Keep AI files with the project
Store prompts, rules, memory files, instructions, and skill files in the repo. A skill should explain what it does and why the project needs it.
AI Best Practices
Build AI skills that are clear, safe, and reusable.
A good AI workflow should not be lost in chat history. Turn it into a small, documented skill that can be installed, tested, and updated.
Make each skill small
A good skill does one repeated job: review code, summarize issues, write release notes, generate tests, or explain product context. Small skills are easier to test and update.
Write for people and AI
Include the goal, context, limits, examples, expected output, and known problems. A teammate should understand the skill before running it.
Name the tools it supports
Say whether the package supports Cursor, Claude, Codex, or another assistant. Tool files should go into predictable folders.
Test with real work
Before publishing, try the skill on work that looks like real use. Keep a short checklist for expected behavior, wrong behavior, and edge cases.
Keep secrets out
Do not package API keys, customer data, private prompts, credentials, or internal documents. Use ignore files and review the preview before publishing.
Use examples
Input and output examples help assistants follow the right pattern. Keep examples short, specific, and easy to scan.
Version important changes
Changing a skill can change how teammates work. Use a new version and explain what changed.
Keep people in control
For important work, make the skill draft, inspect, and explain. Do not make it take irreversible action silently.
Explain how to recover
Say what to do if install fails, a tool is unsupported, a generated file is wrong, or a user needs an older version.
AIPM skill quality checklist
- The skill has a clear name, description, supported targets, and entry file.
- The package includes only files that should be public.
- The instructions include context, constraints, examples, and expected output.
- The skill was tested against at least one realistic task before publishing.
- The version number matches the size of the change.
Source-backed guidance
This page turns public AI guidance into practical AIPM advice. These references are good starting points for teams building reusable AI workflows.
OpenAI prompting guideUse clear instructions, context, examples, and repeated testing.Anthropic prompt engineeringStructure prompts, add context, and test changes on real tasks.Google People + AI GuidebookDesign AI products around people, feedback, and trust.NIST AI Risk Management FrameworkFind and manage AI risks before broad rollout.Microsoft Responsible AIFairness, reliability, privacy, inclusion, transparency, and accountability.Stanford AI IndexYearly data about AI progress, use, evaluation, and governance.