Claude Code Config

Explore codebase and create structured implementation plan

Run quality, review, and review-tests checks in parallel, then fix, lint and format

Use when Claude wants a read-only second opinion from OpenAI Codex CLI on: exploring an unfamiliar codebase, reviewing a plan/design .md, or reviewing a PR diff. Codex runs sandboxed read-only (no writes, no prompts, no network).

When you get a failed result/error/user says its wrong/its a bug, use this skill to systematically debug and fix the problem.

Run comprehensive code review on current branch changes

Run quality checks, fix code quality issues, check code style, format code, or perform lint checks

Adversarial devil's advocate. Use when the user wants their idea, plan, PR, code review, architectural decision, or claim deliberately attacked — 'poke holes in X', 'tell me why I'm wrong about Y', 'red-team this', 'what am I missing'. Distinct from /ask (neutral), /plan (constructive), /brainstorm (open exploration), /review (structured code review). Use when the user explicitly wants pushback.

Create a pull request. MUST be invoked for ANY request that creates a PR — including 'create a PR', 'open a PR', 'gh pr create', 'submit for review', or any equivalent phrasing. Do NOT call `gh pr create` directly; always run this skill instead.

Run the CI watcher script for the current or specified branch. Use `/ci-watcher stop` to kill the watcher for the current session.

Surface every OPEN end blocking work — next actions, open tickets, pending decisions — as a short numbered list (title, one-line description, one-line why-it's-open, short recommendation) grouped under Open Ends. Never lists what's already done. Ends by offering to walk through them one by one. Use when the user asks 'what's next', 'next steps', 'open ends', 'what decisions are left', 'what do you need from me', or /next-steps.

Present a list one item at a time, every item numbered. User says 'next' or 'go' to advance, or asks a question to stay on the item. Claude's own questions favor numbered alternatives, yes/no, or 'go' for a stated default, to keep replies short. Uses the adhd-structure skill to format each item's content. Logs answers without acting until every item is resolved, and survives interruptions (subagent/monitor notifications) by logging them too and reprinting the pending question.

Plan and build tests for the current feature

Re-authenticate an expired gcloud session. Use PROACTIVELY, without being asked by name, whenever a `gcloud`/`gsutil`/`bq`/`pulumi` command fails on credentials — 'reauthentication required', 'invalid_grant', 'invalid_rapt', 'credentials not found', 'token has expired', 'Reauthentication failed', ADC/refresh-token errors, or a `pulumi` command (e.g. `pulumi whoami`, `pulumi login`, `pulumi up` against the GCS state backend) failing with `invalid_grant`/`invalid_rapt` — or whenever the user asks to re-auth, re-login, or log in to gcloud.

Start a new feature with full planning. Use PROACTIVELY, without being asked by name, whenever the user asks to start any new feature or substantial piece of work — 'build X', 'add X feature', 'implement X', 'create X', 'let's start working on X', 'new feature' — anything beyond a trivial one-file edit.

Assign or update a short label describing what this Claude Code session is currently doing, stored in a per-session sidecar file so status_line.sh and the tab title can show it

Plain-language format for one chunk of content, with escalating detail: 1-line summary, then a 5-line version, then a 3-part deep dive added only on request. Used by the one-by-one skill for each item. Use proactively — by default, not only when asked — for every markdown file and ticket (Linear issues, PR descriptions, plan/review docs).

Create isolated git worktree for feature

Resume work on existing feature worktree

Get a secret value that only exists in a third-party dashboard (Supabase, GCP, etc.) and seed it into keyshelf. Use whenever a fix needs a credential/API key/password that has no CLI or Management-API retrieval path — resetting a DB password, revealing a service-role/secret API key, copying an OAuth client secret, etc.

DevOps playbook for risky multi-step rollouts, migrations, and cutovers — incremental single-purpose PRs, inert-but-present fallbacks for one-line rollback, self-validation in prod (Chrome walkthroughs, /healthz probes, expected-vs-observed GO/NO-GO tables, negative controls), human-gated irreversible steps, CI-guard discipline, and durable milestone memory. Invoke for any high-blast-radius change e.g. auth/issuer cutovers, signing-key revokes.

End-to-end feature development pipeline: plan → implement → build tests → post checks → PR → summary. Invoked by /new-feature and /continue-feature. this skill should not be invoked directly but rether by the other skills.

Fast feature development: worktree → implement → PR (no planning, no tests)

Decompose a large, multi-part feature or initiative into decoupled groups (epics), research each with parallel subagents (get a second opinion from Codex on genuinely hard architecture calls), draft MVP-scoped tickets under each group — created in the project's ticket tracker if one exists, else a single markdown plan file as fallback — and gate ALL implementation behind a full `/one-by-one` user-approval pass. Never let implementation start before that gate completes.

Use whenever a task might require a human-only action — approving an OAuth/app install, clicking a payment or install confirmation, revealing/copying a one-time credential, solving a CAPTCHA, or anything else Shopify/GCP/etc. structurally require a real logged-in human for. Do the entire task yourself first via Chrome MCP, right up to the blocked step, and only then ask — with the browser already sitting on the exact page, and the ask trimmed to the 1-2 clicks left.