Fixing code

Skill alexei-led/cc-thingz/dist/claude/dev-flow/skills/fixing-code

Portable skills, agents, hooks, and Pi extensions for Claude Code, Codex CLI, Copilot, Cursor, Grok, and Pi.

Install

npx -y skills add alexei-led/cc-thingz --skill fixing-code

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

SKILL.md

3.2 KB, as published. Nobody here has run it

Fix and Diagnose Code

Follow the base skill. This Claude overlay only defines tool use and execution details.

Fix the requested defect or failing gate one verified issue at a time. Do not expand to unrelated failures without asking. No guessing.

Arguments

diagnose or investigate: use the hard-bug workflow with hypotheses and probes.
team: spawn read-only analysis agents to challenge root cause before editing.

Use TaskCreate and TaskUpdate when the fix has more than two steps:

Reproduce or run the failing gate.
Read the failing path.
Diagnose root cause.
Patch one issue.
Verify narrow and broad checks.
Clean up probes.

Tool order

Run the smallest known failing command or project gate. Prefer focused tests, lint, or typecheck while editing and the broader project gate before final output.
Use Read, Grep, and Glob to inspect the exact failing path.
Use AskUserQuestion if logs, payloads, access, repro steps, environment, or instrumentation approval are missing.
Use Edit for existing files and Write only for new files or complete rewrites.
Run the narrow repro after each patch.
Run the relevant broader check before final output.

Do not use destructive git commands. Do not use --no-verify. Do not clear caches, disable rules, or skip fast tests as a routine speed fix.

Repro commands

Prefer configured project commands. Examples:

make lint
make test
go test ./...
ruff check .
pytest -q --maxfail=1 --tb=short
bunx vitest run path/to/file.test.ts
bun test path/to/file.test.ts
npm test

Use only commands supported by the repo and available tools. Browser-only debugging belongs in browser-automation unless a cheaper CLI/unit signal exists.

Hard-bug mode

Before editing, write 3-5 ranked falsifiable hypotheses. Probe one at a time. If you add temporary logs, tag them with [DEBUG-<short-id>] and remove them before final output.

If team is set, agents analyze only. Ask for root cause, evidence, suggested fix, priority, and confidence. Verify their claims before editing.

Scope control

If all checks pass and no bug is reproduced, report that and ask for a repro artifact.
If the failing gate contains unrelated failures, fix the requested issue first and ask before expanding.
If the only test seam is too shallow, say so; do not create fake-confidence tests.
If a patch causes new failures, diagnose that failure before continuing.