# Claude Code Skills Pack

**18 production-grade skills for Claude Code. 5 free.**

Each skill is a `SKILL.md` that Claude Code loads automatically when you hit
the situation it covers — and then follows as an operating procedure, not a
suggestion. Rollout percentages for flag-gated releases. The exact loop for
killing a flaky test. The backwards-compatibility checklist for API changes.
The triage ladder for a pile of security findings. Written by someone who has
done these things at 3am, for the agent that has to do them next.

## What's in the pack

| Skill | What it does |
|-------|--------------|
| `shipping-behind-feature-flags` | Rollout schedule with exact percentages and dwell times, written error-rate thresholds, and a rollback that's a 60-second flag flip. |
| `postmortem-writing` | Blameless format with a sourced timeline, root-cause discipline past "bug in X," and a 5-rule action-item rubric that rejects "improve monitoring." |
| `dependency-upgrade-triage` | A T1–T5 severity ladder (exploitable+reachable first), what to pin vs. skip vs. defer, and suppressions that require an owner and a review date. |
| `git-rebase-survival-guide` | Pre-flight checks, the conflict loop with test-after-every-continue, reflog recovery for bad force-pushes, and the rebase-vs-merge decision. |
| `perf-profiling-loop` | Measure→hypothesize→profile→fix loop, flamegraph reading heuristics, per-stack profiling commands. |
| `test-pyramid-refactor` | What to cut from an inverted suite, the bug-reintroduction check that makes cutting safe, and the time budgets per layer. |
| `api-contract-review` | The additive/risky/breaking classification, the deprecation-window rules, and the `oasdiff`/`buf breaking` CI gate. |
| `incident-runbook-authoring` | Runbook structure built for the tired 3am reader, the 5-minute triage rule, and the quarterly drill that keeps runbooks honest. |
| `code-review-discipline` | The B/M/m/nit severity rubric, the three-pass review structure, and the letting-go rules that keep PRs shipping Thursday. |
| `mcp-server-authoring` | Tool-schema design rules (one tool = one intent, enums over open-ended params), output byte-budgets, and the 5-prompt smoke test. |
| `claude-md-wrangling` | What belongs in CLAUDE.md and what doesn't, the 150-line budget, subdirectory splitting, and the fresh-session verification loop. |
| `flaky-test-hunting` | The N-run reproduction loop, the closed classification set (timer/order/state/race/network/locale), and the 200-run exit criteria. |
| `migration-rollback-plan` | Expand→migrate→contract sequencing, chunked idempotent backfills with lock-avoidance settings, and the rollback table written *before* anything runs. |
| `security-scan-triage` | Reachability-ranked severity (exploitability before CVSS), the source-to-sink check for SAST, and suppression entries with owners and dates. |
| `observability-instrumentation` | SLO-first instrumentation, 14.4×/6× burn-rate alerts, RED/USE metrics, and the cardinality trap that breaks TSDBs. |
| `release-notes-crafting` | Commit-to-user-sentence translation, the three-part breaking-change form, and audience altitude per reader. |
| `docs-drift-audit` | Executing documented commands in a clean clone, verifying defaults against source, and doc-tests that make drift a CI failure. |
| `ai-agent-eval-design` | Golden-set construction from production failures, anchored rubrics, LLM-judge bias controls, and the calibration bar (κ ≥ 0.6) before you trust a score. |

## Free (5)

`flaky-test-hunting`, `code-review-discipline`, `postmortem-writing`,
`dependency-upgrade-triage`, `claude-md-wrangling` — the five highest-traffic
skills, included in the zip under `skills-pack/lite/` as a lead magnet.

## Install

Download `skills-pack.zip` from this repository, unzip, then copy the
`skills/` directory into your Claude Code skills location:

```bash
# Global (all projects)
cp -r skills ~/.claude/skills/

# Project-scoped (this repo only)
cp -r skills .claude/skills/
```

Each skill activates automatically when your task matches its trigger — say
"this test is flaky, find out why" and `flaky-test-hunting` loads. You can
also invoke explicitly: "use the postmortem-writing skill."

## License

Personal, non-commercial. Use in unlimited personal and work projects. Don't
resell the pack (or a substantially modified version of it) as-is. You may
use the skills' output (your rollouts, postmortems, runbooks) however you
like — those are yours.

---

*Part of the [freelance portfolio](https://huggingface.co/datasets/protoneko/freelance-samples)
of Minh Tran, automation & data engineer.*