Skip to content

colorpulse6/claude-skills

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

21 Commits
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Claude Skills

A personal Claude Code plugin marketplace, 19 plugins deep: a code-review ladder (risk-gate → pr-review → review-board, with Codex as a second model), a quality-gated push pipeline, session continuity (rehydrate / handoff / open-threads / session-recap), the Fable operating method (fable-method / architect), operator rhythms (marketing-ops), and meta-skills for authoring more (create-skill / marketplace-lint / skill-starter). Auto-detects project type — no per-project configuration needed.

Skills

/pr-review

AI-powered PR review that combines automated rule checking with interactive walkthrough.

Features:

  • Inbox mode - /pr-review with no args discovers all PRs awaiting your review
  • Single PR mode - /pr-review 42 or /pr-review <url>
  • Incremental re-review - only reviews new commits if you've already reviewed the PR
  • Contract verification - checks alignment between GraphQL schema, resolvers, services, and Prisma/DB layers
  • Frontend + backend rules - auto-detects changed file types and applies relevant rules
  • Interactive walkthrough - finding-by-finding with Accept/Reject/Edit/Details
  • GitHub suggestion blocks - one-click apply for simple fixes
  • Line-specific comments - never posts code feedback as general PR comments

Usage:

/pr-review           # Inbox mode - pick from PRs awaiting your review
/pr-review 42        # Review PR #42 in current repo
/pr-review https://github.com/owner/repo/pull/42  # Review by URL

/pr-respond

Interactive walkthrough for addressing review comments on your own PRs. The counterpart to /pr-review.

Features:

  • Inbox mode - /pr-respond with no args finds all your open PRs with unresolved review threads
  • Single PR mode - /pr-respond 42 or /pr-respond <url>
  • Per-thread actions - Fix / Push back / Defer / Skip each unresolved comment
  • Batched implementation - one subagent handles all Fix items in a single pass, produces a coherent commit
  • Auto-detect type check - finds the right command (pnpm/yarn/npm with check-types or typecheck)
  • Diff review before commit - see the full diff before anything gets pushed
  • Smart replies - Fix replies include commit SHA with a link to the exact line; push-back replies use your reasoning verbatim
  • Auto-resolve - Fix and Push back threads are resolved, Defer and Skip stay open

Usage:

/pr-respond           # Inbox mode - pick from your PRs with unresolved threads
/pr-respond 42        # Respond to PR #42
/pr-respond https://github.com/owner/repo/pull/42  # By URL

/push

Quality-gated push pipeline that runs checks before pushing and creates PRs automatically.

Features:

  • Auto-detection - detects package manager (pnpm/yarn/npm) and available scripts
  • Parallel quality gates - type check, lint, tests, build
  • Auto-fix loop - retries failed gates after fixes (max 3 iterations)
  • Conventional commits - enforces commit message format
  • Branch policy - blocks direct push to main
  • Secret detection - scans staged files for credentials
  • PR creation - creates PR with summary and test plan
  • CI watch - polls CI and auto-fixes failures (max 2 iterations)

Usage:

/push  # Run the full pipeline

/session-recap

A browser-viewable HTML timeline of your recent Claude Code sessions — a dark "flight recorder" view to use as a reference point for what you worked on.

Features:

  • Daily or weekly - defaults to the last 7 days; pass a number or daily/weekly/month
  • Timeline by day - sessions grouped by day with project, branch, and time
  • Reference points - PRs (clickable), commits, files touched, and active time per session
  • Key decisions - one bounded LLM pass adds a summary + key decisions for notable sessions
  • Active time, not wall-clock - inter-message gaps over 30 min are dropped, so resumed sessions don't show absurd spans
  • Resume from a card - each session exposes its claude --resume <id> command (click to copy)
  • Filter + search - project chips (shift-click to solo), full-text search, PRs/commits-only toggle
  • Self-contained - one HTML file with data inlined; works offline

Usage:

/session-recap          # Last 7 days
/session-recap daily    # Last 24 hours
/session-recap 30       # Last 30 days

codex-second-opinion

An independent, separate-model (Codex CLI, flagship GPT tier — sol in current generations) review of code — to complement your own Claude subagent audits with a model that has different blind spots. In practice a separate model surfaces real bugs that parallel same-model reviews all miss.

Features:

  • Independent findings first, then confirm/refute - a neutral brief structure that gets genuine discovery, not a yes-man echo
  • Read-only - codex exec --sandbox read-only; reads + runs read-only commands, never edits
  • Reconcile, don't paste - convergence (high-confidence), net-new-from-Codex (verify yourself), and sharpened disagreements
  • For the risky stuff - concurrency, process lifecycle, async timing, money/LLM spend — where same-model reasoning shares blind spots

Usage: ask for "a codex second opinion" / "what does Codex think" before stacking work on a just-shipped slice. Requires the Codex CLI installed.

/skill-starter

A template plugin you copy to start a new skill. It's a complete, working example of the orchestrator pattern: one user-invocable skill resolves a target, fans the work out to parallel specialist subagents, and aggregates their findings into a single report.

Demonstrates:

  • plugin.json + SKILL.md frontmatter - the minimum to make a skill discoverable and user-invocable
  • Deterministic script layer - scripts/collect.sh resolves units of work as JSON; judgment stays in markdown
  • Parallel subagents - spawns one starter-worker agent per batch (tool- and model-scoped) in a single message so they run concurrently
  • Aggregation + output contract - workers write findings/<batch_id>.md, the skill merges them into REPORT.md
  • Error-handling table - predictable behavior for empty matches, failed workers, unwritable output

TEMPLATE: markers throughout show exactly where to swap in your own logic.

Usage:

/skill-starter                  # default: audit repo-tracked text files
/skill-starter "src/**/*.ts"    # audit a glob
/skill-starter ./docs           # audit a directory

/review-board

A multi-lens, multi-model code review conductor — the "parent" that composes parallel agents with codex-second-opinion.

What it does: on a diff / PR / branch, it spawns several scoped Claude review agents in parallel (each hunting one failure-class) while a separate model (Codex) reviews independently, then reconciles everything into one ranked, cross-model verdict.

Two kinds of diversity at once:

  • Lens diversity - security, concurrency, performance, contract lenses each tunnel-vision on one bug class, so nothing gets averaged away
  • Model diversity - Codex runs concurrently (its latency hides behind the lens work) and catches what same-model reasoning misses

Output: REVIEW-BOARD.md with a verdict, a cross-model table, convergence (high-confidence), net-new-from-Codex (verify yourself), and sharpened disagreements. Falls back to Claude-only if the Codex CLI isn't installed.

Customize: lenses live in references/lenses.md (one section each) and the agent is parameterized — add a lens without writing a new agent.

Usage:

/review-board           # review the working diff
/review-board 42        # review PR #42
/review-board my-branch # diff against the merge base

/risk-gate

Triage-then-escalate: scores a diff cheaply, sends only the risky parts to deep review.

What it does: a deterministic scorer (risk_score.py) rates each changed file on risk = likelihood × blast-radius (danger paths like auth/payments/migrations, churn, added control flow), buckets them Low/Medium/High, and escalates only HIGH-risk files to review-board (or codex-second-opinion). Spends the expensive model-diversity where it pays.

Usage:

/risk-gate            # triage the working diff
/risk-gate 42         # triage PR #42

/plan-harden

Adversarially review a design doc / spec / RFC before you build it — the cheapest place to fix a flaw.

What it does: a context gate (what is it / who's affected / success), a pre-mortem ("it's 6 months out and this failed — why?"), and parallel design-lens agents (edge-cases, failure-handling, data-integrity, rollout-rollback, observability, scaling, security/STRIDE, dependencies). Reconciles into a five-part hardening report + an APPROVED/REVISE verdict. All findings are [PLAN_RISK] (prospective).

Usage:

/plan-harden                 # newest design doc in docs/
/plan-harden docs/rfc.md     # a specific doc

/dep-audit

Audit dependency-manifest / lockfile changes in two passes — cheap offline signal, then real tools.

What it does: an offline pass (manifest_scan.py, zero network) flags semver deltas, range loosening, lockfile-integrity changes, new install scripts, and typosquats; then per-ecosystem dep-scanner agents run the real tools (osv-scanner, npm audit, pip-audit, cargo audit, govulncheck) plus maintainer-health checks, scoring each package on four axes (security / license / maintenance / breaking-change) with a Block/Warn/Monitor/Ignore verdict. Honest about offline limits.

Usage:

/dep-audit            # audit the working diff's dependency changes
/dep-audit 42         # audit PR #42

/test-gap

Find the most important missing tests in a change — ranked by risk, not coverage %.

What it does: fans out gap-finder agents over the changed files; each traces codepaths (changed branches, error paths, boundaries) and ranks untested ones by risk = impact × likelihood, using the surviving-mutant test to catch covered-but-unasserted code. Optional coverage (diff-cover) and mutation (Stryker/PIT/mutmut/cargo-mutants) passes confirm gaps. Outputs prioritized tests-to-add (P0→P2) with proposed cases.

Usage:

/test-gap             # find gaps in the working diff
/test-gap 42          # PR #42

/create-skill

A meta-skill for authoring new skills/plugins in this repo the right way — it bakes in the research behind every other plugin here.

What it does: interviews for intent + trigger phrases, picks the right shape (inline skill vs orchestrator + parallel agents), scaffolds the files (scaffold.sh), writes them against a distilled authoring guide (frontmatter, descriptions, progressive disclosure) and patterns reference (inline-vs-agents rubric, parameterized-agent pattern, confidence/severity/status-enum conventions), self-reviews against an anti-pattern checklist, and registers the plugin in the marketplace + README.

Two bundled references hold the substance:

  • authoring-guide.md — frontmatter rules, the description discovery surface, progressive disclosure (<500-line SKILL.md, one-level refs, TOC over 100 lines), subagent scoping, and the anti-pattern checklist
  • patterns.md — when to fan out vs stay inline, the parameterized-agent pattern, scripts-for-determinism, and the shared quality conventions

Usage:

/create-skill                       # interview-driven
/create-skill "audit GraphQL schemas"   # seed it with the intent

/marketplace-lint

Validate that every plugin in this repo is structurally sound and properly registered — built with /create-skill as a dogfood.

What it checks: valid plugin.json whose name matches its directory, registration in both marketplace.json and the README, SKILL.md frontmatter (name + description), scripts that are executable and (for Python) compilable, and references over 100 lines that open with a ## Contents TOC. Emits findings as error (breaks discovery/use) or warn (drift), exits non-zero on errors. Inline skill + one deterministic script — no agents.

Usage:

/marketplace-lint              # lint all plugins
/marketplace-lint risk-gate    # lint just one

/fable-method

The operating method Claude Fable 5 left behind on its last publicly-available day — how it worked, written for whatever agent reads it next (Claude, Codex, or anything newer).

What it teaches: an evidence-first operating loop (orient → map → decide → act → verify → report), calibration rules for when to act vs ask vs stop, and red flags that catch rationalization ("it probably works" → run it). Four references carry the depth: operating-principles.md (the judgment layer — root cause, scope discipline, reversibility gate, when to stop), research-method.md (mapping unknown systems, delegation briefs with output contracts, the dual-review workflow), communication.md (outcome-first reporting; visual-first for this user), and codex-adaptation.md (running all of it outside Claude Code).

Usage:

/fable-method                          # load the method at task start
/fable-method "migrate the billing service"   # seed it with the task

/rehydrate

Restore full working context for a project after time away — a returning human or a fresh agent gets briefed in minutes.

What it does: reads the project's dated context pack (~/kb/docs/handoff/), verifies every load-bearing claim against live git/filesystem state (packs are maps, not territory), then delivers a visual status brief: where you left off, what changed since, landmines, and ONE recommended next thread — plus the full open-threads list. Heals stale packs on every use; builds a pack from live sources if none exists.

Usage:

/rehydrate               # full sweep across all packs
/rehydrate jt3           # one project

/marketing-ops

The operator layer for an autonomous content/growth engine — the machine researches, writes, and renders; this runs the four things machines shouldn't do alone: verify health, approve truth, distribute, decide from metrics.

What it does: a five-station weekly check-in (automation health, queue runway, the rendered-but-unposted pile, product-truth gate, funnel pulse), a posting ritual (prepare, never publish), and launch mode. References: channel-playbooks.md (Reddit, Product Hunt, HN, LinkedIn, faceless short-form, cold email — all rules-aware) and launch-checklist.md (gates → assets → sequencing → day-of runbook → T+7 review).

Usage:

/marketing-ops           # weekly check-in
/marketing-ops post      # today's posting ritual
/marketing-ops launch job-toast

/handoff

Close a working session so nothing is forgotten — the counterpart to /rehydrate.

What it does: inventories the session (done-verified vs unverified vs abandoned — tiers never blur), leaves the repo consistent (branch WIP, never silent dirt), logs each thread to the knowledge base via kblog (one line: thread, state, next move), names EVERY open thread in a closing table, and heals any context pack the session made stale.

Usage:

/handoff                             # full closing protocol
/handoff "shipped billing-v2 gate A" # seed the summary

/open-threads

Cross-project forgotten-work radar — the counterweight to "you finish what you remember."

What it does: a deterministic sweep (sweep.py) finds every local git repo and measures the forgotten work in it (dirty files, unpushed commits, never-pushed branches, stashes, staleness); the skill cross-references the context packs' declared threads, ages everything, and reports the top 5 threads at risk with one 15-minute first move each — plus pack-drift findings (pack claims that no longer match live git state). Read-only: it points at /handoff and /rehydrate for healing.

Usage:

/open-threads                     # full radar sweep
/open-threads ~/somewhere/else    # add extra scan roots

/architect

Frontier-model-plans, cheap-model-executes — spend expensive tokens on judgment, not typing.

What it does: writes a decision-complete spec (five-part contract: context, task, constraints, output contract, done criteria — every architectural fork decided), dispatches implementation to a cheap lane (cheaper Claude subagent or Codex CLI), runs fresh-context adversarial review (implementer never reviews; reviewer never implements), and accepts work only at the top of the claim ladder (WRITTEN < RUNS < VERIFIED — the architect re-runs the done-criteria commands itself). Lanes fail loudly; a BLOCKED executor means the spec was incomplete, and the spec gets fixed, not the guess.

Usage:

/architect "migrate the billing service to the new API"

Installation

Via Plugin Marketplace

# In Claude Code
/plugin marketplace add colorpulse6/claude-skills

# Install the plugins you want
/plugin install pr-review@colorpulse6-skills
/plugin install pr-respond@colorpulse6-skills
/plugin install push@colorpulse6-skills
/plugin install session-recap@colorpulse6-skills
/plugin install codex-second-opinion@colorpulse6-skills
/plugin install skill-starter@colorpulse6-skills
/plugin install review-board@colorpulse6-skills
/plugin install risk-gate@colorpulse6-skills
/plugin install plan-harden@colorpulse6-skills
/plugin install dep-audit@colorpulse6-skills
/plugin install test-gap@colorpulse6-skills
/plugin install create-skill@colorpulse6-skills
/plugin install marketplace-lint@colorpulse6-skills
/plugin install fable-method@colorpulse6-skills
/plugin install rehydrate@colorpulse6-skills
/plugin install marketing-ops@colorpulse6-skills
/plugin install handoff@colorpulse6-skills
/plugin install open-threads@colorpulse6-skills
/plugin install architect@colorpulse6-skills

Using Codex or another non-Claude harness? See AGENTS.md — the skills are written tool-agnostically and load fine as plain instructions.

Once installed, the skills are available in any project on your machine.

How It Works

Project Detection

The skills auto-detect project characteristics by looking at:

Signal Indicates
pnpm-lock.yaml pnpm package manager
yarn.lock yarn package manager
package-lock.json npm package manager
tsconfig.json TypeScript project
prisma/schema.prisma Prisma ORM
*.graphql files GraphQL schema
.tsx / .jsx files React frontend
apps/api/, services/, server/ Backend code
package.json scripts Available commands (test, lint, check-types, build)

Commands that don't exist are skipped gracefully.

Review Rule Selection

When reviewing a PR, the skill inspects the changed files and conditionally loads rules:

  • Frontend files changed → loads frontend-rules.md
  • Backend files changed → loads backend-rules.md
  • Both → loads both sets

Universal rules (secrets, any types, error handling) always apply.

Local Overrides

If a project has its own .claude/skills/pr-review/ or .claude/skills/push/ directory, it takes precedence over the global skill. This lets you keep project-specific rules (e.g., custom terminology, domain-specific checks) separate from the generic skills.

Repo Structure

claude-skills/
├── .claude-plugin/
│   └── marketplace.json           # Marketplace manifest listing all plugins
├── plugins/
│   ├── pr-review/
│   │   ├── .claude-plugin/
│   │   │   └── plugin.json        # Plugin manifest
│   │   └── skills/
│   │       └── pr-review/
│   │           ├── SKILL.md
│   │           ├── review-lens.md
│   │           ├── output-contract.md
│   │           └── references/
│   │               ├── backend-rules.md
│   │               ├── frontend-rules.md
│   │               ├── contract-verification.md
│   │               └── severity-definitions.md
│   ├── pr-respond/
│   │   ├── .claude-plugin/
│   │   │   └── plugin.json
│   │   └── skills/
│   │       └── pr-respond/
│   │           └── SKILL.md
│   ├── push/
│   │   ├── .claude-plugin/
│   │   │   └── plugin.json
│   │   └── skills/
│   │       └── push/
│   │           ├── SKILL.md
│   │           └── scripts/
│   │               ├── preflight.sh
│   │               └── secret-scan.sh
│   ├── session-recap/
│   │   ├── .claude-plugin/
│   │   │   └── plugin.json
│   │   └── skills/
│   │       └── session-recap/
│   │           ├── SKILL.md
│   │           └── scripts/
│   │               ├── build_timeline.py
│   │               ├── build_digests.py
│   │               └── render_html.py
│   ├── skill-starter/                 # template plugin (orchestrator + parallel agents)
│   │   ├── .claude-plugin/
│   │   │   └── plugin.json
│   │   ├── agents/
│   │   │   └── starter-worker.md      # parallel specialist subagent
│   │   └── skills/
│   │       └── skill-starter/
│   │           ├── SKILL.md
│   │           └── scripts/
│   │               └── collect.sh     # deterministic unit collector
│   ├── review-board/                  # multi-lens + multi-model review conductor
│   │   ├── .claude-plugin/
│   │   │   └── plugin.json
│   │   ├── agents/
│   │   │   └── review-lens.md         # parameterized single-lens reviewer (parallel)
│   │   └── skills/
│   │       └── review-board/
│   │           ├── SKILL.md
│   │           └── references/
│   │               └── lenses.md      # per-lens checklists (add a lens here)
│   ├── risk-gate/                     # triage a diff, escalate only risky parts
│   │   ├── .claude-plugin/plugin.json
│   │   └── skills/risk-gate/
│   │       ├── SKILL.md
│   │       ├── references/risk-signals.md
│   │       └── scripts/risk_score.py  # deterministic per-file risk scorer
│   ├── plan-harden/                   # adversarial pre-build design review
│   │   ├── .claude-plugin/plugin.json
│   │   ├── agents/design-lens.md      # parameterized adversarial lens (parallel)
│   │   └── skills/plan-harden/
│   │       ├── SKILL.md
│   │       └── references/review-checklist.md
│   ├── dep-audit/                     # dependency-change audit (offline + tools)
│   │   ├── .claude-plugin/plugin.json
│   │   ├── agents/dep-scanner.md      # per-ecosystem scanner (parallel)
│   │   └── skills/dep-audit/
│   │       ├── SKILL.md
│   │       ├── references/dep-checklist.md
│   │       └── scripts/manifest_scan.py  # offline diff scanner
│   ├── test-gap/                      # find the most important missing tests
│   │   ├── .claude-plugin/plugin.json
│   │   ├── agents/gap-finder.md       # per-shard gap finder (parallel)
│   │   └── skills/test-gap/
│   │       ├── SKILL.md
│   │       └── references/test-gap-heuristics.md
│   ├── create-skill/                  # meta-skill: author new skills the right way
│   │   ├── .claude-plugin/plugin.json
│   │   └── skills/create-skill/
│   │       ├── SKILL.md
│   │       ├── references/
│   │       │   ├── authoring-guide.md # frontmatter, disclosure, anti-patterns
│   │       │   └── patterns.md        # inline-vs-agents, quality conventions
│   │       └── scripts/scaffold.sh    # generates a new plugin skeleton
│   ├── marketplace-lint/              # validate plugin structure + registration
│   │   ├── .claude-plugin/plugin.json
│   │   └── skills/marketplace-lint/
│   │       ├── SKILL.md
│   │       └── scripts/marketplace-lint.py  # deterministic consistency checker
│   ├── fable-method/                  # the operating method Fable left behind
│   │   ├── .claude-plugin/plugin.json
│   │   └── skills/fable-method/
│   │       ├── SKILL.md
│   │       └── references/
│   │           ├── operating-principles.md  # the judgment layer
│   │           ├── research-method.md       # mapping, delegation, dual review
│   │           ├── communication.md         # outcome-first, visual-first reporting
│   │           └── codex-adaptation.md      # running it outside Claude Code
│   ├── rehydrate/                     # verified context restore after time away
│   │   ├── .claude-plugin/plugin.json
│   │   └── skills/rehydrate/SKILL.md
│   ├── marketing-ops/                 # operator rhythm for the growth engine
│   │   ├── .claude-plugin/plugin.json
│   │   └── skills/marketing-ops/
│   │       ├── SKILL.md
│   │       └── references/
│   │           ├── channel-playbooks.md     # Reddit/PH/HN/LinkedIn/short-form/email
│   │           └── launch-checklist.md      # gates → assets → day-of → T+7
│   ├── handoff/                       # end-of-session protocol (pairs with rehydrate)
│   │   ├── .claude-plugin/plugin.json
│   │   └── skills/handoff/SKILL.md
│   ├── open-threads/                  # cross-project forgotten-work radar
│   │   ├── .claude-plugin/plugin.json
│   │   └── skills/open-threads/
│   │       ├── SKILL.md
│   │       ├── references/triage.md         # aging thresholds, noise filters, drift rules
│   │       └── scripts/sweep.py             # deterministic git-repo sweep
│   └── architect/                     # frontier-plans / cheap-executes loop
│       ├── .claude-plugin/plugin.json
│       └── skills/architect/
│           ├── SKILL.md
│           └── references/spec-contract.md  # five-part contract, lanes, briefs
├── AGENTS.md                          # entry point for Codex / non-Claude agents
└── README.md

Contributing

This is a personal skills repo. Feel free to fork it and adapt to your own needs.

License

MIT

About

Personal Claude Code plugin marketplace

Resources

Stars

1 star

Watchers

0 watching

Forks

Releases

No releases published

Packages

 
 
 

Contributors