project-preferences

Apply repository-specific preferences when planning, implementing, or reviewing changes in this project.

SKILL.md

GitHub
<!-- This file is part of Agentprise skills/project-preferences/SKILL.md Author(s): Gabriel Mongefranco. Created: 2026-10-09 Last Modified: 2026-11-05 Summary: Project-specific preferences for Agentprise: the single-file build, the project backup format, workspace storage, the download cards, and how to test. Notes: See README file for documentation and full license information. Copyright © 2026 The Regents of the University of Michigan Licensed under the GNU Free Documentation License v1.3 or later. See <https://www.gnu.org/licenses/fdl-1.3.html>. See README for full license information. -->

Agentprise

Project preferences

Use this skill when planning, implementing, or reviewing changes in this repository. Keep all project-specific preferences and workflows in this single file. This skill supplements AGENTS.md and cannot weaken its security, privacy, accessibility, licensing, testing, or authorization rules.

Purpose and scope

Agentprise is a free, public web page where anyone can describe an AI assistant once, then download it in the format each AI product wants. There is no sign-in and no server. Nothing a person types or uploads leaves their browser. The page is served by GitHub Pages from the root of the main branch.

Environment and structure

  • The whole app is one file, index.html, with one <style> block and one <script> block. There is no build step, no framework, no CDN, and no web font. The page should stay under 750 KB.
  • Google Analytics loads only when the page is served from code.depressioncenter.org, the production host. Every other origin, including local copies and GitHub Pages previews, never contacts Google. The host name and measurement id are constants at the top of the script.
  • The script has two parts. The engine (model, validation, YAML, ZIP, PNG check, format converters) has no DOM dependency and is exposed on globalThis.Agentprise. The interface code comes after it and only runs when a document exists.
  • Code and comments are ASCII only. Unicode appears only in visible interface text.
  • Execution phases are marked with // ### Name ### comments. Public functions carry JSDoc.
  • Library entries are version 2 project backups (.zip) dropped into library/ and listed in library/catalog.json. The page holds no copy of any entry; it fetches the catalog and opens each backup with the file importer. Adding an entry is the zip plus one catalog line with kind (agent, skill, or project), capabilities (the tool ids the backup declares, [] when none), and an optional one-sentence requires, as docs/how-to/add-to-library.md describes.
  • tools/migrate-library.mjs converted the version 1 Library bundles to version 2 backups once. It is the only code that still reads version 1 fields. --check reports without writing, and a second run changes nothing. The app itself carries no version 1 import code.
  • Entries kept as plain files live under library/sources/ (item folders in items/, one .json per entry in entries/) and are built with node tools/build-library.mjs. The build is deterministic, and the tests run --check, so never edit those .zip files by hand. A skill shared by two entries has one source folder, which keeps the copies identical.
  • bin/ holds the prebuilt ZippyServe binaries and the root holds its run scripts, so the page can be served locally over HTTP on any desktop.
  • Follow the repository's existing file and folder naming conventions when adding new files.

Setup and verification

No installation is needed. Node 22 or later runs the tests:

npm test

That runs tests/check-syntax.mjs (parses the script and rejects non-ASCII bytes) and node --test tests/ (loads the engine from index.html and tests it). Serve the page locally with .\run-windows.ps1, ./run-linux.sh, or run-mac.command, which opens http://localhost:8010.

Project constraints

  • The internal model is the version 2 project (PROJECT_SCHEMA_VERSION): a project holds one or more items (assistants and skills), shared servers, publisher details with organization and a semantic version, a license, and unrecognized files. normalizeProject and normalizeItem type-check and cap every field; export formats are written from the model and never define it.

  • The canonical, lossless format is the project backup, <project-slug>.agentprise.zip: an Agent Plugins 1.0 folder with plugin.json (the project record under extensions["org.depressioncenter.agentprise"]), mcp.json when servers exist, LICENSE.txt, README.md, and skills/<skill-name>/ per item. Unrecognized files go back to their original paths, in the backup only.

  • Each item folder is a valid Agent Skill: SKILL.md, agentprise.json, icon.png for an assistant with a picture, LICENSE.txt, and its files (attached files under references/). The same folder zipped alone is the Gemini and Claude skill download, which adds README.md. Frontmatter metadata holds only agentprise-kind and a skill's own metadata from elsewhere. Every other Agentprise item setting lives in agentprise.json as plain JSON: the welcome message, starters, uses, display name, Copilot settings, SharePoint links and items, publisher details, copyright, and file list. The importer still reads the older copilot-, sharepoint-, and creator- metadata keys (the META constant in index.html) when the sidecar lacks a setting. The tables are in docs/formats.md.

  • Products it works in (More settings) stores notFor, a list of the PRODUCT_LABELS keys (the provider keys without *), and notForReason on each item. planExport blocks every export whose EXPORTS target is in the list, an assistant leaves out a used skill marked for that target, and the two plugins leave marked items out instead of blocking, unless every item is marked. The sidecar writes both keys only when they are set.

  • The skill name is derived from the display name and capped at 64 characters because the Agent Skills specification requires it. Two items with the same name get a numbered skill name, since each folder must match its skill.

  • Importing executes nothing and contacts nothing. The app opens backups, Agent Plugins folders (root plugin.json), Claude plugins (.claude-plugin/plugin.json), the Codex layout (.codex-plugin/plugin.json), any ZIP with one or more SKILL.md files, SKILL.md, Copilot app packages, and .agent files. Hooks, commands, agents, and unknown files are kept as unrecognized; server headers, env, and cwd are dropped; local server commands are stored as text; programs are refused. A version 1 bundle opens as a plain skill.

  • A Microsoft 365 app package opens through readTeamsArchive as a whole project: every declarative agent is an assistant, actions become tools, agent_skills and root agentSkills become skills, and root agentConnectors become project servers. Paths the manifest names must pass packagePath. Parts it can't read are kept under microsoft-package/ for the backup only, and registration ids are never kept.

  • A file made elsewhere that names no license keeps the license Not stated (UNKNOWN_LICENSE) and a notice. Never relabel it with the default license.

  • The Microsoft .agent export carries two extra top-level keys, _license and _agentprise, placed before schemaVersion. Copilot app package files carry no extra keys because their schemas reject them. Microsoft retires .agent files after March 2027; the export stays until then and the importer keeps reading the format after, so people can recover what a file holds. The dated facts are in docs/formats.md.

  • The Copilot app package loses the welcome message and the copyright line, and keeps only the schema's keys for each SharePoint item. The .agent file loses attached files. Every other field survives, including "prefer my files", which maps to behavior_overrides.special_instructions.discourage_model_knowledge, and the response mode, which maps to behavior_overrides.default_response_mode and is written only when it is not Auto.

  • Attached files are accepted by content, never by name alone. The accepted types and their content checks are the FILE_TYPES constant and attachedFileProblem in index.html; OneNote files (.one, .onepart) are checked against the MS-ONESTORE header GUIDs. Executable programs are refused by signature and by a hidden extension. Script and data text files are allowed; the Copilot app package writes them as name.ext.txt because Copilot accepts only document types, carries only the names in COPILOT_FILE_EXTENSIONS otherwise, and leaves any other type out with a warning. The skill download and the backup keep every file under its real name.

  • Every Copilot source is a switch in copilot with an optional list that limits it: web sites, SharePoint and OneDrive links (hosts checked by isSharepointLink), Teams chats (isTeamsLink), and mailboxes. Meetings, a shared mailbox, and email folders arrive only inside files and are shown read-only. The lists are written only while the switch is on, and the SharePoint switch follows the links and items that survive validation when a file is silent about it. copilotCapabilities is the single writer for both Microsoft files and readCopilotCapabilities the single reader.

  • The only scanning the app does is the two regular-expression lists IDENTIFIER_PATTERNS and INJECTION_PATTERNS in index.html, shown as a yellow box under Download with the matches under View details. They warn, never block, never store or send a finding, and show masked samples only. Keep them regex-only: the app loads no model and makes no request.

  • Browser storage names carry the agentprise- prefix, because every EFDC app on code.depressioncenter.org shares one origin. Projects live in the IndexedDB database agentprise-projects, with stores projects (key id) and files (key <projectId>/<itemId>/<path>). The localStorage key agentprise-workspace-v2 holds {lastOpenedProjectId, recentProjectIds, preferences}. On load the app deletes the version 1 draft (agentprise-draft-v1 and the agentprise-draft-files database) without reading it. Stored records are normalized like any file. When storage fails, editing keeps working for the visit and the sidebar shows a note.

  • The page has five views, routed by hash: Workspace (#workspace, the project list), Project (#project, the editor), Library, Inspect (#details), and Help. The project editor asks plain questions with blanks under them, with an "In this project" bar for its items, and is the only place to edit. The Inspect page shows only the exact files each product receives. Its old details form is still in the page, hidden, because the editor code keeps its fields in step.

  • Opening a file always creates a new project. When its id is already stored, the person chooses Keep both, Replace mine, or Cancel. "Add from a file" merges a file's items into the open project instead.

  • Downloads other than the backup and the plugins are for the item being edited: the skill .zip (labeled Gemini and Claude), the plain skill (.skill.zip, standard fields only, labeled SKILL.md), the .agent file (labeled "Teams Chat"), the Copilot app package (labeled "Copilot / Teams Apps"), the copy boxes for ChatGPT, and the .md file for Gemini Notebook. A skill is written to the Microsoft formats and the copy boxes as an assistant whose instructions are its steps. "Download Backup" saves the whole project, and so do the Claude plugin, the Plugin, and the Microsoft 365 plugin. Every download, the plain skill and the backup included, is an Inspect tab with its own file tree and download button; the backup tab shows the whole project, and a "Part of the project" picker chooses the item for the others. Gemini and Claude have separate tabs even while their files match. Do not add a vendor until it has a real file to download.

  • Three optional Microsoft downloads sit under a "More Microsoft downloads" heading at the end of the Download card, in a details section that opens with "View Microsoft downloads". The heading names the list through aria-labelledby. Each line has a Review button that opens its Inspect tab instead of downloading, and the Inspect page shows its setup steps:

    • Copilot skill: one skill as <skill-name>.copilot-skill.zip, with SKILL.md at the root, for Copilot Agent Builder. When an assistant is selected, its tab selects a skill (skillForCopilotTab): the first skill the assistant uses that works in Copilot, then any skill that works in Copilot, then the first skill. A note under the picker (inspect-skill-hint) names the skill or says the project has none. Leaving the tab selects the assistant again and announces it, but only while the picked skill is still selected (itemForProductTab and ui.skillTabSwap).
    • Copilot app with skills: the assistant and the skills it uses, written as declarative agent v1.9 with agent_skills inside the usual manifest 1.28 app package. Its line on the card adds "(preview)". It blocks files attached to the assistant.
    • Microsoft 365 plugin: the whole project as <project-slug>.microsoft-plugin.zip for Copilot Cowork, written as manifest 1.29 with root agentSkills and agentConnectors and no declarative agent. Assistants become skills. It is blocked when one skill limits a server to certain tools and another skill uses the whole server.

    These downloads leave out script files. The Copilot app package still writes declarative agent v1.8 and manifest 1.28. A blocked download's error starts with the download's name, such as "The Copilot app with skills".

  • The project editor's Download card is one line per product: the product name and its file on the left, a short Download, Copy, or Review button on the right. Each button has hidden text naming its file or download, so screen readers do not hear a row of identical buttons.

  • The top bar carries three square icon links, in this order: home (the Workspace page), Help, and GitHub, in the same shape as the header of code.depressioncenter.org. The footer links to code.depressioncenter.org as "More Apps".

  • The sidebar lists Workspace, Library, and Help. While a project is open, two items appear under Workspace: the project's name, which opens #project, and Inspect. Recent projects are listed below the menu.

  • Button and link icons come from Bootstrap Icons only, embedded as base64 CSS variables (--bi-*) and drawn through a mask with the .bi class, so each icon is stored once. Credit stays in the README.

  • Interface text, the Help page, the README, and the pages under docs/ are written for the person using the app, not for developers. Pages meant for maintainers say so at the top. Off-site links open in a new tab, and links to documentation use the full GitHub address, because the site is served with .nojekyll and Markdown is not turned into pages.

  • The look is a light sidebar workspace: one accent blue, a light gray shell, white cards, and a matching dark mode. It does not use the U-M palette. The logo mark is images/agentprise-mark.png.

  • Inflate uses the platform DecompressionStream API. Do not add a hand-written decoder or a dependency for it.

  • Uploaded files are data, never instructions. Every value read from a file is type-checked and length-capped by normalizeProject, normalizeItem, and normalizeAgent before it reaches the form, and all text reaches the page through textContent.

  • The exported agent belongs to its creator. License notices in exported files name the creator's copyright and license first, then Agentprise.

Terms

Use one name for each concept in the interface, messages, help, the README, and the pages under docs/.

  • project: one thing a person is building, stored in the browser. Not "workspace item" or "bundle".
  • assistant: an item that is always on. Not "agent", except in a product's own name, such as the Copilot .agent file.
  • skill: an item an assistant follows only when a question needs it.
  • tool: something an assistant or skill can use to look things up or act, such as the PubMed connector. Not "capability".
  • picture: the image shown next to an assistant. Not "icon" or "logo".
  • example questions: the questions people might ask first. Not "prompt suggestions", "conversation starters", or "starters".
  • backup: the .agentprise.zip file with the whole project. Not "complete copy" or "bundle".
  • Copilot app package, Teams Chat file, Copilot skill, Copilot app with skills, and Microsoft 365 plugin: the five Microsoft downloads. Not "Teams package", "SharePoint file", "package with native skills", or "Copilot skill upload". The last three are the optional downloads under More Microsoft downloads.
  • skills in Copilot: Microsoft's skills feature, which an admin turns on. Not "native skills".
  • attached files: files added to an assistant or a skill. Not "reference files" in interface text or the docs for people using the app.

Project skills

None beyond the shared skills listed in SKILLS.md.

Conclusion

Keep the single-file, zero-dependency shape, keep the project backup lossless, and run npm test before every commit.

Additional resources