason.
Then move every processed idea file to
docs/skill-ideas/_processed/with the verdict appended. External mode stops after step 7 — report only.
- add — write
Constraints
- Proposer's full case is written to the report before the Skeptic pass begins.
- The Skeptic argues from "this should not exist"; a weak defence is a finding, not a pass.
- Synthesis must address every overlap the Skeptic named.
- External mode never writes outside this repo and never archives the scanned files.
- Never delete an idea file —
_processed/only. - Don't design the skill in depth here —
addproduces a first draft; a real design pass is/designit.
Done criteria
- Every
*.mdin scope has a verdict and a 5-axis rating in the report. - Confirmed
add/foldactions applied; registry updated. - Processed idea files moved to
_processed/(local mode).