| model | sonnet |
| name | my-implement |
| description | Execute an approved plan phase by phase. Dispatches one implementation-executor at a time, enforces RED -> GREEN -> VALIDATE, re-verifies results, and stops on repeated failure. |
Implement Plan
Orchestrate an approved implementation plan. The executor writes code; this skill slices phases, dispatches one executor at a time, re-verifies, updates the plan, and owns loop detection.
Load Rules
Read these first:
~/.claude/rules/tdd-phase.md
~/.claude/rules/subagent-contract.md
~/.claude/rules/loop-detection.md
~/.claude/rules/no-outward-actions.md
~/.claude/rules/context-checkpoint.md
If running through Codex, use the same files under ~/.agents/rules/.
For unusual or ambiguous implementation runs, read references/protocol-index.md.
Flow
- Resolve the plan path from
$ARGUMENTS; otherwise list plans in ~/.claude/thoughts/shared/plans/ and ask which one.
- Read the plan fully. Resume from the first unchecked phase.
- Create one todo per phase.
- For each phase, build a compact slice for
implementation-executor:
- phase overview
- RED tests
- GREEN changes
- allowed paths
- success criteria
- verification commands
- architectural constraints
- relevant gotchas
- Dispatch exactly one executor.
- Re-run the phase success criteria yourself and read the diff.
- If the phase conforms, mark its checklist items complete in the plan and continue.
- If it fails, apply the loop-detection rule. Retry only with new information.
- After all phases, run the plan’s full validation strategy and mark the plan implemented.
Stop Conditions
Stop instead of dispatching when:
- A phase has no RED tests or no success criteria.
- The phase is too broad to fit a compact slice.
- Required files or APIs differ from the plan in a major way.
- The same failure reaches the loop-detection limit.
- A fix requires an outward action or product/scope decision.
Output
Summarize phases completed, deviations, tests run, failed checks if any, files changed, and whether /my-validate should run next.