AI Agent Skills

Portable, installable skills for AI coding agents such as Codex and Claude.

npm package	Install
@anuragsarkar97/ai-agent-skills	npx -y @anuragsarkar97/ai-agent-skills@latest install --agent codex --write --force

This repository gives agents reusable operating modes for review, planning, implementation, testing, product thinking, communication, reliability, observability, and stack-specific engineering judgment. Skills are intentionally small. Deeper guidance lives in curated knowledge packs that agents load only when needed.

Install

Install everything into Codex:

npx -y @anuragsarkar97/ai-agent-skills@latest install --agent codex --write --force

Install everything into Claude:

npx -y @anuragsarkar97/ai-agent-skills@latest install --agent claude --write --force

That single command copies all skills plus shared _knowledge/ into the agent skill directory.

Use

After installing, ask the agent normally. The router skill helps the agent select the right skill without you naming it every time.

Examples:

Review this API design for production readiness.

Grill this migration before I merge it.

Draft a firm but respectful Slack message explaining why we should narrow this scope.

Design the tables, indexes, constraints, and migration plan for organization invitations.

Production checkout is returning 500s after the last deploy. Triage whether we should rollback or hotfix.

I want to upskill myself in Kubernetes; find recent blogs, papers, docs, and videos and make me a 30-day plan.

/remember I need to send email to product manager at 10 pm tonight

Search our prior decisions before changing payment retry behavior.

Review this webhook endpoint for authorization, replay, and secret-exposure risks.

For explicit testing, name a skill:

Use the critical-thinking skill to challenge this design before we implement it.

Use caveman mode: answer only what I directly ask and stop adding unrelated explanation.

What You Get

Area	Skills
Routing and workflow	agent-skill-router, repository-map, implementation-plan
Critical judgment	critical-thinking, wtf-check, oppenheimer-simplifier
Product and communication	product-competitive-thinking, product-communication
Learning and growth	upskilling-research
Reliability and operations	incident-response, observability-design, security-review
Personal productivity	macos-reminder
Implementation writers	service-writer, api-writer, utility-writer, test-writer
Review and quality	code-review, api-review, database-schema-design, design-principles-review, test-design-review, naming-review
Documentation and delivery	code-documentation, commit-pr-writer
Context and maintenance	decision-memory, caveman-mode, smriti-shruti, self-amending-skill

Skill Highlights

• agent-skill-router: chooses and combines skills proactively.
• caveman-mode: keeps answers terse, directly relevant, and silent on unasked tangents.
• critical-thinking: challenges weak designs and says no to anti-patterns.
• wtf-check: interrupts human or AI drift, thrash, over-scoping, and context switching with a blunt reset question.
• product-competitive-thinking: reviews work through product, startup, user, and competitor lenses.
• product-communication: writes respectful but firm Slack, email, issue, PR, and stakeholder messages.
• upskilling-research: finds fresh blogs, papers, docs, talks, courses, and videos and turns them into a learning path.
• database-schema-design: designs tables, entities, migrations, indexes, constraints, and retention.
• api-review: reviews API contracts, auth behavior, pagination, compatibility, and error shapes.
• code-review: performs normal reviews and adversarial grills for risky production changes.
• incident-response: triages outages, rollback/hotfix decisions, severity, handoff, and post-mortems.
• repository-map: discovers repository stack, commands, entry points, operational files, and a per-file CSV map (purpose/role plus optional symbols).
• decision-memory: records and retrieves engineering decisions, rationale, rejected alternatives, and evidence.
• security-review: finds exploitable security and privacy risks with concrete attack paths and fixes.
• macos-reminder: schedules local macOS notification reminders from /remember style prompts.
• observability-design: designs logs, metrics, traces, SLOs, dashboards, and alerts.
• service-writer, api-writer, utility-writer, test-writer: keep implementation work split by responsibility.
• smriti-shruti: reduces noisy context while preserving the facts needed for the current task.
• self-amending-skill: improves the skill system safely from real usage evidence.

Knowledge Packs

Installed skills include shared _knowledge/ references. Agents load these only when the task calls for them, which keeps the active context lean.

Knowledge Pack	Use For
architecture/principles.md	SOLID, YAGNI, dependency direction, service boundaries
api/api-review.md	API contracts, auth, pagination, idempotency, errors
database/schema-design.md	entities, constraints, indexes, migrations, retention
communication/product-communication.md	Slack, email, feedback, conflict, decision messages
incident-response/incident-response.md	severity, incident roles, rollback/hotfix, post-mortems
observability/observability-design.md	golden signals, USE/RED, SLOs, logs, metrics, traces, alerts
golang/go-engineering.md	Go services, APIs, contexts, errors, concurrency, tests
react/react-engineering.md	React components, hooks, state, effects, accessibility, tests
python/python-scripting.md	Python CLIs, filesystem work, subprocesses, logging, script tests
kubernetes/kubernetes-operations.md	workloads, probes, resources, rollouts, RBAC, secrets
cloud/aws-azure-architecture.md	AWS, Azure, EKS, AKS, managed services, identity, networking
product/startup-pm.md	MVP scope, startup product bets, competitors, learning metrics
security/security-review.md	auth, permissions, secrets, PII, tenant isolation, webhooks
testing/testing-strategy.md	test levels, contract tests, regression coverage, false confidence

Common Commands

List available skills:

npx -y @anuragsarkar97/ai-agent-skills@latest list

Run package validation:

npx -y @anuragsarkar97/ai-agent-skills@latest check

Search skills and knowledge:

npx -y @anuragsarkar97/ai-agent-skills@latest search "api auth pagination"

Print bootstrap instructions for an agent:

npx -y @anuragsarkar97/ai-agent-skills@latest bootstrap --agent claude

Generate repository-specific context:

npx -y @anuragsarkar97/ai-agent-skills@latest index-project --path .

Record or search an engineering decision:

npx -y @anuragsarkar97/ai-agent-skills@latest decision-memory add --path . \
  --title "Do not retry payment creation" \
  --decision "Payment creation is never retried automatically" \
  --rationale "Retries can create duplicate charges"

npx -y @anuragsarkar97/ai-agent-skills@latest decision-memory search --path . --query "payment retry"

Install selected skills:

npx -y @anuragsarkar97/ai-agent-skills@latest install --agent codex --skills agent-skill-router,critical-thinking --write --force

Install the frontend-focused local profile:

npx -y @anuragsarkar97/ai-agent-skills@latest install --agent claude --profile frontend --write --force

Install third-party frontend skills from a Git URL:

npx -y @anuragsarkar97/ai-agent-skills@latest install --agent claude --external git@github.com:org/agent-skills.git#skills/ux-pro --write --force

Install several additional third-party frontend skills from a manifest:

npx -y @anuragsarkar97/ai-agent-skills@latest install --agent claude --profile frontend --external-manifest examples/frontend-external-skills.json --write --force

Install as symlinks from a local clone:

npm run skills:install -- --agent codex --mode symlink --write --force

Install into a custom target:

npx -y @anuragsarkar97/ai-agent-skills@latest install --target /path/to/agent/skills --write --force

How It Works

Each skill is a folder under skills/<skill-name>/ with a canonical SKILL.md.

skills/<skill-name>/
├── SKILL.md
├── agents/
│   ├── openai.yaml
│   └── claude.md
├── references/
├── scripts/
└── assets/

The SKILL.md file defines the trigger, workflow, checks, and output shape. Agent-specific metadata lives under agents/. Detailed reusable guidance lives in shared knowledge/ and is installed as _knowledge/ beside the skills.

Local Development

Clone and validate:

git clone git@github.com:anuragsarkar97/skills.git
cd skills
npm test

Create a skill:

npm run skills:create -- my-new-skill --description "Use when an AI agent needs to..."

Then edit skills/my-new-skill/SKILL.md with concrete triggers, workflow, checks, and output shape.

Create from a stronger template:

npm run skills:create -- my-review-skill --template review --description "Use when..."
npm run skills:create -- my-writer-skill --template writer --description "Use when..."
npm run skills:create -- my-ops-skill --template ops --description "Use when..."
npm run skills:create -- my-knowledge-skill --template knowledge-backed --description "Use when..."

Build And Package

One command runs the full local packaging flow:

npm run skills:package-all

It runs validation, regenerates the catalog and graph, packages Claude bundles, and performs an npm pack dry run.

For a publishable release, bump the version as part of the same flow:

npm run skills:package-all -- --patch

Use --minor, --major, or --bump 1.2.3 when the release needs a different version change.

Create Claude-ready ZIP bundles:

npm run skills:package-claude

Create selected Claude bundles:

npm run skills:package-claude -- --skills agent-skill-router,critical-thinking

Create an npm tarball locally:

npm run skills:package-all -- --pack-tarball

Publish:

npm run skills:package-all -- --patch
npm publish --access public

Full release automation:

npm run skills:release -- --patch

The release script runs package-all with a version bump, writes the changelog, commits the release files, publishes to npm, tags the release, and pushes the branch and tag. Use --dry-run to test publish packaging without changing git or npm.

Maintenance Scripts

npm run skills:audit
npm run skills:bootstrap -- --agent claude
npm run skills:catalog
npm run skills:changelog
npm run skills:check
npm run skills:duplicates
npm run skills:evaluate
npm run skills:examples
npm run skills:graph
npm run skills:index-project
npm run skills:knowledge-index
npm run skills:marketplace-manifest
npm run skills:refresh-knowledge
npm run skills:repository-map -- --path .
npm run skills:search -- "go service tests"
npm run skills:verify-sources

• skills:check runs validation, audit, prompt examples, evaluation hooks, and source checks.
• skills:bootstrap prints copy-paste agent instructions for proactive skill routing.
• skills:catalog writes skills/catalog.json.
• skills:changelog generates changelog text from commits since the last tag.
• skills:graph writes skills/graph.json and skills/graph.mmd.
• skills:evaluate checks prompt coverage and shared knowledge hooks.
• skills:knowledge-index writes knowledge/index.json from curated references.
• skills:search searches skills and shared knowledge references.
• skills:verify-sources verifies source metadata in knowledge references.
• skills:refresh-knowledge writes a review queue for source refresh work.
• skills:repository-map writes .skill-context/repo-files.csv, repo-symbols.csv, and repo-map.state.json, using Git changes and content hashes for verified incremental regeneration.
• skills:repository-map:enrich upgrades pending repo-files.csv rows with AI summaries (--mock for offline/tests; CURSOR_API_KEY + @cursor/sdk for live enrichment).
• skills:index-project writes .skill-context/project-context.json.

Skill Contract

Every skill must include:

• A lowercase hyphenated directory name.
• A SKILL.md file.
• YAML frontmatter with only name and description.
• A name value that matches the directory.
• A description that clearly says what the skill does and when an agent should use it.

Keep SKILL.md concise. Move bulky examples, schemas, policies, framework-specific details, and source-backed guidance into shared knowledge references.

Repository Layout

.
├── agents/                 # Agent-specific usage notes
├── AGENTS.md               # Root instructions for Codex-style agents
├── CLAUDE.md               # Root instructions for Claude-style agents
├── bin/                    # npm CLI entrypoint
├── examples/               # Prompt examples for validation
├── knowledge/              # Shared curated references
├── profiles/               # Install profiles such as frontend
├── scripts/                # Maintenance, packaging, and validation scripts
├── skills/                 # Reusable skills
└── templates/              # Skill templates

Contribution Notes

• Prefer small, focused skills over broad instruction dumps.
• Add or update prompt examples when adding a skill.
• Add source notes to every shared knowledge reference.
• Run npm run skills:package-all before publishing or opening a large skill-system change.
• Regenerate catalog and graph artifacts when inventory, metadata, or routing changes.