agentleFS
pandazkipandazki/pneuma-skills/.agents/skills/create-mode/SKILL.md

create-mode

Skill161 starsChanged 10 days agoMITView on GitHub

Create or fork a Pneuma mode, scaffold its manifest, viewer, skill, seeds, and showcase. Use for new-mode authoring in this repository; discover unresolved requirements, document the design brief, then implement and verify.

The file

We split this skill into the rules it gives an agent

12 rules. About 0 are common, 4 are distinctive, and 8 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈4

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 8

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    If the user *only* asks about an existing mode's behavior, this skill is not the right tool — direct them to …

    line 31
  2. Very specific:

    If you ever find yourself stuck choosing between two patterns, open references/case-studies.md — it indexes which existing mode made which …

    line 64
  3. Distinctive:

    Use the visibility decision recorded in the brief.

    line 202
  4. Very specific:

    If the mode is not hidden, add its row to the "Built-in Modes" table and update CLI usage in both README.md and README.zh.md.

    line 247
  5. Distinctive:

    Verification obligation (hosted player): Never whitelist a mode on the strength of reading code.

    line 276
  6. Very specific:

    Use the visibility decision already recorded in the brief.

    line 292
  7. Distinctive:

    Don't claim the mode is ready until you verify these:

    line 305
  8. Very specific:

    The launcher's /api/registry includes the new entry: Test via curl -s http://localhost:17996/api/registry | jq '.builtins[].name' (or …

    line 309
  9. Very specific:

    Cloud surfaces match the brief: If the brief said *no* to both, verify the mode's name appears in neither core/player-support.ts nor either …

    line 312
  10. Very specific:

    One noun for "which object" — ViewerAddress: Every action that takes an object reference, every notification that reports one, every …

    line 332
  11. Distinctive:

    Each action serves a concrete task: Start with the smallest sufficient action space.

    line 333
  12. Very specific:

    manifest.ts declares; pneuma-mode.ts implements: Keep the split. Manifest is read by skill-installer + backend.

    line 334

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Create Mode

A guided journey for adding a new mode to Pneuma Skills. The journey has three phases — Discovery (ask the right questions), Brief (write down key choices and resolve material unknowns), Implementation (generate files). Each phase has a clear handoff to the next; never skip Brief. Pneuma has a stable contract layer; a concise design brief keeps the viewer aligned with the user's intended work.

This skill runs in Claude Code and Codex. Use the active harness's file, planning, browser, and question tools; native Claude tool names are not prerequisites. Resolve paths from this skill directory.

Apply the root product and architecture guidance and Engineering Judgment to the brief: identify domain invariants and ownership before choosing Source kinds or actions, and inspect existing modes and mature implementations before building anew.

The reference material in references/ is where the knowledge lives — go read the relevant one whenever you're about to make a meaningful decision. SKILL.md is the journey, not the textbook.


When to use

Trigger this skill when the user asks for any of:

  • "create a new mode for X"
  • "fork slide / webcraft / … for a different domain"
  • "scaffold a mode"
  • "add a [mindmap | spreadsheet | timeline | annotator | …] mode"
  • "design the viewer for a mode that …"

If the user only asks about an existing mode's behavior, this skill is not the right tool — direct them to docs/reference/viewer-agent-protocol.md or the mode's own SKILL.md.


Phase 1 — Discovery interview

Goal: fill the design brief from the user's request, existing decisions, and repository evidence. Ask only for missing choices that materially affect the result. Use the question tool available in the active harness, or a concise chat question; never require a tool named AskUserQuestion. Bundle closely related unknowns and continue independent work while waiting.

Discovery checklist (reuse known answers; ask only where needed)

  1. Identity — name (kebab-case), one-line displayName, two-line description, intended icon style. This is the only question you can pose as a single multi-line form.
  2. Domain in one sentence — what is the user creating with this mode? A document? A canvas of objects? A timeline? Let the user answer free-form before you offer Source-kind options. Read references/domain-and-sources.md while they think.
  3. Inspiration vs original — does this mode borrow content (commands, references, design language, taxonomy) from an existing tool, library, or project? If yes, ask the upstream's name + URL + license. This determines whether you'll write NOTICE.md and set inspiredBy. Read references/external-integrations.md for the borrow-vs-inspiration line.
  4. Source kind (branch on Q2 + Q3) — present file-glob / json-file / aggregate-file / memory with the one that fits Q2's domain pre-selected as "Recommended". Explain why it fits in the option's description. If aggregate-file wins, note that you'll also generate domain.ts.
  5. Workspace model (when Q4 is not memory) — "all" / "manifest" / "single"; do users author many independent files, an ordered/structured set, or one main document? See references/viewer-contract-patterns.md for the FileWorkspaceModel matrix.
  6. ViewerAddress vocabulary — "what's the smallest thing the user can point at?" Propose a draft { contentSet?, ... } based on Q2's domain noun (slide / page / row / node / heading). Confirm with the user; explicitly name the coarse "where" key and any fine "within" key. See references/viewer-contract-patterns.md::ViewerAddress.
  7. Initial action space — propose the smallest sufficient set with id / label / category / agentInvocable, tying each action to a concrete task. Include navigate-to when users need addressable navigation; add ui and custom for demonstrated needs. Don't list capture — it's framework-built-in.
  8. External integrations (conditional — only ask if Q2 or Q3 implied an external API / SDK / CDN / library / API key) — does the viewer fetch external APIs (→ proxy)? does the agent or viewer need API keys (→ init.params with sensitive: true + envMapping)? does this need an MCP server (→ skill.mcpServers)? Read references/external-integrations.md for the proxy / Babel-JIT / NOTICE patterns.
  9. Cloud surfaces — resolve both choices in every brief; reuse existing decisions and ask only when unclear. Two independent questions, in this order: (a) should a finished piece of work in this mode be shareable as a read-only page a stranger can open with no Pneuma installed? — that's the hosted player, and it obligates the viewer to render from workspace files alone, with no live backend; (b) does the mode produce a deployable static site? — that's Vercel / Cloudflare Pages deploy. Put the real trade-off in each option's description: "yes" buys shareability and costs a read-only degradation path plus a browser verification pass on every viewer change; "no" costs nothing and can be revisited in a later release. Read references/cloud-surfaces.md before you ask — the compatibility checklist there is what "yes" actually commits to.
  10. Seed strategy — single file, multiple use-case content sets, or language×theme matrix? What's the first seed's narrative — what story does it tell to a brand-new user? See references/seed-and-showcase.md.
  11. Evolution directive — give the evolve agent a one-sentence "what should it learn for this mode?" (e.g., "Learn the user's slide design preferences: typography, palette, density, structure"). This is what makes the mode personalize over time.

What to read while interviewing

When you're about to ask …Read first
Q2 / Q4 (domain → source kind)references/domain-and-sources.md
Q5 / Q6 / Q7 (workspace / address / actions)references/viewer-contract-patterns.md
Q3 / Q8 (inspiration / external deps)references/external-integrations.md
Q9 (cloud surfaces)references/cloud-surfaces.md
Q10 (seed strategy)references/seed-and-showcase.md
Q11 (evolution directive)references/skill-md-patterns.md (evolution section)

If you ever find yourself stuck choosing between two patterns, open references/case-studies.md — it indexes which existing mode made which choice, so you can read that mode's manifest as a concrete precedent.


Phase 2 — Design brief & user confirmation

Goal: write down every key choice with a one-line rationale in one place. Use the brief for Phase 3; label assumptions and resolve material unknowns before dependent implementation. An approved brief or explicit instruction to implement already supplies authorization.

Brief structure

Present a concise brief in chat, or link the design document if the task already uses one. Cover the applicable fields below without asking the user to repeat known decisions:

# Mode design brief — <displayName>

## Identity
- name: <kebab-case>
- displayName: <string or LocalizedString>
- description: <one sentence>
- icon: <SVG approach: e.g. "lucide-style line icon, single path">

## Domain
<one paragraph — what the user is creating; what the viewer renders; what the agent does>

## Invariants and implementation choice
- required invariants: <what must remain true, and how it will be verified>
- state and writers: <persistent work, transient UI state, and who may change each>
- reuse: <existing mode / project implementation / library, or the concrete gap>
- added abstraction or complexity: <its contract or real variation and benefit; omit if none>

## Source layer
- kind: <file-glob | json-file | aggregate-file | memory>
- domain type T: <the TypeScript type the viewer subscribes to, sketched>
- why this kind: <one sentence — see references/domain-and-sources.md>
- domain.ts needed: <yes | no>

## Workspace model
- type: <"all" | "manifest" | "single">
- multiFile: <true | false>
- ordered: <true | false>
- hasActiveFile: <true | false>
- supportsContentSets: <true | false>

## ViewerAddress vocabulary
- coarse keys: <e.g. `contentSet?`, `slide`>
- fine keys: <e.g. `selector?`, `anchor?`>
- example address: `{ contentSet: "en-light", slide: 3 }`
- documented in: skill/SKILL.md (will write a sub-section)

## Action space
| id | label | category | agentInvocable | params |
|----|-------|----------|----------------|--------|
| navigate-to | Go to … | navigate | true | { address: object } |
| … | … | … | … | … |

(framework provides `capture` automatically — not listed)

## Seed strategy
- shape: <single | content-sets-by-use-case | language×theme>
- content sets: <list, with each name + one-line purpose>
- first seed narrative: <one sentence>

## External integrations
- proxy: <none | list routes>
- init.params: <none | list with sensitive flag>
- skill.mcpServers: <none | list>
- viewer.refreshStrategy: <"auto" | "manual">
- NOTICE.md required: <yes | no — if yes, upstream name + license + version pinned>
- inspiredBy: <none | { name, url }>
- external effects (if any): <observable completion, retry/idempotency, cancellation, and recovery or undo limits>

## Cloud surfaces
- hosted player: <yes | no — and the reason, in the vocabulary of references/cloud-surfaces.md>
- artifact deploy: <none | vercel + cf-pages>
- obligations (only when either is yes):
  - registration: <`core/player-support.ts` whitelist entry / `compatibleModes` entry in BOTH deploy plugins + an `/export/<name>` route>
  - read-only degradation: <which affordances hide when `editing === false`; which `/api/*` calls gate on the `staticPlayer` store flag>
  - verification: <build the player, load a real package for THIS mode in a browser, exercise it read-only, console clean>

## Launcher surface
- visibility: <public (in gallery) | hidden (internal-only, manifest.hidden=true)>
- featured-eligible: <yes (default; showcase highlights present) | no (no showcase or hidden mode)>

## Evolution directive
> <one sentence to the evolve agent>

## Open questions / deferred
- <anything we punted on; e.g. "showcase imagery defers to /showcase">

Before implementation

If the user has authorized implementation and the brief fits that scope, proceed. Ask only when a material unresolved choice would change the product or exceed the agreed scope. Keep the brief current as the user steers; do not request approval again for decisions already made.


Phase 3 — Implementation

Once the brief is sufficiently resolved and implementation is authorized, generate files in this order. Use templates from assets/templates/; replace the TODO: placeholders against the brief. Don't ad-lib structure — the templates encode the conventions extracted from existing modes.

Step 1 — Scaffold the directory

modes/<name>/
├── manifest.ts          ← from assets/templates/manifest.ts.template
├── pneuma-mode.ts       ← from assets/templates/pneuma-mode.ts.template
├── domain.ts            ← only if Source kind is aggregate-file; from domain.ts.template
├── skill/
│   └── SKILL.md         ← from assets/templates/SKILL.md.template
├── seed/
│   └── <content sets per brief>
├── viewer/
│   └── <ModeName>Preview.tsx   ← scaffold a stub PreviewComponent
└── showcase/
    └── showcase.json    ← from assets/templates/showcase.json.template (with concept descriptions)

NOTICE.md                ← only if brief said "NOTICE.md required: yes"; from NOTICE.md.template

Step 2 — Wire up file-by-file

For each file, fill in templates against the brief. Specifics:

  • manifest.ts — every brief field maps to a manifest field. The template has marked sections (// TODO: identity, // TODO: sources, etc.) — fill each from the brief. Don't add fields the brief doesn't have; brevity over completeness for v0.1.0.
  • pneuma-mode.ts — the ModeDefinition binding: import manifest, wire it to a stub ViewerContract that imports the PreviewComponent and implements extractContext, workspace.resolveItems, workspace.createEmpty. See references/viewer-contract-patterns.md::pneuma-mode.ts for the binding pattern.
  • domain.ts (aggregate-file only) — write the load(files) → T | null and save(value, current) → { writes, deletes } pair as pure functions. Read existing modes' domain.ts for the pattern (slide / illustrate / kami use this).
  • skill/SKILL.md — follow references/skill-md-patterns.md: Scene → Viewer Contract → Core Rules → Workflow → Commands → References. Include a ## ViewerAddress vocabulary sub-section that names every key from the brief and a one-line meaning per key.
  • viewer/<Name>Preview.tsx — stub. Renders a placeholder ("Mode initialized — start authoring"). Imports the Source from props.sources via useSource. The user (or you in a follow-up) will flesh this out.
  • seed/ — write the first content set's files per the brief's narrative.
  • showcase/showcase.json — from template, with brief's tagline + 3 highlight concept descriptions. Images are generated in Step 4.
  • NOTICE.md (if required) — pin upstream name + URL + license + version + sync date; include the "what we borrowed / what we adapted / what we dropped" mapping table. Template at assets/templates/NOTICE.md.template.

Step 3 — Register the mode and its supported surfaces

Complete frontend registration (3a), launcher discovery (3b), and public documentation (3c) as applicable. These serve different consumers: a mode can launch successfully while remaining absent from the gallery or docs.

Use the visibility decision recorded in the brief. Public modes need all three; hidden modes still need frontend registration but stay out of public catalogs and may omit the gallery entry as described below. Resolve visibility only if it is still unknown. Cloud registration (3d) is conditional on the brief's ## Cloud surfaces decision and the verification pass described there.

3a. Frontend dynamic-import registry — core/mode-loader.ts (bundled modes only)

builtinModes lists exactly the bundled modes in modes/distribution.json (core/__tests__/mode-loader-external.test.ts pins the equality). A catalog mode gets no entry here: it is discovered from its on-disk manifest.ts and loaded as a built catalog bundle, so skip this step for it. For a bundled mode, add an entry to the builtinModes: Record<string, ModeSource> map so the frontend can dynamic-import the mode's manifest and viewer. Without it, a bundled mode 404s when a user opens its URL ("Unknown mode: <name>").

// core/mode-loader.ts — inside `const builtinModes: Record<string, ModeSource> = { ... }`
<name>: {
  type: "builtin",
  manifestLoader: () =>
    import("../modes/<name>/manifest.js").then((m) => m.default),
  definitionLoader: () =>
    import("../modes/<name>/pneuma-mode.js").then((m) => m.default),
},

Copy the shape from the neighboring entry rather than from memory — the field names are manifestLoader / definitionLoader, and the type: "builtin" discriminant is required.

3b. Launcher gallery discovery and distribution

The launcher's /api/registry scans on-disk modes/*/manifest.ts files. A new directory is discovered automatically; do not add a hardcoded name to server/index.ts. hidden: true is the public-picker filter.

For a bundled mode, add its name to modes/distribution.json and its directory to package.json's files, then add the frontend entry in Step 3a. A catalog mode follows the catalog publishing workflow instead. Hosted-player support requires a viewer in the separate player build registry. Verify both the registry response and the actual gallery; source discovery and release packaging are separate concerns.

3c. Docs — public mode catalogs

If the mode is not hidden, add its row to the "Built-in Modes" table and update CLI usage in both README.md and README.zh.md. Hidden modes stay out of the public catalog. AGENTS.md links to the catalog; it does not maintain another mode list. CLAUDE.md remains the one-line @AGENTS.md import: never add content to it or copy it over AGENTS.md.

3d. Cloud surfaces — conditional, driven by the brief

Unlike 3a–3c, this one is not universal. Add each entry only if the brief's ## Cloud surfaces section said yes; a mode that answered "no" is correctly absent from both files, and adding it speculatively ships a broken share link.

  • Hosted player — add the viewer import to src/player/player-modes.ts so the player build contains it, then append the mode name to WEB_PLAYER_SUPPORTED_MODES in core/player-support.ts after verification. The whitelist entry is the last thing you do: the whitelist is a claim that the viewer has been exercised in a real player build. See the verification obligation below.
  • Artifact deploy — add the mode name to compatibleModes in both plugins/vercel/manifest.ts and plugins/cf-pages/manifest.ts. Membership there only makes the deploy providers resolve for the session; the button itself lives on the mode's /export/<name> page (server/routes/export.ts + server/routes/deploy-ui.ts), which needs a mode-specific collectDeployFiles(). Listing the mode without building that page produces nothing — doc and gridboard are both listed today and neither has an export route.

Verification obligation (hosted player). Never whitelist a mode on the strength of reading code. Build the player (bunx vite build --config vite.player.config.ts), materialize a real package for this mode and serve it from one origin (copy scripts/smoke-player.ts; scripts/smoke-webcraft.ts and scripts/smoke-kami.ts are the mode-specific precedents), open it in a browser, and exercise the viewer read-only — content sets, item navigation, timeline scrub — with the console clean. The failure modes here all look fine in source: an empty viewer because the mode's file extension isn't in the package's text allowlist, an asset path the content service worker can't resolve, a viewer stuck on "Loading…" waiting for a signal the player never sends. references/cloud-surfaces.md carries the full checklist.

Featured vs. hidden

Use the visibility decision already recorded in the brief. Public modes with showcase highlights are eligible for the random featured slot. hidden: true removes the mode from user pickers entirely; it is not a separate "do not feature" switch. Do not ask again after registration or hide a public mode to avoid featuring.

Step 4 — Generate showcase imagery

Hand off to the existing showcase workflow. Read .agents/skills/showcase/SKILL.md and execute its Step 3 (Generate Showcase Images) for the new mode — hero + 3 highlight images, 1376×768, "Ethereal Tech Dark Mockup" style, saved to modes/<name>/showcase/. The descriptions you put in showcase.json during Step 2 become the briefs for image generation.

This is the only Phase-3 step that takes appreciable time. If image generation isn't available right now (no API key, offline), surface that to the user and let them decide whether to defer — showcase.json with the right descriptions but missing images is a valid intermediate state.

Step 5 — Sanity check

Don't claim the mode is ready until you verify these:

  1. modes/<name>/manifest.ts type-checks against core/types/mode-manifest.ts (bun run typecheck runs clean from the repository root).
  2. A disposable bun run dev <name> --no-open --viewing session starts without error. Use an isolated workspace; do not resume an unrelated agent just to inspect the viewer.
  3. The launcher's /api/registry includes the new entry. Test via curl -s http://localhost:17996/api/registry | jq '.builtins[].name' (or whatever port the launcher is on). If the name is missing, check its on-disk manifest and public visibility; Step 3b describes current discovery and packaging.
  4. The launcher's mode gallery shows the new entry (same — say so if you can't run the launcher).
  5. There are no lingering TODO: comments from the template you didn't address.
  6. Cloud surfaces match the brief. If the brief said no to both, verify the mode's name appears in neither core/player-support.ts nor either deploy plugin's compatibleModes — a speculative entry ships a broken share link or a dead Deploy button. If the brief said yes to the hosted player, the browser pass from Step 3d must have actually happened: player built, real package for this mode loaded, viewer exercised read-only, console clean. If you couldn't run it, say so explicitly and leave the whitelist entry out until someone can — an unverified whitelist entry is worse than a missing one, because supported is baked into every package at share time and a package exported while the flag was wrong stays wrong until it's re-shared.

Closing principles

These show up in every existing mode; honor them in the one you're creating too.

  1. Domain-first, transport later. Define the domain type T before choosing how it serializes. Source kind is a consequence of T, not a prior decision.
  2. One noun for "which object" — ViewerAddress. Every action that takes an object reference, every notification that reports one, every locator card that points to one, must use the same address shape. Mode owns the vocabulary; framework owns the slot.
  3. Each action serves a concrete task. Start with the smallest sufficient action space; two to five actions is a common pattern, not a quota. Distinguish agent-to-viewer actions from user-to-agent commands (⑥), and reuse built-in actions.
  4. manifest.ts declares; pneuma-mode.ts implements. Keep the split. Manifest is read by skill-installer + backend; pneuma-mode.ts is read by the frontend mode-loader. Don't put React imports in manifest.ts.
  5. SKILL.md is the agent's project guide for this mode — it follows the same "scene → contract → rules → examples → references" rhythm as the root AGENTS.md does for the project. Put depth in skill/references/<topic>.md files, not in the main body.
  6. Borrowed content needs a NOTICE.md; borrowed ideas don't. Direct transcription, license excerpts, command tables, font subsets → declare upstream + license + version. Architectural metaphors, aesthetic direction, workflow philosophy → no notice needed.
  7. Showcase is mandatory, but imagery can defer. showcase.json with descriptions and a tagline is the minimum bar (so the launcher gallery has copy); imagery generation can happen later via the existing /showcase flow.

References

Open the matching file when you're about to make the corresponding decision. Don't load them all eagerly — progressive disclosure.

FileWhen to read
references/mode-anatomy.mdFirst touch — overview of the directory shape, required vs optional files, manifest field matrix
references/domain-and-sources.mdPicking Source kind, designing domain type T, writing domain.ts
references/viewer-contract-patterns.mdWiring ViewerContract, choosing ViewerAddress vocabulary, designing workspace.resolveItems
references/skill-md-patterns.mdWriting skill/SKILL.md and the evolution directive
references/seed-and-showcase.mdDesigning seed content sets and showcase.json
references/external-integrations.mdproxy routes, JIT compilation, API-key params, NOTICE.md mechanics
references/cloud-surfaces.mdDeciding hosted-player support and artifact deploy — what the player environment is, the viewer compatibility checklist, the static-web fast path, the disqualifiers, how to verify before whitelisting
references/case-studies.md"Where did <existing mode> make this choice?" — index by pattern, not by mode

Templates in assets/templates/ are the concrete files you'll write from. Each template has TODO: markers where the brief plugs in.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.