agentleFS
fortunto2fortunto2/solo-factory/skills/build/SKILL.md

solo-build

Skill18 starsChanged 4 months agoMITView on GitHub

Use when "build it", "start building", "execute plan", "implement tasks", "ship it", track ID referenced, or plan tasks need execution. Do NOT use for planning (/plan) or scaffolding (/scaffold).

  • Deletes or force-pushes

The file

Show its rules and text

We split this skill into the rules it gives an agent

53 rules. About 0 are common, 24 are distinctive, and 29 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈24

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 29

    Very specific

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

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

    After /plan has created a track with spec.md + plan.md.

    line 27
  2. Distinctive:

    If MCP tools are not available, fall back to Glob + Grep + Read.

    line 37
  3. Very specific:

    Check docs/plan/*/plan.md — standard location

    line 42
  4. Very specific:

    DO NOT: search for conductor/ or any other directory — only docs/plan/.

    line 44
  5. Distinctive:

    Commit strategy (conventional commits format)

    line 48
  6. Very specific:

    If docs/workflow.md missing: use defaults (moderate TDD, conventional commits).

    line 51
  7. Very specific:

    Read the stack YAML (templates/stacks/{stack}.yaml) — the pre_commit field tells you which system and what it runs:

    line 55
  8. Very specific:

    Don't use --no-verify on commits — if hooks fail, fix the issue and commit again.

    line 75
  9. Very specific:

    If not found: search docs/plan/*/plan.md for partial matches, suggest corrections.

    line 81
  10. Very specific:

    Search for plan.md files in docs/plan/.

    line 87
  11. Very specific:

    Read each plan.md, find tracks with uncompleted tasks.

    line 88
  12. Distinctive:

    If multiple, ask via AskUserQuestion.

    line 89
  13. Very specific:

    If zero tracks: "No plans found.

    line 90
  14. Distinctive:

    Do NOT read source code files at this stage: Only docs. Source files are loaded per-task in the execution loop (step 3 below).

    line 108
  15. Very specific:

    If a task is marked [~] in plan.md:

    line 112
  16. Distinctive:

    Ask via AskUserQuestion, then proceed.

    line 123
  17. Very specific:

    Read references/context-engineering.md for full rules on observation masking, attention positioning, and plan recitation.

    line 127
  18. Very specific:

    IMPORTANT — All-done check: Before entering the loop, scan plan.md for ANY - [ ] or - [~] tasks.

    line 136
  19. Very specific:

    Update plan.md: [ ] → [~] for current task.

    line 146
  20. Distinctive:

    Read ONLY the files explicitly mentioned in the task description (file paths).

    line 159
  21. Very specific:

    If the task doesn't mention files, use Grep with a narrow pattern on src/ or app/ — never /*.

    line 161
  22. Very specific:

    Read references/quality-tools.md for full commands per stack.

    line 167
  23. Distinctive:

    Run after each task implementation, before git commit.

    line 176
  24. Very specific:

    Read references/tdd-seams.md before the first test of a feature — it covers where seams go, vertical vs horizontal slicing, and the four …

    line 192
  25. Distinctive:

    Run existing tests to check nothing broke.

    line 197
  26. Very specific:

    If the task touches core business logic (pipeline, algorithms, agent tools), run make integration (or the integration command from …

    line 201
  27. Distinctive:

    After implementation, run a quick visual smoke test if tools are available:

    line 205
  28. Distinctive:

    If you have Playwright MCP tools or browser tools available:

    line 208
  29. Very specific:

    Start the dev server if not already running (check stack YAML for dev_server.command)

    line 209
  30. Distinctive:

    Check the browser console for errors (hydration mismatches, uncaught exceptions, 404s)

    line 211
  31. Distinctive:

    If the task affects responsive layout, resize to mobile viewport (375px) and check

    line 213
  32. Distinctive:

    If instructed to use iOS Simulator in the pipeline prompt:

    line 216
  33. Very specific:

    Check simulator logs: xcrun simctl spawn booted log stream --style compact --timeout 10

    line 220
  34. Distinctive:

    If instructed to use Android Emulator in the pipeline prompt:

    line 223
  35. Very specific:

    Graceful degradation: If browser/simulator/emulator tools are not available or fail — skip visual checks entirely.

    line 229
  36. Distinctive:

    Capture SHA: after commit:

    line 241
  37. Distinctive:

    SHA annotation in plan.md: After every task commit:

    line 246
  38. Very specific:

    After each task, check if all tasks in current phase are [x].

    line 254
  39. Very specific:

    Run verification steps listed under ### Verification for the phase.

    line 259
  40. Distinctive:

    Run full test suite.

    line 260
  41. Very specific:

    Commit plan.md progress: git commit -m "chore(plan): complete phase {N}".

    line 263
  42. Distinctive:

    Ask via AskUserQuestion. Do NOT automatically continue past failures.

    line 294
  43. Very specific:

    When all phases and tasks are [x]:

    line 298
  44. Distinctive:

    Check acceptance criteria from spec.md:

    line 315
  45. Very specific:

    Update spec.md checkboxes: - [ ] → - [x] for verified criteria.

    line 318
  46. Distinctive:

    If any criterion FAILS → do NOT mark plan complete.

    line 319
  47. Very specific:

    Commit spec updates: git add docs/plan/*/spec.md && git commit -m "docs: update spec checkboxes (verified by build)"

    line 320
  48. Distinctive:

    If the same build/test error occurs 3+ times in a row, STOP retrying.

    line 330
  49. Distinctive:

    Do NOT burn iterations on identical failures.

    line 334
  50. Very specific:

    On session start: Read plan.md, find all incomplete tasks ([ ] and [~]).

    line 394
  51. Distinctive:

    Foundational principle: If you wouldn't mass-delete production data without checking, don't skip a test without evidence it's unnecessary.

    line 425
  52. Very specific:

    Fix: Run /plan "your feature" first to create a track.

    line 444
  53. Distinctive:

    Fix: Use the error handling flow — attempt fix, rollback if needed, pause for user input.

    line 448

≈ “Common” is an estimate: at least 1 in 5 of the 372,429 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

/build

This skill is self-contained — follow the task loop, test rules, and completion flow below instead of delegating to external build/execution skills (superpowers, etc.).

Execute tasks from an implementation plan. Finds plan.md (in docs/plan/), picks the next unchecked task, implements it with tests at agreed seams, commits, and updates progress.

Live Context

  • Branch: !git branch --show-current 2>/dev/null
  • Status: !git status --short 2>/dev/null | head -10
  • Last commit: !git log --oneline -3 2>/dev/null

When to use

After /plan has created a track with spec.md + plan.md. This is the execution engine.

Pipeline: /plan → /build → /deploy → /review

MCP Tools (use if available)

  • session_search(query) — find how similar problems were solved before
  • symbol_search("<words>") — find reusable code across projects (symbol names, every project; Grep inside one)
  • Grep — file dependencies, imports, callers (a structural graph answered these no better: measured 27 Oct 2026)

If MCP tools are not available, fall back to Glob + Grep + Read.

Pre-flight Checks

  1. Detect context — find where plan files live:
    • Check docs/plan/*/plan.md — standard location
    • Use whichever exists.
    • DO NOT search for conductor/ or any other directory — only docs/plan/.
  2. Load workflow config from docs/workflow.md (if exists):
    • TDD strictness (strict / moderate / none)
    • Commit strategy (conventional commits format)
    • Verification checkpoint rules
    • Integration Testing section — if present, run the specified CLI commands after completing tasks that touch the listed paths If docs/workflow.md missing: use defaults (moderate TDD, conventional commits).
  3. Verify git hooks are installed:

    Read the stack YAML (templates/stacks/{stack}.yaml) — the pre_commit field tells you which system and what it runs:

    • husky + lint-staged → JS/TS stacks (eslint + prettier + tsc)
    • pre-commit → Python stacks (ruff + ruff-format + ty)
    • lefthook → mobile stacks (swiftlint/detekt + formatter)

    Then verify the hook system is active:

    # husky
    [ -f .husky/pre-commit ] && git config core.hooksPath | grep -q husky && echo "OK" || echo "NOT ACTIVE"
    # pre-commit (Python)
    [ -f .pre-commit-config.yaml ] && [ -f .git/hooks/pre-commit ] && echo "OK" || echo "NOT ACTIVE"
    # lefthook
    [ -f lefthook.yml ] && lefthook version >/dev/null 2>&1 && echo "OK" || echo "NOT ACTIVE"

    If not active — install before first commit:

    • husky: pnpm prepare (or npm run prepare)
    • pre-commit: uv run pre-commit install
    • lefthook: lefthook install

    Don't use --no-verify on commits — if hooks fail, fix the issue and commit again.

Track Selection

If $ARGUMENTS contains a track ID:

  • Validate: {plan_root}/{argument}/plan.md exists (check docs/plan/).
  • If not found: search docs/plan/*/plan.md for partial matches, suggest corrections.

If $ARGUMENTS contains --task X.Y:

  • Jump directly to that task in the active track.

If no argument:

  1. Search for plan.md files in docs/plan/.
  2. Read each plan.md, find tracks with uncompleted tasks.
  3. If multiple, ask via AskUserQuestion.
  4. If zero tracks: "No plans found. Run /plan first."

Context Loading

Step 1 — Architecture overview (if MCP available)

project_info(name="{project name}")

Returns: stack, languages, directory layers, key patterns, top dependencies, files with the most definitions.

Step 2 — Essential docs (parallel reads)

  1. docs/plan/{trackId}/plan.md — task list (REQUIRED). Read the ## Context Handoff section first — it has a compact summary of intent, key files, decisions, and risks. This is your primary orientation.
  2. docs/plan/{trackId}/spec.md — acceptance criteria (REQUIRED)
  3. docs/workflow.md — TDD policy, commit strategy (if exists)
  4. CLAUDE.md — architecture, Do/Don't
  5. .solo/pipelines/progress.md — running docs from previous iterations (if exists, pipeline-specific). Contains what was done in prior pipeline sessions: stages completed, commit SHAs, last output lines. Use this to avoid repeating completed work.

Do NOT read source code files at this stage. Only docs. Source files are loaded per-task in the execution loop (step 3 below).

Resumption

If a task is marked [~] in plan.md:

Resuming: {track title}
Last task: Task {X.Y}: {description} [in progress]

1. Continue from where we left off
2. Restart current task
3. Show progress summary first

Ask via AskUserQuestion, then proceed.

Context Engineering Rules

Read references/context-engineering.md for full rules on observation masking, attention positioning, and plan recitation. Key points:

  • Observation masking: large outputs (>50 lines) → write to scratch/, keep 5-10 line summary
  • Attention positioning: errors at START, history in MIDDLE, next steps at END
  • Plan recitation: re-read plan.md at START of each task to prevent drift

Task Execution Loop

Makefile convention: If Makefile exists in project root, always prefer make targets over raw commands. Use make test instead of pnpm test, make lint instead of pnpm lint, make build instead of pnpm build, etc. Run make help (or read Makefile) to discover available targets. If a make integration or similar target exists, use it for integration testing after pipeline-related tasks.

IMPORTANT — All-done check: Before entering the loop, scan plan.md for ANY - [ ] or - [~] tasks. If ALL tasks are [x] — skip the loop entirely and jump to Completion section below to run final verification and output <solo:done/>.

For each incomplete task in plan.md (marked [ ]), in order:

1. Find Next Task

Parse plan.md for first line matching - [ ] Task X.Y: (or - [~] Task X.Y: if resuming).

2. Start Task

  • Update plan.md: [ ] → [~] for current task.
  • Announce: "Starting Task X.Y: {description}"

3. Research (smart, before coding)

Do NOT grep the entire project or read all source files. Load only what this specific task needs.

If MCP available (preferred):

  1. Grep the project for the task keywords — exact names beat any index. If the name is uncertain: symbol_search("<words>", project="{name}"). Read only the top 2-3 files.
  2. session_search("{task keywords}") — check if you solved this before.
  3. Grep the imports of the files you'll modify, and Grep for their module name to find who depends on them.

If MCP unavailable (fallback):

  1. Read ONLY the files explicitly mentioned in the task description (file paths).
  2. Glob for the specific module directory the task targets (e.g., src/auth/**/*.ts), not the entire project.
  3. If the task doesn't mention files, use Grep with a narrow pattern on src/ or app/ — never **/*.

Never do: Grep "keyword" . across the whole project. This dumps hundreds of lines into context for no reason. Be surgical.

Platform-Specific Quality Tools

Read references/quality-tools.md for full commands per stack. Quick reference:

StackLintFormatType-checkPre-commit
Pythonuv run ruff check --fix .uv run ruff format .uv run ty check .uv run pre-commit run --all-files
JS/TSpnpm lint --fixpnpm formatpnpm tsc --noEmithusky + lint-staged
iOSswiftlint lint --strictswift-format—lefthook
Android./gradlew detekt./gradlew ktlintCheck—lefthook

Run after each task implementation, before git commit. If any fail, fix before proceeding.

4. Tests (policy from workflow.md)

Who owns the expectation decides the workflow. A red test proves something only when someone checked why it went red; when you write the test, run it and watch it fail, all it proves is that you ran it.

Agree the seam first: name the public interface this task's tests will observe, and confirm it with the user before writing a test. No test at an unconfirmed seam.

Design before code. Write down the data model, edge cases and contracts for the task before any code or test. In Böckeler's eval (TDD inside the agent loop, martinfowler.com, Aug 2026) this up-front design is what moved quality; a test-by-test loop let the first test fix the shape, and behaviour nobody wrote a test for was never built.

Strict — the user owns the test. Write the failing test at the agreed seam, show it to the user, and implement only after they confirm it tests the wanted behaviour. Then green, then refactor while tests stay green. Use this when the user asked for it in docs/workflow.md, or when the expectation comes from acceptance examples or a contract written before the code.

Moderate (default) — you own the test. No red-green ritual: implement and test together at the agreed seams, one behaviour at a time. The same eval found agent-run TDD cost 2.96x to 8.5x the tokens with no better mutation scores or design, and agents skipped or faked the red step. Test business logic and API routes; skip UI layout and config. Check the suite instead of the ritual: at the end of a phase, run the stack's mutation tool on the changed business logic if it is installed (mutmut, Stryker, cargo-mutants) and kill the surviving mutants that matter. With no mutation tool, break one line of the new logic by hand and confirm a test goes red, then restore it.

One behaviour → code + test → next (vertical slices). Writing all tests first then all implementation verifies imagined behaviour, not real.

Read references/tdd-seams.md before the first test of a feature — it covers where seams go, vertical vs horizontal slicing, and the four ways a test lies (implementation-coupled, side-channel, tautological, insensitive).

5. No new tests (policy "none", or a trivial task)

  • Implement the task directly.
  • Run existing tests to check nothing broke.

5.5. Integration Testing (CLI-First)

If the task touches core business logic (pipeline, algorithms, agent tools), run make integration (or the integration command from docs/workflow.md). The CLI exercises the same code paths as the UI without requiring a browser. If make integration fails, fix before committing.

5.6. Visual Verification (if browser/simulator/emulator available)

After implementation, run a quick visual smoke test if tools are available:

Web projects (Playwright MCP or browser tools): If you have Playwright MCP tools or browser tools available:

  1. Start the dev server if not already running (check stack YAML for dev_server.command)
  2. Navigate to the page affected by the current task
  3. Check the browser console for errors (hydration mismatches, uncaught exceptions, 404s)
  4. Take a screenshot to verify the visual output matches expectations
  5. If the task affects responsive layout, resize to mobile viewport (375px) and check

iOS projects (simulator): If instructed to use iOS Simulator in the pipeline prompt:

  1. Build for simulator: xcodebuild -scheme {Name} -sdk iphonesimulator build
  2. Install on booted simulator: xcrun simctl install booted {app-path}
  3. Launch and take screenshot: xcrun simctl io booted screenshot /tmp/sim-screenshot.png
  4. Check simulator logs: xcrun simctl spawn booted log stream --style compact --timeout 10

Android projects (emulator): If instructed to use Android Emulator in the pipeline prompt:

  1. Build debug APK: ./gradlew assembleDebug
  2. Install: adb install -r app/build/outputs/apk/debug/app-debug.apk
  3. Take screenshot: adb exec-out screencap -p > /tmp/emu-screenshot.png
  4. Check logcat: adb logcat '*:E' --format=time -d 2>&1 | tail -20

Graceful degradation: If browser/simulator/emulator tools are not available or fail — skip visual checks entirely. Visual testing is a bonus, never a blocker. Log that it was skipped and continue with the task.

6. Complete Task

Commit (following commit strategy):

git add {specific files changed}
git commit -m "<type>(<scope>): <description>"

Types: feat, fix, refactor, test, docs, chore, perf, style

Capture SHA after commit:

git rev-parse --short HEAD

SHA annotation in plan.md. After every task commit:

  1. Mark task done: [~] → [x]
  2. Append commit SHA inline: - [x] Task X.Y: description <!-- sha:abc1234 -->

Without a SHA, there's no traceability and no revert capability. If a task required multiple commits, record the last one.

7. Phase Completion Check

After each task, check if all tasks in current phase are [x].

If phase complete:

  1. SHA audit — scan all [x] tasks in this phase. If any are missing <!-- sha:... -->, capture their SHA now from git log and add it. Every [x] task MUST have a SHA.
  2. Run verification steps listed under ### Verification for the phase.
  3. Run full test suite.
  4. Run linter.
  5. Mark verification checkboxes in plan.md: - [ ] → - [x].
  6. Commit plan.md progress: git commit -m "chore(plan): complete phase {N}".
  7. Capture checkpoint SHA and append to phase heading in plan.md: ## Phase N: Title <!-- checkpoint:abc1234 -->.
  8. Report results and continue:
Phase {N} complete! <!-- checkpoint:abc1234 -->

  Tasks:  {M}/{M}
  Tests:  {pass/fail}
  Linter: {pass/fail}
  Verification:
    - [x] {check 1}
    - [x] {check 2}

  Revert this phase: git revert abc1234..HEAD

Proceed to the next phase automatically. No approval needed.

Error Handling

Test Failure

Tests failing after Task X.Y:
  {failure details}

1. Attempt to fix
2. Rollback task changes (git checkout)
3. Pause for manual intervention

Ask via AskUserQuestion. Do NOT automatically continue past failures.

Track Completion

When all phases and tasks are [x]:

1. Final Verification

  • Run local build — must pass before deploy:
    • Next.js: pnpm build
    • Python: uv build or uv run python -m py_compile src/**/*.py
    • Astro: pnpm build
    • Cloudflare: pnpm build
    • iOS: xcodebuild -scheme {Name} -sdk iphonesimulator build
    • Android: ./gradlew assembleDebug
  • Run full test suite.
  • Run linter + type-checker.
  • Visual smoke test (if tools available):
    • Web: start dev server, navigate to main page, check console for errors, take screenshot
    • iOS: build + install on simulator, launch, take screenshot, check logs
    • Android: build APK + install on emulator, launch, take screenshot, check logcat
    • Skip if tools unavailable — not a blocker for completion
  • Check acceptance criteria from spec.md:
    • For EACH - [ ] criterion: verify it's met (search code, run command if specified).
    • If criterion contains a runnable command (make task, cargo test, benchmark, passes N/N) → RUN IT. Do NOT skip as "unverifiable".
    • Update spec.md checkboxes: - [ ] → - [x] for verified criteria.
    • If any criterion FAILS → do NOT mark plan complete. Fix or note the failure.
    • Commit spec updates: git add docs/plan/*/spec.md && git commit -m "docs: update spec checkboxes (verified by build)"
  • Map every criterion to the test that proves it. Print one table in the summary, criterion → test (file::name) → last result. A criterion with no test is a red check, listed by name under UNPROVEN, even if the code looks right; a test that never ran in this session counts as no test. Criteria a test cannot reach (a store listing, a human sign-off) are listed as manual: <who checks it>, never silently ticked. Why: a checkbox ticked by reading code is the agent grading its own claim.

Build failure escalation

If the same build/test error occurs 3+ times in a row, STOP retrying. Instead:

  1. Output the error message clearly.
  2. Suggest a fix (e.g., cargo clean, rm -rf node_modules, disk space check).
  3. Output <solo:redo/> to let review handle it. Do NOT burn iterations on identical failures.

2. Update plan.md header

Change **Status:** [ ] Not Started → **Status:** [x] Complete at the top of plan.md.

3. Signal completion

Output pipeline signal ONLY if pipeline state directory (.solo/states/) exists:

<solo:done/>

Do NOT repeat the signal tag elsewhere in the response. One occurrence only.

4. Summary

Track complete: {title} ({trackId})

  Phases: {N}/{N}
  Tasks:  {M}/{M}
  Tests:  All passing

  Phase checkpoints:
    Phase 1: abc1234
    Phase 2: def5678
    Phase 3: ghi9012

  Revert entire track: git revert abc1234..HEAD

Next:
  /build {next-track-id}  — continue with next track
  /plan "next feature"    — plan something new

Reverting Work

SHA comments in plan.md enable surgical reverts:

Revert a single task:

# Find SHA from plan.md: - [x] Task 2.3: ... <!-- sha:abc1234 -->
git revert abc1234

Then update plan.md: [x] → [ ] for that task.

Revert an entire phase:

# Find checkpoint from phase heading: ## Phase 2: ... <!-- checkpoint:def5678 -->
# Find previous checkpoint: ## Phase 1: ... <!-- checkpoint:abc1234 -->
git revert abc1234..def5678

Then update plan.md: all tasks in that phase [x] → [ ].

Never use git reset --hard — always git revert to preserve history.

Progress Tracking (TodoWrite)

At the start of a build session, create a task list from plan.md so progress is visible:

  1. On session start: Read plan.md, find all incomplete tasks ([ ] and [~]).
  2. Create TaskCreate for each phase with its tasks as description.
  3. TaskUpdate as you work: in_progress when starting a task, completed when done.
  4. This gives the user (and pipeline) real-time visibility into progress.

Rationalizations Catalog

These thoughts mean STOP — you're about to cut corners:

ThoughtReality
"This is too simple to test"Simple code breaks too. Write the test.
"I'll add tests later"Tests written after pass immediately — they prove nothing.
"I already tested it manually"Manual tests don't persist. Automated tests do.
"The test framework isn't set up"Set it up. That's part of the task.
"This is just a config change"Config changes break builds. Verify.
"I'm confident this works"Confidence without evidence is guessing. Run the command.
"Let me just try changing X"Stop. Investigate root cause first.
"Tests are passing, ship it"Tests passing ≠ acceptance criteria met. Check spec.md.
"I'll fix the lint later"Fix it now. Tech debt compounds.
"It works on my machine"Run the build. Verify in the actual environment.

Red Flags — STOP immediately if you catch yourself thinking:

  • "I'll skip the test for this one" — No. Write it.
  • "Let me keep the old code as reference" — Delete means delete. You'll adapt it instead of writing fresh. No exceptions.
  • "I'm following the spirit, not the letter" — Violating the letter of the rules IS violating the spirit.
  • "This is different because..." — It's not. Follow the process.
  • "Just one more quick change before committing" — Commit what you have. Scope creep kills tracks.
  • "The plan says X but Y would be better" — Follow the plan. If it's truly wrong, update plan.md FIRST, then implement.

Foundational principle: If you wouldn't mass-delete production data without checking, don't skip a test without evidence it's unnecessary.

Critical Rules

  1. Run phase checkpoints — verify tests + linter pass before moving to next phase.
  2. STOP on failure — do not continue past test failures or errors.
  3. Keep plan.md updated — task status must reflect actual progress at all times.
  4. Commit after each task — atomic commits with conventional format.
  5. Research before coding — 30 seconds of search saves 30 minutes of reimplementation.
  6. One task at a time — finish current task before starting next.
  7. Keep test output concise — when running tests, pipe through head -50 or use --reporter=dot / -q flag. Thousands of test lines pollute context. Only show failures in detail. If output is large, use observation masking (write to scratch/, keep summary).
  8. Verify before claiming done — run the actual command, read the full output, confirm success BEFORE marking a task complete. Never say "should work now".
  9. Recite the plan — re-read plan.md at the start of each task. Don't rely on memory of what comes next.
  10. Three rounds, then a human — when a reviewer (agent, /review, /code-review) and the author keep answering the same finding, stop after the third round on it: write the finding, both positions and what would settle it, and hand the decision to the human. A fourth round costs a context and decides nothing.

Common Issues

"No plans found"

Cause: No plan.md exists in docs/plan/. Fix: Run /plan "your feature" first to create a track.

Tests failing after task

Cause: Implementation broke existing functionality. Fix: Use the error handling flow — attempt fix, rollback if needed, pause for user input. Never skip failing tests.

Phase checkpoint failed

Cause: Tests or linter failed at phase boundary. Fix: Fix failures before proceeding. Re-run verification for that phase.

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.