| name | klimkit-implement |
| description | Run Klimkit's full implementation workflow for code, docs, release, configuration, UI, backend, security-sensitive, or proof-producing changes. Use when the user asks Codex to implement, fix, refactor, release, migrate, wire up, or complete non-trivial repository work with checklist, verification, reflection, and final review. |
Klimkit Implement
Use this as the skills-first replacement for the old global AGENTS.md implementation workflow. Keep the work in the main agent by default. Do not use subagents for exploration, checklists, debugging, security, or reflection unless the user explicitly asks; the only built-in subagent use is the final-review gate.
Defaults
- Treat Klimkit as a Vercel Skills CLI package first: install and update the root
skills/ library with npx skills.
- Use normal Codex app or CLI sessions plus installed Klimkit skills for day-to-day work.
- Keep task evidence in the project repository, not in generated home files.
- Treat older Klimkit runtime machinery as deprecated legacy material. Do not route new workflows through it unless the user explicitly asks to maintain legacy code.
Workflow
- Intake
- Read the latest user request, repo instructions (
AGENTS.md), relevant docs/work/ LOGs (descend selectively), docs, and nearby code.
- If the repo lacks the
docs/work/ layout, use klimkit-setup.
- Decide whether the work is implementation, debugging, release, review, research, UI/proof, or security-sensitive.
- Explore
- Use
klimkit-code-explorer inline when the code path is unfamiliar, broad, or risky.
- Use direct repo reads for simple tasks.
- Checklist
- Use
klimkit-checklister before code changes for implementation work.
- Treat the checklist as blocking unless the user changes scope.
- Implement
- Choose the narrow method skill when helpful:
klimkit-diagnose, klimkit-tdd, klimkit-walkthrough, klimkit-report-server, or klimkit-create-worktree.
- Keep edits surgical and aligned with existing repo style.
- Do not route new work through deprecated runtime, Switchboard, sync scripts, or plugin machinery unless explicitly maintaining legacy code.
- If a task asks to migrate old Klimkit runtime behavior into skills, make the target skill own any reference scripts, templates, and setup instructions it needs.
- Verify
- Run checks that match the checklist and blast radius.
- Record exact commands, important outputs, skipped checks, and remaining risk.
- For UI/proof work, produce inspectable report evidence when the repo workflow requires it.
- Security
- Use
klimkit-security-auditor before completion when the change touches auth, secrets, permissions, containers, network exposure, infrastructure, user data, or unsafe defaults.
- Reflect
- Use
klimkit-reflector after verification and before final review for non-trivial work.
- If reflection exposes a material gap, fix it and rerun impacted checks.
- Final Review
- Draft the exact final response.
- Run fresh final-review passes with
klimkit-final-reviewer.
- Prefer 3 independent final-review subagents when subagents are available and the user has not forbidden them. Give each only the original request, checklist, changed files, verification evidence, reflection note, proof artifacts, and exact draft response.
- If subagents are unavailable or forbidden, run
klimkit-final-reviewer inline and state that limitation.
- Do not claim completion until final review returns
READY FOR USER or all remaining gaps are explicitly reported.
- Report
- Tell the user what changed, what passed, what was skipped, and how to inspect the result.
- For commits on
main in this repo, create or update the latest GitHub release according to repo instructions.
Skill Routing
klimkit-checklister: acceptance criteria before edits.
klimkit-code-explorer: read-only tracing before edits.
klimkit-security-auditor: security-sensitive review.
klimkit-reflector: cross-task synthesis after verification.
klimkit-final-reviewer: final gate, preferably from fresh subagents.
Read references/artifact-workflow.md when concrete docs/work/ paths, work-note naming, proof reports, or memory/log/reflection details matter.