Executing Plans

Prerequisites

Execution Mode Selection (MANDATORY — before anything else)

Read 03-implementation-plan.md. Count tasks by status and estimate per-task file impact from the Files / Impact field.

Announce: "Mode: executing-plans — N pending tasks, max M files per task." or "Switching to subagent-driven-development — reason: [N tasks / task X touches M files]."

Why this matters: executing-plans runs entirely in the main context. As tasks accumulate, context grows and output quality drops. subagent-driven-development gives each task a fresh context, preventing contamination across tasks. For small plans (≤8 tasks), executing-plans avoids the per-subagent token overhead of cold-starting multiple agents.

Overview

Load plan artifacts, execute tasks in batches, and stop immediately when requirements drift.

Announce at start: "I'm using the executing-plans skill to implement this plan."

Step 0: Load Project Rules

1. Search for rules files in this order:
   • rules/**/*.md
   • .claude/rules/**/*.md
   • docs/rules/**/*.md
   • CLAUDE.md, AGENTS.md (project root)
2. Read all found files.
3. Extract two separate checklists from the loaded content:
   • MUST_CHECK: rules under ## Mandatory sections — objective, verifiable, block task completion if violated
   • SHOULD_CHECK: rules under ## Recommended sections — flag in report but do not block completion
   • SKIP: principles without a clear pass/fail test (e.g., "prefer clarity", "keep it simple")
4. Enforcement:
   • MUST_CHECK violation → task cannot be marked completed; fix before proceeding
   • SHOULD_CHECK violation → record in batch report as flagged item; task may still complete

5. Output a compact summary:

   "Loaded N rules files. Extracted X MUST_CHECK items, Y SHOULD_CHECK items."

   If no rules files are found: note it and continue without a checklist.

This checklist is used in every task DoD and the Post-Batch Rules Sweep.

Step 1: Load And Review Plan Set

Read these files in the same feature folder:

• 01-requirements.md
• 02-design.md
• 03-implementation-plan.md
• 04-verification.md
• 05-change-log.md
• 99-mistake-log.md
• 06-research-notes.md (if present)

Before coding:

1. Confirm every active task maps to REQ-###.
2. Confirm all open CR-### and DR-### items are represented in tasks.
3. Raise gaps before execution.

Step 2: Execute Batch

Default batch size: first 3 pending tasks.

For each task:

1. Check Depends on — all listed tasks must be completed before starting. If not, skip and report blocked.
2. Mark task in_progress.
3. Follow plan steps exactly based on task type:
   • code-task: apply agent-workflow:test-driven-development (write failing test → verify failure → implement → verify pass)
   • migration-task: write forward + rollback scripts, verify both
   • config-task: verify affected environments, document change
4. Before marking complete, verify every DoD item from 03-implementation-plan.md:
   • All tests pass; coverage meets targets
   • No TODO / debug / temporary code left
   • Naming follows project rules
   • Error handling complete
   • All callers in Impact handled
   • 04-verification.md evidence updated
   • Commit message: type(scope): one-line summary
   • Rules compliance: check changed files against AUTO_CHECK items from Step 0 checklist; fix violations before marking complete
5. Mark task completed.

Unexpected Situation Protocol (MANDATORY)

When something unexpected is encountered during task execution (design assumption wrong, file missing, interface different from spec):

Decision rule:

Never make silent design decisions that affect scope, contracts, or other tasks.

Plan Drift Gate (MANDATORY)

If a new request, adjustment, or optimization appears during execution:

1. Stop implementation immediately.
2. If scope change: add CR-### in 05-change-log.md, update 01-requirements.md. If design change only: add DR-### in 05-change-log.md, update 02-design.md.
3. Update 03-implementation-plan.md tasks/statuses.
4. Update impacted checklist items in 04-verification.md.
5. Resume execution only after 2-5 are complete.

No coding is allowed for unmapped requirements.

Post-Batch Rules Sweep (MANDATORY before Step 3)

After all tasks in the batch are complete, scan every file changed in this batch against the AUTO_CHECK checklist from Step 0.

Decision table per violation:

Do not ask the user about AUTO_FIX items individually — fix them silently and list what was fixed in the batch report.

If the Step 0 checklist is empty (no rules files found), skip this sweep.

Step 3: Report Batch

After each batch:

• Show completed tasks
• Show verification evidence summary
• List AUTO_FIX rules corrections applied (if any)
• List NEEDS_DECISION rules findings as a numbered batch (if any) — wait for user response before next batch
• List any open CR-###
• Re-state the AUTO_CHECK checklist (compact form) so rules stay in context near-top for the next batch
• Say: "Ready for feedback."

Step 4: Continue

Based on feedback:

• Apply plan updates if needed
• Execute next batch
• Repeat until all required tasks are complete

Step 5: Completion Gate

Before branch-finishing workflow:

• unmapped_req = 0
• open_cr = 0
• pending_recheck = 0

Then announce and invoke:

• agent-workflow:finishing-a-development-branch

When To Stop And Ask For Help

Stop and escalate when:

• Plan gaps block execution
• Verification repeatedly fails
• Dependencies/instructions are unclear
• Requirement drift cannot be safely mapped

Integration

• REQUIRED: agent-workflow:writing-plans
• REQUIRED: agent-workflow:finishing-a-development-branch