lnrd390 Execution Plan
Create a high-confidence implementation plan for:
$ARGUMENTS
Do not implement application code. Do not commit or push.
Your consumer is a fresh implementation agent that should spend as few tokens as practical rediscovering architecture, paths, and requirements.
Planning protocol
- Read
${CLAUDE_SKILL_DIR}/PLAN-SCHEMA.mdbefore finalizing the plan. - Reuse existing project context before broad exploration: applicable
CLAUDE.md, rules, relevant skills/memory, project docs, and.claude/context-system/when present. Do not assume the Context Toolkit exists. - Treat persisted context as routing/decision evidence, not unquestionable truth. Verify task-critical facts against current source.
- Spend discovery progressively: exact known paths/symbols first, then targeted search, then bounded module exploration. Broad scans are a last resort.
- For large independent unknowns, use isolated Explore/Plan subagents when available and require terse findings with exact paths/symbols. Avoid redundant agents exploring the same area.
- Convert the request into explicit requirements, derived requirements, non-goals, constraints, compatibility expectations, and decisions.
- Build a verified change map. Existing paths must exist; new paths must be explicitly
CREATE; deletions must be verified. Prefer symbols over line numbers. - Split into the minimum number of independently valid increments. A small change should remain one increment.
- Each increment must be executable by a fresh agent using mostly the listed files/symbols. Include behavior, decisions/constraints, validation, and done criteria. Reference existing patterns by path+symbol rather than copying code.
- Optimize total planning + execution tokens, not plan length in isolation.
Saving the plan
Prefer the project's configured plansDirectory if discoverable. When project-local plans are configured as .claude/plans, save there. Otherwise save to .claude/plans/ only when doing so is consistent with the project; if not, return the plan and report where Claude Code currently stores plans.
Use a short kebab-case filename. Add <!-- cwt-plan:v1 --> near the top.
Finish by reporting only the plan path, number of increments, and any unresolved discovery that would materially affect execution.
Reference
The complete schema is bundled at ${CLAUDE_SKILL_DIR}/PLAN-SCHEMA.md.