AI skills guide for contributors¶
This page is for humans who want to create or maintain AI skills for SpecStar.
It is not itself a triggerable skill file. If you want to build a real skill, start from the skill folders under skills/ and then install or mirror them into the AI tool's expected location.
Where to start¶
The canonical examples for this repository live under skills/:
skills/specstar-backend/skills/specstar-frontend/skills/README.md
If you want to understand what a good SpecStar skill looks like, read those folders first.
What goes where¶
Human-authored skill source¶
Put reusable skill content under skills/<skill-name>/.
A typical skill package looks like this:
Use this location when you are:
- designing a new skill,
- expanding an existing skill,
- collecting references and examples,
- reviewing the skill with other humans.
Tool-specific installation location¶
If you want GitHub Copilot in this repo to use the skill directly, place the active copy under .github/skills/.
In short:
skills/= human-maintained source and examples.github/skills/= repo-local skill installation for Copilot-style tools
How to create a new AI skill¶
1. Choose one responsibility¶
A skill should own one clear job, for example:
- backend resource lifecycle work,
- frontend generator changes,
- documentation maintenance.
Avoid putting unrelated responsibilities into a single giant skill.
2. Create a SKILL.md¶
Start with a minimal structure like this:
---
name: example-skill
description: Explain what this skill does and when it should trigger.
---
# Example Skill
## Use this when
narrow list of situations where the skill should be loaded.
## Workflow
step-by-step guidance the model should follow.
## Verification
the checks or commands the model should run before claiming success.
The description field matters a lot. It should clearly say both:
- what the skill helps with, and
- when the assistant should use it.
3. Add references only when needed¶
If the skill needs lots of background material, put it under references/ instead of making SKILL.md huge.
4. Reuse real project conventions¶
A good SpecStar skill should reflect the real repo workflow, such as:
- use
uv runfor Python commands; - follow TDD for bug fixes;
- add targeted tests for new behavior;
- update docs and docstrings when public behavior changes;
- respect the generated versus customizable boundary in the web app.
5. Test the skill on real prompts¶
Before considering the skill finished, try it against a few realistic prompts and check whether it:
- triggers at the right time,
- follows the correct project conventions,
- suggests or runs the right verification steps.
Existing SpecStar skill map¶
Backend examples¶
skills/specstar-backend/SKILL.md
Frontend examples¶
skills/specstar-frontend/SKILL.md
Repo-installed skills for Copilot¶
The repo-local installed skills live under .github/skills/.
These are the versions the assistant can discover directly while working inside this repository.
Recommended workflow for contributors¶
- read the examples under
skills/; - draft or update the new skill there first;
- keep the trigger description concrete;
- verify the skill with realistic prompts;
- if needed, mirror the finalized version into
.github/skills/for repo use.
Need help writing one?¶
If you want the assistant to help design a skill, ask it to create or refine a skill for a specific workflow and point it to the examples in skills/.
That usually works best when you provide:
- the type of task,
- when the skill should trigger,
- the expected workflow,
- the checks the assistant must perform before saying the work is done.