Execute Plan

/execute-plan

Implement the chunk at the path provided by the user. Every plan produced by /deep-plan is a suite — a directory containing index.md and one or more chunk files. Even a single-chunk plan is a suite (the chunk lives at /plan.md).

Usage

/execute-plan

Examples (paths shown use the default plans directory; substitute the configured plansDir if different):

• Single-chunk suite: /execute-plan .claude/plans/2026-04-30-153022-add-rate-limiter/plan.md
• One chunk in a multi-chunk suite: /execute-plan .claude/plans/2026-04-30-153022-mqtt-sync/listener.md

If the user passes the suite directory or index.md, ask which chunk they want to run.

Plans directory

Resolve the plans directory before any path-touching action: read .claude/plans-config.json if present and use its plansDir; otherwise default to .claude/plans. If the config file is missing or malformed, silently fall back to the default. The user-supplied chunk path is authoritative — this resolution only matters when this skill needs to interpret or print paths that aren't given by the user.

Asking questions

This skill prefers the AskUserQuestion tool for interactive prompts. If AskUserQuestion is not available (older Claude Code versions, restricted environments), fall back to plain text: print the question, list the options as a numbered list with the recommended option marked (Recommended), and wait for the user's reply (a number or the option label). The skill proceeds normally in either mode — every call site below that says "call AskUserQuestion" follows this fallback rule.

Phases

Phase 1 — Read

Read the chunk file as your first action. Do not explore or implement before reading it. If the path is missing, the file doesn't exist, or the file is malformed, stop and tell the user.

Required sections in every chunk file:

• ## Context
• ## Approach
• ## Files Changed
• ## Implementation Steps
• ## Risks & Edge Cases
• ## Acceptance Criteria

Required frontmatter (a block of > **Field:** value lines near the top):

• > **Status:**
• > **Depends on:**
• > **Part of suite:**

If any required section or frontmatter line is missing, stop and tell the user the chunk is malformed.

Also read CLAUDE.md at the project root for project conventions.

Phase 1.5 — Suite dependency check

Every chunk belongs to a suite, including single-chunk suites. Always run this phase.

1. Locate the suite index using the > **Part of suite:** line. If unset, fall back to index.md in the same directory as the chunk.
2. Read the index file.
3. Find this chunk's entry in the index.
4. Check the chunk's current Status: field in the index:

• If done: stop. Tell the user this chunk is already marked done. Ask via plain text whether they want to re-run it (which requires manually editing the index back to pending first).
• If in-progress: warn the user that the chunk is already in progress (possibly from a previous failed run). Ask via AskUserQuestion: "Continue (resume / re-run), or stop?" with options "Continue", "Stop". If continue, proceed.
• If blocked: stop. Tell the user the chunk is marked blocked and needs human resolution before it can run.
• If pending: proceed.

1. Check this chunk's Depends on: list. For each dependency, look up its status in the index.

• If any dependency is not done, stop and print:

``` 🚫 Cannot execute : depends on chunks that are not done yet.

• :
• :

Run those chunks first, or use /plan-tracker to see suite status. ```

• End your turn. Do not proceed to implementation.

1. If all dependencies are done (or the chunk has no dependencies), mark this chunk's status as in-progress in both the index file and the chunk file's frontmatter. This is the only automatic status update — it does not claim correctness, only that work has started.

For single-chunk suites this phase is trivially satisfied (Depends on: is (none)), but the status update from pending → in-progress still happens.

Phase 2 — Confirm

Print:

📄 Chunk loaded: 
   Suite: 
   Status: in-progress (was pending; updated on run start)
   Approach: 

   Files to change (in order):
   1.  — 
   2. ...

   Acceptance criteria: 

Then ask in plain text: "Plan looks current and correct? Reply 'go' to proceed, or describe what's stale or wrong."

Stop and wait. Do not proceed to Phase 3 until the user explicitly approves.

If the user reports that the plan is stale (e.g. "the file structure has changed since this was written"), stop. Tell the user the plan should be re-generated with /deep-plan. Do not patch a stale plan in execution mode. Revert the chunk's status from in-progress back to pending in both the index and the chunk file before stopping.

Phase 3 — Implement

Implement the steps in the order specified by the chunk. Do not reorder, skip, or merge steps without explicit user approval.

For each step:

1. Make the change.
2. Verify it builds / parses / passes locally if applicable.
3. Continue to the next step.

If a step is impossible or wrong-as-specified (the plan calls for a file that doesn't exist, specifies an interface that conflicts with reality, etc.), stop and ask the user. Do not silently work around plan specifications.

The chunk provides interfaces, contracts, and intent — not implementation bodies. You are responsible for designing implementations that satisfy the contracts. Use the project's existing patterns (from CLAUDE.md and the codebase) as a guide.

Phase 4 — Verify

After all steps are complete, walk through the acceptance criteria checklist from the chunk. For each criterion:

• Verify it (run the relevant test, check the relevant behavior)
• Mark ✅ or ❌
• For ❌, briefly state why it can't be satisfied as-specified

Phase 5 — Report

Print a final report:

Implementation report

Acceptance criteria:
  ✅ 
  ✅ 
  ❌  — 

Files changed:
  - 

Notes:
  

If any criteria are ❌, do not claim the work is complete.

Phase 6 — Status prompt

Always prompt the user for the final status — never auto-update from in-progress to done.

Call AskUserQuestion:

• Question: "How should this chunk be marked?"
• Options (single-select):

1. "Mark done"
2. "Mark blocked"
3. "Leave as in-progress"
4. "Discard run (revert to pending)"

Update the chunk's status in both the suite index file and the chunk file's frontmatter:

• "Mark done" → set status to done in both files. Print: "✅ Chunk marked done in suite index."
• "Mark blocked" → set status to blocked in both files. Print: "⚠️ Chunk marked blocked. Other chunks depending on this one cannot run until resolved."
• "Leave as in-progress" → no change (already in-progress). Print: "ℹ️ Status left as in-progress."
• "Discard run" → set status back to pending in both files. Print: "↩️ Status reverted to pending. (Code changes were NOT reverted — use git to undo if needed.)"

End your turn after printing the acknowledgment.

Constraints

• Read the chunk first, before any other action.
• Always check suite dependencies — even for single-chunk suites (the status transition still applies).
• Refuse to run if dependencies aren't done.
• Implement steps in the order specified.
• Don't deviate from the plan silently — ask the user when reality and the plan diverge.
• Don't claim completion when acceptance criteria are not met.
• Never auto-mark a chunk as done. Always prompt the user.
• The only automatic status update is pending → in-progress at run start.
• Status must always be kept in sync between the index and the chunk file frontmatter.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: vpaivag
• Source: vpaivag/skills
• License: MIT

Install and usage instructions live in the source repository linked above.