| name | impl |
| description | Implement a change with a pipeline scaled to its complexity. Single entry point for all code changes — auto-detects effort or accepts small|mid|tuff override. |
Implement a Change
Single entry point for all code changes. Select effort explicitly or let Claude auto-detect from the request.
Arguments
$0 -- (optional) Effort tier override: small, mid, or tuff. Omit to auto-detect.
- All remaining words form the change description.
Effort Tiers
| Tier | Scope | Pipeline |
|---|
small | 1–2 files, modifying existing logic, no new abstractions | Read → 1 question max → implement → just check |
mid | Existing pattern, chain of /add-* skills, 3–8 files | Quick research → chain /add-* skills in order → just check |
tuff | New domain, new abstractions, or no prior pattern to follow | Full pipeline: questions → research → plan → plan-review → user-approval → implement → verifier |
Auto-Detection Rules
Choose tuff if any of:
- New entity or Alembic migration is required
- New port, adapter, repo, or event (new abstraction, not adding to an existing one)
- No existing
/add-* skill covers the task
- Request says "new domain", "from scratch", "design", or "architect"
Choose mid if any of:
- Existing entity is being extended with new behavior (new handler, new endpoint, new event)
- One or more
/add-* skills directly match the task
- 3–8 files touched, all following established patterns
Choose small if all of:
- Scoped to 1–2 existing files
- No new files, no new abstractions
- Modifying a value, condition, field, or small logic block
When ambiguous between mid and tuff, choose mid. When ambiguous between small and mid, choose mid.
small pipeline
- Read the file(s) involved.
- If something is genuinely ambiguous, ask at most one clarifying question. Otherwise proceed immediately.
- Make the change.
- Run
just check. Fix any issues before reporting done.
No planning. No agents. No approval step.
mid pipeline
- Identify which
/add-* skills apply (e.g., /add-handler + /add-endpoint + /add-test).
- Run
/research <domain> to confirm existing patterns — never invent structure when an example already exists.
- Execute the identified
/add-* skills in dependency order: entity before repo, repo before handler, handler before controller, then tests.
- Run
just check. Fix all issues.
- Delegate to implementation-verifier agent — fix any drift before reporting the task done.
No plan file. No approval step.
tuff pipeline
Follow the full workflow:
- Ask clarifying questions: business rules, scope, data model, side effects, error cases, naming
- Run
/research <topic> — read all relevant existing code thoroughly
- Run
/plan <feature> — produce the structured plan with every file, field, method signature
- Confirm (plan mode only): if the session is running in plan mode, call
ExitPlanMode to present the plan and wait for approval. Outside plan mode, skip — proceed directly to implementation.
- Implement in dependency order — run
just check after each major step; do not deviate from the plan without asking
- Delegate to implementation-verifier agent — fix any drift before reporting the task done
- Delegate to architecture-reviewer agent — fix any layer violations or pattern drift before reporting done
Run the plan-reviewer agent only when the user explicitly asks for a plan review.