Build with intention. Review with evidence. Launch with fewer blind spots.
A curated, evidence-oriented checklist library for production Generative AI, LLMOps, autonomous agents, and the supporting software needed to ship them safely.
91 scoped checklists · 4,215 unique checks · 4 complementary review scopes
Great products are rarely derailed by one spectacular mistake. More often, the problems are questions nobody asked, assumptions nobody tested, and responsibilities nobody owned.
Awesome Checklists helps teams ask those questions before they become expensive surprises. Use it to turn vague readiness discussions into structured, evidence-based reviews—without reinventing a process for every project.
Short curated Awesome index: AWESOME.md is the linted, phase-oriented selection; this README introduces the source library and agent skill; DEVELOPMENT.md contains technical and maintenance details.
- Production GenAI lifecycle
- Start with 25–50 checks
- Agent skill for Codex and Claude Code
- Explore the checklists
- How to use these checklists
- Development and maintenance
These phase-oriented drop-in PR/issue templates complement, rather than replace, the existing 91 numbered source checklists. Each item links to its canonical stable ID and explains the failure mode it helps detect. [PR Template], [Security Audit], [Eval Benchmark], and [Runtime Guardrail] are editorial navigation tags, not certifications or claims that a check was run.
[PR Template] Prompt and context checklist — verify trust boundaries, strict output schemas, context-token budgets, truncation and injection resistance before a prompt or tool schema change. Source: Prompt engineering and versioning.
[Eval Benchmark] RAG and data checklist — prevent broken document chunking, incompatible embeddings, partial index swaps, unmeasured retrieval loss and stale/deleted vector data. Sources: RAG and knowledge bases and AI evaluation.
[Runtime Guardrail] Agent safety checklist — validate tool registry membership, operation-level permissions, prompt-injection defenses, handoff traces and partial-stream action boundaries. Sources: Tool orchestration, red-teaming and streaming inference.
[Eval Benchmark] Evaluation and observability checklist — catch regressions by keeping golden cases, schema and tool assertions, retrieval quality, user-visible latency and usage costs separate. Sources: Evaluation and regression and AI monitoring and drift.
[Security Audit] Production readiness checklist — verify cold-start behavior, secrets scanning, quotas, dependency/container patching and rate-limit recovery before promoting a release. Sources: Backend operations, performance and inference serving.
For Pro-GenAI implementation examples and where to apply these checks, see development companions.
Do not attempt all 4,215 checks. Start with the Core 50 curated starter or the 25-check starter for each relevant scope. You can also generate a focused, UNKNOWN-first project review rather than copying the entire source library.
Only the Core 50 priority tiers are individually curated; all other tiers are provisional, text-derived suggestions. The generator does not use provisional P0/P1/P2 labels for selection ranking. Determine project-specific severity from real threats, obligations, evidence and qualified human review. No independent field validation or external reviewer sign-off is claimed.
Use these checklists directly in your coding agent. The awesome-checklists agent skill teaches Codex, Claude Code, and other Agent Skills-compatible tools to choose relevant reviews, retrieve revision-pinned checklist criteria, inspect real project evidence, and report actionable findings with stable IDs.
From the project you want to review, install the skill using the Skills CLI:
npx skills add https://github.com/Pro-GenAI/awesome-checklists --skill awesome-checklists -a codex -a claude-codeTo install for just one agent, use -a codex or -a claude-code; add -g for a global installation. Discover the packaged skill without installing it with npx skills add Pro-GenAI/awesome-checklists --list. The CLI requires a supported Node.js release (currently Node.js 22.20+).
After installation, ask an agent: “Review this app's release readiness using awesome-checklists; cite actual checklist IDs and evidence for each finding.” Other examples: “Audit our API security and database controls,” or “Assess this RAG/agent implementation for evaluation and guardrail gaps.”
For source pinning, anonymous access, and skill fetch-helper internals, see agent skill internals.
What is new here—and what is not: OWASP ASVS, NIST AI RMF, accessibility guidance and established release-readiness practices already cover many of these topics. The repository's contribution is a cross-scope synthesis with stable IDs, agent-assisted retrieval, bounded review generation and evidence-record conventions—not novel security standards or an independent certification. Its source files have not all received independent expert or field validation; low or high GitHub engagement is not a substitute for that evidence. See authorship and validation limits.
- Spot blind spots earlier. Surface overlooked requirements, security risks, accessibility gaps, operational dependencies, and AI-specific failure modes.
- Make reviews actionable. Checks are written to prompt verification, not just agreement. Record evidence, ownership, exceptions, and follow-up work in your project's review records.
- Create a shared language. Stable checklist IDs make it easier to discuss findings across engineering, product, QA, security, and operations.
- Start wherever you are. Use relevant checks during discovery, implementation, release review, or post-launch changes—not only on launch day.
- Adapt to your context. Choose the scopes and checks that matter for your product, rather than treating every item as mandatory.
All 91 checklists live in lists/, grouped into shared, frontend, backend, and AI scopes. AI-specific interface checks are included in the relevant AI domain files.
| Scope | Files | Checks | What it covers |
|---|---|---|---|
| Shared / cross-cutting | 14 | 701 | Planning, legal considerations, architecture, security, testing, reliability, operations, metrics, and launch decisions |
| Frontend | 20 | 684 | UI/UX, accessibility, browser/API integration, performance, SEO, operations, and release checks |
| Backend | 35 | 1,118 | APIs, data management, services, infrastructure, security, operations, and applicable AI integrations |
| AI | 22 | 1,712 | Model selection, training, RAG, agents, evaluation, guardrails, monitoring, and their AI-specific frontend integration checks |
| Total | 91 | 4,215 |
Counts reflect the current inventory and may change.
For inventory generation, stable ID ownership, and other technical details, see DEVELOPMENT.md.
For a first pass, use the Core 50 starter review. Choose only applicable scopes and add checks for your system's architecture, risks, and release requirements. This is a starting point, not a certification or universal minimum.
Priority labels are suggestions, not audited severity determinations or permission to skip a control. Reviewers must assess the actual project risks and applicable release gates.
For an evidence-linked example, see the sample review of this repository.
| Review depth | Practical entry point | What it establishes |
|---|---|---|
| First pass | 25 per scope or Core 50 | A bounded set of candidate risks, with all project results initially UNKNOWN |
| Release review | Scoped 25–100-check generator, additional applicable controls, all release gates | A project-specific, evidence-backed go/no-go decision by an authorized reviewer |
| Expanded audit | Wider checklist selection, versioned primary requirements and independent testing | Deeper assurance for the actual system, not automatic certification from checklist completion |
No fixed number of checks guarantees readiness. Priorities outside the Core 50 are provisional; scope, architecture and threat model determine what must be reviewed.
| If you're working on… | Open these checklists |
|---|---|
| A new project | Requirements & planning · Architecture & code quality |
| A product or release review | Security · Testing · Release & launch readiness |
| A user-facing application | Frontend UX & usability · Accessibility |
| An API or data platform | Backend API design · Data & database management |
| An AI-powered product | AI evaluation & governance · RAG & knowledge bases · Agents & tool orchestration |
Browse the complete frontend, backend, and AI directories.
- Select the relevant scopes. Start with applicable shared checks, then add frontend, backend, and/or AI checklists based on the actual components and risks of your project. Not every folder applies to every project.
- Create versioned project review records. Keep the reusable source templates unchecked; preferably generate a project-specific UNKNOWN-first review and retain the pinned JSON catalog, source SHA, product revision, reviewers and evidence. If you copy Markdown, compare upstream revisions before reusing prior PASS decisions.
- Verify—don't assume. Mark
[x]only when a check has actually passed and supporting evidence has been recorded in your project's review notes, test artifacts, or issue tracker. - Explain exceptions. Keep failed or pending checks open. Document
N/Awith a reason, and track material issues or accepted risks with owners, deadlines, and the relevant stable checklist IDs. - Review the release gates. Evaluate each domain's release gate across all applicable scope files, with appropriate human approval. A repeated gate is not a separate approval.
- Revisit after meaningful changes. New features, vendors, models, policies, markets, deployments, or incidents can invalidate yesterday's assumptions.
A checked box should mean “we verified this,” not “we hope this is fine.”
Projects that have actually carried out and documented an applicable awesome-checklists review can display a voluntary, self-declared badge in their own README. The larger badge includes a checkmark to indicate a checklist-based review—not approval, certification, or a verified passing result:
Copy this Markdown into your project's README (no badge hosting required):
[](https://github.com/Pro-GenAI/awesome-checklists)For a self-hosted SVG badge option, see badge assets.
Important
This is a maintainer's self-declaration of having used the checklists, not an approval issued by Pro-GenAI or an independent audit. The badge does not certify security, compliance, accessibility, AI safety, or release readiness. Display it only after relevant checks were evaluated and results, outstanding risks, and exceptions were documented. Include a link to the project's own review summary where appropriate, and update or remove the badge if its basis is no longer current.
CC0 does not require attribution, but a link helps others discover and improve the source. If you use the checklists without performing a documented review, you can voluntarily write:
Uses [awesome-checklists](https://github.com/Pro-GenAI/awesome-checklists) by Pro-GenAI.This credit is not a certification, endorsement or substitute for the documented-review badge above. See the licensing guide for details.
Warning
These checklists are a starting point—not the ultimate, exhaustive, universally applicable, or authoritative source of truth.
They represent review prompts and practical references, not official standards, certifications, legal advice, or a guarantee of security, compliance, accessibility, reliability, AI safety, product quality, or release readiness.
Requirements depend on your architecture, users, risk profile, industry, jurisdiction, and the versions of standards and tools you use. Some checks may be inapplicable; others may be missing, incomplete, or become outdated. Always verify important claims against current primary sources and involve qualified reviewers where needed.
Use professional judgment. Challenge the checklist. Never mistake checklist completion for proof that a system is safe or ready.
Repository architecture, ID registries, source pinning, export/generator commands, validation, CI, source provenance and technical implementation examples are documented in DEVELOPMENT.md. For contribution requirements see CONTRIBUTING.md, and for agent-maintainer instructions see AGENTS.md.
Original checklist text, editorial curation and documentation for which contributors have sufficient rights are offered under CC0 1.0 Universal for frictionless internal and commercial reuse. Original executable source, agent-skill scripts, website tooling and automation use the MIT License. See the file-by-file licensing and SPDX guide for the scope table, mixed files, third-party exclusions and contributor-rights procedure. Neither license grants rights to linked standards, trademarks, logos or uncleared third-party material.
Found a gap? Disagree with a check? Have a better reference or clearer way to verify something? We want to hear it.
This collection is intended to improve through practical experience, constructive criticism, and fresh evidence. Suggestions are welcome whether they propose a new check, correct an existing one, clarify applicability, or remove a misleading requirement.
- Open an issue to propose an improvement or report an outdated, ambiguous, duplicated, or missing check.
- Make the suggestion specific. Include the relevant file and checklist ID (if any), the proposed change, why it matters, where it applies, and a current primary reference when possible.
- Contribute or report a gap. See CONTRIBUTING.md for contribution and review procedures, or AGENTS.md for stable-ID conventions and local checks.
- Small improvements count. Clearer wording, corrected links, stronger caveats, and well-supported counterexamples make the collection more useful.
The goal is not to defend every existing item. The goal is to make the questions better.
Use what helps. Question what doesn't. Contribute what is missing.