Discover Conventions
Generate or refresh the project's AI-facing spec files. Keep everything short — these files load into every session, so every line costs tokens forever.
Scan (read evidence, don't guess)
- Manifests and configs: package.json / pyproject.toml / go.mod / Cargo.toml / pom.xml, lockfiles, linter/formatter/tsconfig, CI files
- Existing docs: README, CONTRIBUTING, docs/, existing AGENTS.md / CLAUDE.md / editor rules
- Code sample: 3-5 representative source files plus 1-2 test files; note naming, module structure, error handling, test style
- Git:
git log --oneline -20 for the commit message convention
Write
.superskills/conventions.md — single source of truth, max 80 lines:
- Commands: install / build / test / lint, the real ones from manifests
- Structure: one line per top-level directory that matters
- Conventions: only rules evidenced in code or docs — naming, imports, error handling, test layout, commit style
- Don'ts: only if evidenced (e.g. a lint rule or doc forbids it)
Every line must be project-specific and evidenced. Drop generic advice ("write clean code"). Prefer one line over three.
AGENTS.md (if missing) — max 20 lines: one-paragraph project description, key commands, then exactly these pointers:
Read .superskills/conventions.md before writing code.Check .superskills/learnings/INDEX.md for past learnings; read a linked entry when relevant.If anything in a request is unclear, do not guess — proactively trigger the superskills clarify skill to ask before coding; when the request is already specific, just implement.Before reporting any coding task as done you must have executed the changed code: run the documented examples plus boundary cases (empty, extreme, malformed, repeated/trailing input) via the test suite or a throwaway script you then delete, fix failures at root cause, and include the run output in your final reply. For larger changes apply the superskills test skill.
If AGENTS.md exists, only append the pointer lines when missing. Never rewrite existing user content.
CLAUDE.md (if missing) — exactly these imports:
@AGENTS.md
@.superskills/conventions.md
If it exists, append missing imports only.
.superskills/learnings/INDEX.md — create with a # Learnings header if missing.
Refresh mode
If .superskills/conventions.md already exists:
- Diff reality against it (recent commits, new configs, changed structure); update changed lines only, keep it within 80 lines.
- Fold learnings from
.superskills/learnings/ that have hardened into stable conventions directly into conventions.md, then delete those learning files and their INDEX lines.
Finish by suggesting a commit of .superskills/, AGENTS.md, and CLAUDE.md.
1---2name: discover3description: Discover an existing project's conventions and generate or refresh minimal AI spec files (.superskills/conventions.md, AGENTS.md, CLAUDE.md). Use when a project lacks AGENTS.md/CLAUDE.md, when the user asks to generate conventions ("生成规范", "init ai docs", "discover conventions"), or when conventions are reported stale.4---56# Discover Conventions78Generate or refresh the project's AI-facing spec files. Keep everything short — these files load into every session, so every line costs tokens forever.910## Scan (read evidence, don't guess)1112- Manifests and configs: package.json / pyproject.toml / go.mod / Cargo.toml / pom.xml, lockfiles, linter/formatter/tsconfig, CI files13- Existing docs: README, CONTRIBUTING, docs/, existing AGENTS.md / CLAUDE.md / editor rules14- Code sample: 3-5 representative source files plus 1-2 test files; note naming, module structure, error handling, test style15- Git: `git log --oneline -20` for the commit message convention1617## Write18191. `.superskills/conventions.md` — single source of truth, max 80 lines:20 - **Commands**: install / build / test / lint, the real ones from manifests21 - **Structure**: one line per top-level directory that matters22 - **Conventions**: only rules evidenced in code or docs — naming, imports, error handling, test layout, commit style23 - **Don'ts**: only if evidenced (e.g. a lint rule or doc forbids it)2425 Every line must be project-specific and evidenced. Drop generic advice ("write clean code"). Prefer one line over three.26272. `AGENTS.md` (if missing) — max 20 lines: one-paragraph project description, key commands, then exactly these pointers:28 - `Read .superskills/conventions.md before writing code.`29 - `Check .superskills/learnings/INDEX.md for past learnings; read a linked entry when relevant.`30 - `If anything in a request is unclear, do not guess — proactively trigger the superskills clarify skill to ask before coding; when the request is already specific, just implement.`31 - `Before reporting any coding task as done you must have executed the changed code: run the documented examples plus boundary cases (empty, extreme, malformed, repeated/trailing input) via the test suite or a throwaway script you then delete, fix failures at root cause, and include the run output in your final reply. For larger changes apply the superskills test skill.`3233 If AGENTS.md exists, only append the pointer lines when missing. Never rewrite existing user content.34353. `CLAUDE.md` (if missing) — exactly these imports:3637 ```38 @AGENTS.md39 @.superskills/conventions.md40 ```4142 If it exists, append missing imports only.43444. `.superskills/learnings/INDEX.md` — create with a `# Learnings` header if missing.4546## Refresh mode4748If `.superskills/conventions.md` already exists:49- Diff reality against it (recent commits, new configs, changed structure); update changed lines only, keep it within 80 lines.50- Fold learnings from `.superskills/learnings/` that have hardened into stable conventions directly into conventions.md, then delete those learning files and their INDEX lines.5152Finish by suggesting a commit of `.superskills/`, `AGENTS.md`, and `CLAUDE.md`.
mrlyk