swingerman/engineer/refine
Claude Code and agent skill by swingerman/engineer
- community
Checking that the SKILL.md file is still in the repository.
Use after a feature's code is implemented and passing, to clean up the changed code before verification. Triggers — "/engineer.refine", "refine this code", "clean up the feature", "refactor what we built".
Install
npx vskill@latest install swingerman/engineer/refineSKILL.md
refine
Charter-bound clean-up of a feature's changed code — Checkpoint 6. Modeled on Claude Code's stock /simplify (a three-subagent parallel review) with two layers the stock skill lacks: charter validation of every proposal, and graceful breaking changes.
The behavior contract (ACs + specs) is sacred — a refactor that breaks it is a category error. New shapes are introduced alongside old, never instead of.
When to use
Checkpoint 6, after atdd:atdd-team produces passing code, before crap-analyzer (Checkpoint 7).
Verification independence: refine runs on a non-implementer agent (agent_id ≠ CP5's; enforced by dae_handoff.py gate()). The three review subagents (Step 2) are fresh regardless, so independence is satisfied by construction.
Not for: changing behavior (feature-edit); risk analysis (crap-analyzer); code that isn't implemented/green yet.
Workflow
Step 0 — Entry gate. Before starting, verify the prior checkpoint is complete: run ${CLAUDE_PLUGIN_ROOT}/scripts/dae_handoff.py <feature-dir> --through 5. On a non-zero exit, stop and surface the gap to the human — do not proceed.
Verify branch hygiene: run ${CLAUDE_PLUGIN_ROOT}/scripts/dae_branch.py <feature-dir>. On a non-zero exit, stop and surface the message to the human — switch branches and re-invoke. The check honors the git.manual: true manifest opt-out.
After the gate passes, show the pipeline breadcrumb: run ${CLAUDE_PLUGIN_ROOT}/scripts/dae_progress.py <feature-dir> and present its output to the human — it shows where this checkpoint sits in the DAE pipeline. The breadcrumb is advisory: a non-zero exit or a missing progress.md never blocks the skill. Then create one TodoWrite todo per workflow step below. See ${CLAUDE_PLUGIN_ROOT}/references/progress-indicator.md.
- Resolve + scope — resolve the methodology root + manifest via
${CLAUDE_PLUGIN_ROOT}/scripts/dae_resolve.py(seereferences/resolving.md); locate the feature. Scope = the feature branch's changed code (git diffagainst the branch point). Loadfeature.md,acs.md,spec.md,CHARTER.md. - Dispatch three parallel review subagents: three Agent-tool calls in one message, each
subagent_type: engineer:reviewer(or the project override; see${CLAUDE_PLUGIN_ROOT}/references/host-capabilities.md, "Role agents"), each over the same changed code and each with its lens named in the brief. The agent is read-only and caps its reply, so three results fit in context. Before dispatching, run${CLAUDE_PLUGIN_ROOT}/scripts/dae_dup.py <methodology-root>to get deterministic project-wide duplicate findings; pass the JSON result into the Reuse subagent's context as a tool input (the subagent treats it as one signal alongside its own LLM judgment, and produces a unified set of findings — not a separate list). Ifdae_dup.pyreturns anything other thanstatus: ok—status: unavailable(the backend tool isn't installed),status: skipped(manifestduplication.skip: true),status: unsupported(the configured backend isn't implemented in this version), orstatus: error(scan failed) — surface that status in the Reuse subagent's prompt and proceed with its existing LLM-only judgment.
- Reuse — duplication (LLM +
dae_dup.pyfindings), reinvented wheels, dead code, missed existing utilities. LSP-first lookup — at dispatch time, check whether an LSP MCP capability is available; pass that detection result and the LSP-first preference into the Reuse subagent's prompt so it uses find-references / workspace-symbols when available, with graceful fallback to grep + Read. See${CLAUDE_PLUGIN_ROOT}/references/code-lookup.md. - Quality — clarity, structure, naming, incidental complexity, maintainability
- Efficiency — redundant computation, repeated lookups, visible performance smells
- Consolidate — merge findings; dedup overlaps (note multi-lens hits as strong signals); keep genuine conflicts as alternatives for the human.
- Charter filter — check every proposal against
CHARTER.md(architecture, conventions, methodology, the ACs+specs contract). Charter-violating proposals are rejected internally — never shown. This filter is the DAE-specific layer the stock skill lacks. - Classify breaking changes — consumer-facing (something outside the blast radius depends on the old shape) → graceful path: deprecation-marked forwarding shim + migration note, new shape alongside old. Internal-only → hard change is fine.
- Present; human picks — show charter-compliant proposals (change, why, blast radius, churn). The human selects which are worth the churn.
- Apply — apply selected; install graceful paths where needed; re-run both test streams. Before running the test streams, ensure required infra is up: read
manifest.yml'sinfra:section, then call${CLAUDE_PLUGIN_ROOT}/scripts/dae_infra.py ensure <names>for each declared dependency the tests need. On astart-failedstructured failure → stop and surface the diagnosis. On a missing manifest declaration for required infra → stop with the "declare in manifest" message. Any failure → revert that proposal and report. - Handoff — emit a summary.
Prose pass (before Step 8). Run ${CLAUDE_PLUGIN_ROOT}/scripts/dae_prose.py --changed <feature-dir> and fix every hit. Then reread the feature's changed .md artifacts against ${CLAUDE_PLUGIN_ROOT}/references/plain-prose.md: restate each sentence as a fact, number or instruction, or delete it. Never change what an AC or spec requires, only how it is worded.
If a charter rule itself blocks a genuinely better design, surface it in the handoff (→ feature-edit / charter amendment). refine does not amend the charter.
Handoff
Emit per ${CLAUDE_PLUGIN_ROOT}/references/handoff-summary.md. agent_id must differ from the implementer if checkpoint 6 is independence-gated. checkpoint: 6; recommended_next: "/engineer.arch-check (CP7 Light Verify; it runs crap-analyzer too)".
The handoff MUST include the exit_criteria block asserting each of Checkpoint 6's exit criteria (Foundation Design Section 8) with verified_by, met, and evidence. For verified_by: tool criteria, the evidence MUST be the tool's actual output. The checkpoint is marked done only when every criterion is met.
Before stopping, apply the dispatch rule — see ${CLAUDE_PLUGIN_ROOT}/references/handoff-dispatch.md. CP7 verify needs a fresh agent (charter §6). Dispatch subagent_type: engineer:verifier (or the project override), which cannot edit code and writes the CP7 handoff. If ${CLAUDE_PLUGIN_ROOT}/scripts/dae_diverse.py <feature-dir> says "diverse": true, also dispatch a second verifier on the diverse class and merge its findings per ${CLAUDE_PLUGIN_ROOT}/references/review-panel.md ("Second model family"). Do not ask the human "want me to dispatch?" — auto-dispatch at autonomy medium/high; confirm-then-dispatch at low. If verify needs resources you can't reach (live emulators, prod creds), write the exact dispatch command in the handoff and stop.
References
${CLAUDE_PLUGIN_ROOT}/agents/reviewer.md: the Step 2 role agent${CLAUDE_PLUGIN_ROOT}/references/parallelism.md: Tier 2, parallel Agent calls in one message${CLAUDE_PLUGIN_ROOT}/references/handoff-dispatch.md— when to dispatch vs stop- Foundation Design — charter format, verification independence
engineer/references/handoff-dispatch.md— infra-ensure-before-stop rule
Security
- Security scanPassed, 100 of 100Pattern checks on the skill's files.
- Model reviewNot runA model review is scheduled for the next rescan.
Quality evaluation
- Task evaluationNot runNo evaluation yet. Run one in Skill Studio.
Source
- SKILL.md in the repositoryCheckingengineer/skills/refine/SKILL.md
Popularity
- 7-day trend score
- 0
Not refreshed yet.
Works with
Any agent vskill installs to: 56 targets in all. These 39 have a page on this site.
- Claude Code
- Cursor
- GitHub Copilot
- Windsurf
- Codex
- Gemini CLI
- Cline
- Amp
- Roo Code
- Goose
- Continue
- Kilo Code
- Junie
- Trae
- OpenHands
- Qwen Code
- AI Agent
- Tabnine Chat
- Sourcegraph Cody
- Amazon Q
- Firebase Genkit
- Sweep
- Devin
- Phind
- Mentat
- Claude Desktop
- Aider
- GPT Pilot
- Plandex
- bolt.new
- v0
- Replit Agent
- Anse
- TypingMind
- BetterChatGPT
- LibreChat
- Warp AI
- Pieces
- Fig AI
Embed badge
[](https://verified-skill.com/skills/swingerman/engineer/refine)