Awesome Agent Recipes

Copy-paste AI agent workflows for developers who want working automation in minutes, not another prompt graveyard.

A curated collection of production-shaped recipes for AI coding agents, MCP servers, RAG, GitHub Actions, Cursor, Claude Code, Codex, OpenHands, aider, n8n, and adjacent tools.

Why this exists

Most agent examples are screenshots, hype threads, or half-finished configs.

This repo is built around one rule: every recipe must be concrete, copyable, and useful in 5 minutes.

Use it to:

Need	Copy
Review PRs automatically	recipes/github-pr-review-agent
Triage issues	recipes/github-issue-triage-agent
Query Postgres safely through MCP	recipes/mcp-postgres-analyst
Debug tests with an agent	recipes/test-failure-debugger
Add SaaS-grade Cursor rules	recipes/cursor-rules-saas
Build local docs RAG	recipes/local-docs-rag

Quickstart

git clone https://github.com/igla/awesome-agent-recipes.git
cd awesome-agent-recipes
python3 scripts/validate-recipes.py

Pick a recipe:

mkdir -p ../target-repo/.github/agent-recipes/pr-review
cp -R recipes/github-pr-review-agent/. ../target-repo/.github/agent-recipes/pr-review/

Then read the recipe README and copy only the files you need.

Recipes

Recipe	Category	Works with	Best for	Time
GitHub PR Review Agent	GitHub Actions	GitHub, any LLM API	Automated PR review comments	5 min
GitHub Issue Triage Agent	GitHub Actions	GitHub, any LLM API	Labels, priority, routing	5 min
MCP Postgres Analyst	MCP	Claude Desktop, Cursor, Codex-style agents	Safe DB analysis	7 min
MCP Browser Automation	MCP	Playwright MCP, browser agents	UI testing and research	7 min
Local Docs RAG	RAG	Qdrant, Ollama, local agents	Private project docs search	10 min
Cursor Rules for SaaS	Editor rules	Cursor	Architecture, tests, security	2 min
Claude Code Commands	Agent commands	Claude Code	Repeatable slash commands	3 min
Codex Task Prompts	Prompt pack	Codex, coding agents	Refactor, tests, bugfix, docs	3 min
AI Commit Message Generator	Dev tools	Git, local CLI	Conventional commit messages	2 min
Test Failure Debugger	Debugging	Any coding agent	Failing tests, CI logs	5 min
Security Review Agent	Security	Any coding agent	OWASP-style review	5 min
Docs Writer Agent	Documentation	Any coding agent	README, ADR, API docs	5 min

Categories

Category	Goal	Recipes
GitHub automation	PRs, issues, CI	PR review, issue triage
MCP	Tool access for agents	Postgres, browser automation
RAG	Local knowledge retrieval	Local docs RAG
Coding-agent prompts	Better outputs from agents	Codex prompts, test debugger, security review
Editor rules	Persistent agent behavior	Cursor SaaS rules, Claude Code commands
Dev workflow	Daily engineering glue	Commit generator, docs writer

Example workflow

Automated PR review agent

name: Agent PR Review

on:
  pull_request:
    types: [opened, synchronize, reopened, ready_for_review]

permissions:
  contents: read
  pull-requests: write

jobs:
  review:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - name: Run review
        uses: actions/github-script@v7
        env:
          REVIEW_PROMPT_PATH: recipes/github-pr-review-agent/prompt.md
        with:
          script: |
            const fs = require('fs');
            const prompt = fs.readFileSync(process.env.REVIEW_PROMPT_PATH, 'utf8');
            core.info(prompt.slice(0, 300));

Full version: recipes/github-pr-review-agent/workflow.example.yml

Copy-paste in 5 minutes

Step	Command
Validate recipes	python3 scripts/validate-recipes.py
Copy a recipe	cp -R recipes/test-failure-debugger .github/agent-recipes/
Copy a prompt only	cp recipes/security-review-agent/prompt.md .github/prompts/security-review.md
Copy Cursor rules	cp recipes/cursor-rules-saas/.cursorrules ../target-saas-app/.cursorrules
Copy Claude commands	cp -R recipes/claude-code-commands/commands ../target-repo/.claude/commands

Who is this for

You are	You get
Solo developer	Agent workflows without setup research
Startup engineer	PR review, issue triage, docs, tests
OSS maintainer	Repeatable automation for contributors
Platform team	Safe MCP and RAG patterns
Staff engineer	Reusable prompts with clear quality bars

Contributing

Good recipes beat long lists.

Before opening a PR:

python3 scripts/validate-recipes.py

Read CONTRIBUTING.md and start from templates/recipe-template.

Roadmap

Status	Area	Recipes
Ready	GitHub automation	PR review, issue triage
Ready	MCP	Postgres analyst, browser automation
Ready	Local-first AI	Docs RAG
Ready	Coding agents	Codex prompts, Claude commands, Cursor rules
Next	Framework packs	Next.js, Django, Rails, Laravel
Next	Agent evals	Regression tests for prompts
Next	Team ops	Release notes, dependency upgrade agents
Next	More MCP	Linear, Slack, Sentry, Notion

Detailed roadmap: docs/roadmap.md

Star history

Star this repo if it saves you time

If one recipe removes one boring hour from your week, the repo did its job.

How to choose the right recipe

If you want	Start here	Why
Faster reviews	github-pr-review-agent	It comments only when the diff contains actionable risk.
Cleaner issue queues	github-issue-triage-agent	It maps reports to labels, priority, and next maintainer action.
Safe data analysis	mcp-postgres-analyst	It forces read-only SQL, limits, assumptions, and query evidence.
UI verification	mcp-browser-automation	It turns browser observations into reproducible bug reports.
Private project Q&A	local-docs-rag	It keeps docs retrieval local and source-cited.
Better editor behavior	cursor-rules-saas	It makes SaaS constraints persistent across tasks.
Repeatable agent commands	claude-code-commands	It gives maintainers slash commands with strict output formats.
Better coding-agent tasks	codex-task-prompts	It adds task-specific constraints for refactors, tests, bugfixes, features, and docs.

What makes a recipe production-shaped

Dimension	Standard in this repo
Prompt	Role, goal, inputs, rules, checklist, output format, and safety boundaries.
Config	Concrete defaults that can be copied and edited without guessing.
Output	Example output showing the expected level of evidence and brevity.
Safety	Explicit least-privilege permissions and blocked actions.
Portability	Recipes avoid unnecessary vendor lock-in and can be adapted to multiple agents.
Maintenance	Each recipe is small enough to review and improve independently.

Adoption playbooks

Solo developer

1. Copy one prompt into your agent.
2. Run it on one small task.
3. Save the prompt in .github/prompts, .claude/commands, or your editor rules.
4. Add one project-specific rule after the first successful use.
5. Keep the recipe narrow instead of turning it into a universal mega-agent.

OSS maintainer

1. Start with issue triage or PR review.
2. Keep write permissions minimal.
3. Use comments before automatic state changes.
4. Watch the first ten agent outputs manually.
5. Convert recurring maintainer feedback into recipe rules.

Platform or engineering team

1. Pick one workflow with measurable toil.
2. Define allowed tools and blocked actions.
3. Run the recipe on a sandbox repo.
4. Add policy to config.example.yml.
5. Require source-cited output before expanding usage.

Safety model

Risk	Default control
Hallucinated facts	Prompts require source evidence and explicit assumptions.
Overbroad permissions	GitHub workflows use scoped permissions.
Secret exposure	Examples use environment variables and warn against logging secrets.
Destructive automation	Recipes default to review, comment, or read-only behavior.
Bad agent output	Output formats make low-quality responses easy to reject.
Vendor churn	Configs are examples, prompts are portable.

FAQ

Are these full products?

No. They are sharp starting points: prompts, configs, workflows, and examples designed to be copied into real repositories.

Can I use another model provider?

Yes. Most recipes separate the prompt from the transport. Replace the API call or agent runtime and keep the behavior contract.

Should agents be allowed to merge, deploy, or mutate production data?

Not by default. Start with comments, read-only analysis, and human review. Add mutations only after you define rollback, audit logs, and approvals.

Why are the prompts strict?

Loose prompts produce confident noise. These prompts force evidence, scope, output structure, and explicit uncertainty.

How should I share a recipe publicly?

Show the exact workflow it solves, the files to copy, the safety boundary, and one realistic output. Avoid screenshots without runnable content.

Ultimate maintainer playbook

Launch positioning

The strongest pitch for this repo is not "awesome AI links." It is:

Copy-paste agent workflows that turn real developer toil into reviewable automation.

That line matters because it promises a concrete artifact: prompts, configs, workflows, and examples that can be used immediately.

Viral surfaces

Surface	What to show	What to avoid
GitHub README	Recipe table, quickstart, copy paths, safety posture.	Long manifesto before usage.
X	One concrete workflow before the link.	Generic "AI agents are the future" framing.
LinkedIn	Team productivity and maintainer quality.	Inflated claims without examples.
Hacker News	Practical repo contents and constraints.	Hype or vendor loyalty.
Reddit	Before/after workflow examples.	Link drop without useful context.

Star-worthy repo signals

Signal	How this repo shows it
Immediate utility	Every recipe has copyable files.
Taste	Strict output formats and safety defaults.
Depth	Prompts, configs, examples, docs, and validation.
Maintenance readiness	Issue templates, PR template, contribution guide, CI.
Shareability	Clear headline and practical categories.

Recipe maturity model

Stage	Description
Seed	Prompt plus basic README.
Useful	Prompt, config, output example, safety notes.
Trusted	Real failure modes, rollout plan, review rubric.
Team-ready	CI, permissions, metrics, and ownership guidance.
Canonical	Used by multiple repos and refined from real outputs.

This repository should only contain recipes at Useful or higher. The best recipes should become Trusted and Team-ready.

Example community pitch

I got tired of AI agent examples that are just screenshots.

So I made awesome-agent-recipes: copy-paste workflows for PR review, issue triage, MCP Postgres analysis, browser automation, local docs RAG, Cursor rules, Claude Code commands, Codex prompts, test debugging, security review, docs writing, and commit messages.

Every recipe includes a prompt, config, example output, safety notes, and validation.

Maintainer north star

Reject anything that is impressive but not useful. Accept anything that makes a developer faster while preserving evidence, safety, and taste.

Extra launch assets

Repository tagline variants

Variant	Use
Copy-paste agent workflows for real developer work.	GitHub description.
Stop saving AI prompt threads. Start copying working agent recipes.	Social launch.
Production-shaped prompts, configs, workflows, and examples for coding agents.	README intro.
The missing cookbook for AI coding-agent automation.	Community posts.

Launch checklist for GitHub Trending potential

• README has one-screen clarity.
• First table shows practical value immediately.
• Every recipe link opens to substantial docs.
• The repo has license, contributing, code of conduct, issue templates, and validation CI.
• Social post shows three concrete recipes, not a generic AI claim.
• The repository topics match search intent.
• The first issues invite specific recipe requests.
• The first PRs improve depth, not random links.

Community response snippets

This is not a link list. Each recipe includes prompt, config, output example, safety notes, and copy paths.

The goal is not autonomous everything. The goal is reviewable automation that saves maintainer time without hiding risk.

If you have one painful repeated workflow, open a recipe request with the input, output, safety boundary, and tool target.

Star conversion levers

Lever	Implementation
Immediate value	Quickstart plus recipe table.
Trust	Safety docs and strict prompts.
Depth	Every markdown file is substantial.
Contribution path	Template, guide, validation, issue forms.
Shareability	Clear tagline and social card.

Long-term moat

The moat is not having the most recipes. The moat is having recipes that working developers trust enough to copy into production-adjacent repositories.

Masterclass appendix

The bar this file should clear

This file should be strong enough that a developer can send only this link to another maintainer and the maintainer can understand the workflow, judge the safety model, and copy the useful parts without asking for a walkthrough.

Real-world operating constraints

Constraint	Practical implication
Maintainers are busy	Put the action before the background.
Agents are confident when wrong	Require evidence and uncertainty labels.
Tooling changes	Keep prompts portable and configs explicit.
Security matters	Prefer reviewable output over hidden mutation.
Teams differ	Show customization patterns, not rigid doctrine.
Public repos attract drive-by PRs	Make contribution standards visible in the file itself.

Evaluation suite

Use these checks after every meaningful edit:

1. Can a user identify the exact file to copy?
2. Can a user identify the exact command to run?
3. Can a user identify required secrets or permissions?
4. Can a maintainer reject unsupported agent output quickly?
5. Can a contributor see what a good example looks like?
6. Can the workflow fail safely?
7. Can the recipe be ported to another model provider?
8. Can the output be pasted into GitHub without extensive cleanup?
9. Can the safety notes prevent the most likely misuse?
10. Can the document survive being read out of context?

Expert review table

Review area	Excellent	Weak
Scope	One workflow with clear boundaries.	Multiple unrelated workflows mixed together.
Evidence	Claims cite concrete sources.	Claims rely on agent confidence.
Safety	Blocked actions are named.	Safety says only "be careful".
Setup	Commands and paths are copyable.	Setup requires guessing.
Output	Structured, short, and actionable.	Long prose with no decision point.
Maintenance	Failure modes feed prompt improvements.	Bad outputs are treated as random.

Negative test cases

A robust file should make these bad outcomes less likely:

Bad outcome	Prevention
Agent invents a dependency.	Require repository evidence before naming dependencies.
Agent says tests passed without running them.	Require exact command and result.
Agent performs risky mutation.	Default to read-only, comment-only, or recommendation-only behavior.
Agent leaks a secret from logs.	Redact or avoid repeating sensitive values.
Agent gives a broad rewrite.	Require minimal fix first.
Agent asks vague questions.	Require one narrow blocking question only.

Production readiness checklist

• The workflow has a clear owner.
• The trigger or invocation is documented.
• The default permissions are least privilege.
• The output has a fixed shape.
• The agent has a refusal path.
• The user can run a dry run.
• The first live outputs are manually reviewed.
• The integration can be disabled quickly.
• Bad outputs become test cases or prompt edits.
• The recipe stays small enough to reason about.

Copy block for maintainers

## Maintainer decision

Status: accept | request changes | reject

Reason:
- Evidence quality:
- Safety posture:
- Copy-paste usability:
- Output quality:

Required change:
- One concrete change, if any.

Portability notes

When moving this workflow to another tool, preserve behavior before changing style:

1. Copy the prompt unchanged.
2. Map config keys to the new tool.
3. Run on a small known case.
4. Compare output against this file's example.
5. Tune only the transport-specific pieces.

Long-form quality rubric

Score	User experience	Maintainer confidence
10	Feels like a polished OSS product.	Safe to recommend publicly.
9	Strong, practical, and slightly opinionated.	Safe after one local review.
8	Useful but missing one advanced case.	Needs small edit.
7	Works for insiders.	Needs maintainer explanation.
6	Interesting but incomplete.	Not ready.
5	Too generic.	Rewrite around a real workflow.

Final polish pass

Before publishing, read only headings and tables. If the file still communicates the workflow, safety model, and output quality from those alone, the structure is strong.