autopilot-execute
Skill0 starsChanged 7 months agoMITView on GitHub
Unattended execution of the next plan item(s). Reads project context and .plan/ files, executes work autonomously without human confirmation, verifies the build, updates all plan files, and exits. Designed to be invoked repeatedly in successive sessions.
In this repository8 other files it gives its agents.
- close-planSkill
- deep-modules-auditSkill
- docs-auditSkill
- execute-phaseSkill
- lint-fixSkill
- living-design-systemSkill
- phased-planSkill
- responsive-auditSkill
The file
Show its rules and text
We split this skill into the rules it gives an agent
40 rules. About 0 are common, 31 are distinctive, and 9 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈31
Distinctive · estimate
This author's own choices. Worth a look.
9
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 10
Do not ask questions. Do not wait for confirmation.
- Very specific: line 14
Read autopilot.config in the project root.
- Very specific: line 18
If autopilot.config does not exist, detect the project's build system:
- Very specific: line 19
If bun.lockb or bun.lock exists → bun run typecheck
- Very specific: line 20
If pnpm-lock.yaml exists → pnpm run typecheck
- Very specific: line 23
If no typecheck script in package.json, try npm run build
- Distinctive: line 27
Read these files in this order:
- Distinctive: line 34
If CLAUDE.md references additional files to read (AGENTS.md, CONVENTIONS.md, etc.), read those too — but be fast.
- Distinctive: line 44
Look at the next unchecked item in the current phase.
- Distinctive: line 47
If the item is small (create a helper function, add a config value, write a type definition) → combine it with the next 1-2 unchecked items …
- Distinctive: line 48
If the item is medium (build a module, implement a feature, create a component) → do just that item
- Distinctive: line 49
If the item is large (complex multi-file feature, integration work) → do just that item, potentially noting sub-progress
- Distinctive: line 50
If it's a verification item (e.g., "verify all endpoints work", "run full test suite") → run the checks, report results
- Distinctive: line 51
If it's a CREATE TASKS item (meta-task) → read the data source, generate the plan items, insert them, then EXIT immediately
- Distinctive: line 59
NEVER pick items from the next phase.
- Distinctive: line 60
If this is the LAST item in the phase, complete it, verify exit criteria from PLAN.md, then update the plan header.
- Distinctive: line 64
Do the work. Be thorough.
- Distinctive: line 67
Read CLAUDE.md for project conventions, key file locations, and commands
- Distinctive: line 68
Check MEMORY.md for architectural patterns and decisions
- Distinctive: line 69
Check CONVENTIONS.md for build rules with ✅/❌ examples
- Distinctive: line 70
Always read existing source files before modifying — understand full context
- Distinctive: line 71
Follow patterns established in previous sessions (documented in MEMORY.md)
- Distinctive: line 72
When building something analogous to a reference repo pattern, read the reference first
- Distinctive: line 76
If you discover something unexpected, make a judgment call and log it in DRIFT.md
- Distinctive: line 77
If a file you need doesn't exist, note it and skip gracefully
- Distinctive: line 78
Follow the architectural patterns from MEMORY.md and CONVENTIONS.md
- Very specific: line 87
Run the build command from autopilot.config to confirm nothing is broken:
- Distinctive: line 96
Read the error output carefully
- Distinctive: line 98
Run the build again after each fix
- Distinctive: line 99
If still failing after 2 attempts:
- Distinctive: line 100
Leave the item unchecked in PLAN.md
- Very specific: line 101
Add a note to the item: - [ ] N.X — {Description} ⚠️ BUILD FAILED: {brief error}
- Distinctive: line 103
Update MEMORY.md with the gotcha
- Very specific: line 111
Check off completed items: - [ ] → - [x]
- Very specific: line 114
If phase is complete: update Current Phase: to next phase, Status: to note completion
- Distinctive: line 115
If this was the last item in the phase, verify exit criteria and note result
- Distinctive: line 118
Add any new decisions to the Key Decisions table (increment the # counter)
- Distinctive: line 119
Add any new gotchas to Gotchas & Warnings
- Distinctive: line 120
Update Architecture Notes or Technical Context if anything changed
- Distinctive: line 123
If anything deviated from the plan or previous patterns, log it
≈ “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.
Files
Autopilot Execute
You are running unattended. No human is present. Be fully autonomous. Do not ask questions. Do not wait for confirmation. Make decisions and log them.
Step 0 — Read Project Config
Read autopilot.config in the project root. It defines:
BUILD_CMD— the command to verify the project builds (may chain multiple checks)PROJECT_DIR— the directory containing the project source (.for monorepos)
If autopilot.config does not exist, detect the project's build system:
- If
bun.lockborbun.lockexists →bun run typecheck - If
pnpm-lock.yamlexists →pnpm run typecheck - If
yarn.lockexists →yarn typecheck - Otherwise look for
package.json→npm run typecheck - If no
typecheckscript in package.json, trynpm run build
Step 1 — Orient (mandatory, fast)
Read these files in this order:
CLAUDE.md— Project identity, current state, commands, conventions.plan/PLAN.md— Find current phase and next uncompleted item.plan/MEMORY.md— Absorb all shared context, decisions, gotchas.plan/DRIFT.md— Check for spec changes
If CLAUDE.md references additional files to read (AGENTS.md, CONVENTIONS.md, etc.), read those too — but be fast. Orientation should take seconds, not minutes.
Then print exactly one status line:
📍 Phase {N} — {Phase Name} | 📋 Next: {N.X} — {Description} | 📝 {count} items left in phaseStep 2 — Scope
Look at the next unchecked item in the current phase.
Sizing judgment:
- If the item is small (create a helper function, add a config value, write a type definition) → combine it with the next 1-2 unchecked items in the same phase
- If the item is medium (build a module, implement a feature, create a component) → do just that item
- If the item is large (complex multi-file feature, integration work) → do just that item, potentially noting sub-progress
- If it's a verification item (e.g., "verify all endpoints work", "run full test suite") → run the checks, report results
- If it's a CREATE TASKS item (meta-task) → read the data source, generate the plan items, insert them, then EXIT immediately
Announce your scope:
🎯 This iteration: {N.X} — {Description} [and {N.Y} if combining]Rules:
- NEVER pick items from the next phase. If all items in the current phase are checked, update the PLAN.md header to mark the phase complete and EXIT immediately. The next session handles phase boundaries.
- If this is the LAST item in the phase, complete it, verify exit criteria from PLAN.md, then update the plan header.
Step 3 — Execute
Do the work. Be thorough.
Reference material:
- Read
CLAUDE.mdfor project conventions, key file locations, and commands - Check MEMORY.md for architectural patterns and decisions
- Check CONVENTIONS.md for build rules with ✅/❌ examples
- Always read existing source files before modifying — understand full context
- Follow patterns established in previous sessions (documented in MEMORY.md)
- When building something analogous to a reference repo pattern, read the reference first
While executing:
- Complete each item fully before moving to the next
- If you discover something unexpected, make a judgment call and log it in DRIFT.md
- If a file you need doesn't exist, note it and skip gracefully
- Follow the architectural patterns from MEMORY.md and CONVENTIONS.md
After each item:
✅ {N.X} — {Description} — DoneStep 4 — Verify
Run the build command from autopilot.config to confirm nothing is broken:
# Execute BUILD_CMD from autopilot.configIf build succeeds: Proceed to Step 5.
If build fails:
- Read the error output carefully
- Attempt to fix the issue (you have 2 fix attempts)
- Run the build again after each fix
- If still failing after 2 attempts:
- Leave the item unchecked in PLAN.md
- Add a note to the item:
- [ ] N.X — {Description} ⚠️ BUILD FAILED: {brief error} - Log the failure details in SESSION-LOG.md
- Update MEMORY.md with the gotcha
- EXIT. The next session will detect you're stuck.
Step 5 — Close (mandatory, NEVER skip)
Even if something went wrong, you MUST update the plan files before exiting.
Update .plan/PLAN.md:
- Check off completed items:
- [ ]→- [x] - Update header:
Last Session:→ increment session number and set today's date- If phase is complete: update
Current Phase:to next phase,Status:to note completion
- If this was the last item in the phase, verify exit criteria and note result
Update .plan/MEMORY.md:
- Add any new decisions to the Key Decisions table (increment the # counter)
- Add any new gotchas to Gotchas & Warnings
- Update Architecture Notes or Technical Context if anything changed
Update .plan/DRIFT.md:
- If anything deviated from the plan or previous patterns, log it
Append to .plan/SESSION-LOG.md:
## Session {N} — {Item Description} ({date})
**What happened:**
- {Brief summary of work done}
**Items completed:**
- [x] {N.X} — {Description}
**Issues encountered:**
- {Any blockers, surprises, or decisions made — or "None"}
**Handoff to next session:**
- Next: {N.Y} — {Next unchecked item description}
- {Any critical context for the next iteration}
---Update .agents/TENSIONS.md:
- Append any friction noticed during execution (format:
- {date}: [{area}] {observation})
Hard Rules
- NEVER skip Step 1 or Step 5 — The plan files are the shared brain. Corrupt them and everything breaks.
- NEVER cross phase boundaries — If you finish the last item in a phase, mark the phase complete and EXIT. Do not start the next phase.
- NEVER ask questions — You are unattended. Make the best decision you can. Log it in MEMORY.md or DRIFT.md.
- ALWAYS verify the build — The BUILD_CMD from autopilot.config must pass before you update plan files with checked items.
- EXIT cleanly — After Step 5, your job is done. The next session will re-invoke you if there's more work.
- Be conservative with scope — It's better to complete 1 item cleanly than attempt 3 and leave a mess. The next session will continue.
- Log everything — Decisions in MEMORY.md, deviations in DRIFT.md, friction in TENSIONS.md. The next iteration has zero memory of this one.
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
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.

