Agent Skills by stark AI
Give coding agents specs, guardrails and local code context
4 followers
Give coding agents specs, guardrails and local code context
4 followers
10 open-source skills for Codex, Cursor and Claude Code. Turn rough requests into verified specs with Codex Spec Interviewer, reuse hand-curated ADRs with Architecture Compass, and index your codebase locally with CodeGraph + ast-grep. Plain Markdown you can read and adapt. Refined through 330+ commits and daily professional use. Clear scope and guardrails help reduce repeated discovery, save tokens and make faster models practical for scoped work. Free under Apache-2.0.
Hi Product Hunt! I'm Marcel. After 330+ commits and months of real-world refinement, I'm sharing Agent Skills: 10 free, open-source skills for Codex, Cursor and Claude Code.
ย
1. Codex Spec Interviewer
Turn a rough request into a user-verified spec with scope, acceptance criteria and validation commands. Our team and other senior developers on enterprise codebases use it multiple times a day.
ย
2. Architecture Compass
50 hand-curated, accepted ADRs (Architecture Decision Records), each with a short summary, full decision and practical guide. For example:
Keep secrets out of browser code.
Validate and authorize writes.
Test affected contracts and preserve mandatory checks.
ย
3. CodeGraph + ast-grep
Guided setup for local indexing, symbol and caller lookup, and structural code search. No hosted indexing service requiredโuseful where public MCP services are prohibited but approved local tools are allowed. Your agent and model still need to meet your organization's data policy.
ย
Why we use them
In our experience, specs and ADRs reduce repeated discovery, save tokens and make faster models practical for scoped implementation. Results vary by task and model.
ย
Evidence you can inspect
Evaluation cases and recorded results live in GitHub, outside installed skills to keep runtime instructions lean. Scenarios cover features, refactors, migrations and failure paths. Daily professional use provides feedback beyond the eval suite.
ย
Skills, ADRs and specs are human-readable Markdown. Adopt and adapt what fits your repository. Free under Apache-2.0.
ย
Get the skills โ
Which would help you most: clearer specs, architecture guardrails, or local code context?
๐๐ฎ๐๐๐ฒ๐๐ ๐ฎ๐ด๐ฒ๐ป๐ ๐ต๐ฒ๐น๐ฝ๐ฒ๐ฟ ๐ถ๐ป ๐๐ผ๐๐ป โฏ
โข ๐ฒ.๐ฐ๐ฌร as fast as native selection: 0.49 vs 3.12 seconds.
โ
Our Jev Capability Advisor uses TypeSafe Jev to recommend suitable skills and MCP tools to agents.
โ
Benchmark: choosing from 132 skills. Times shown are medians:
โ
โข ๐ญ.๐ณ๐ดร as fast as the best comparable skill we tested: 0.49 vs 0.87 seconds.
โข ๐ฐ๐ด/๐ฐ๐ด correct, accepted recommendations versus 46/48.
โข ๐ฒ.๐ฏ% fewer input tokens for selection than the tested alternative.
โ
Under the hood: HTTPS connection and search index reuse, compact selection requests, and up to 16,000 characters of task context.
โ
Same tasks and catalog, separate runs: GPT-6 Luna / medium for native selection, jev-1.13.0 for Jev. We measured selection only, excluding process startup, skill loading and task execution.
โ
๐ ๐๐ผ๐ฟ ๐ป๐ผ๐, ๐๐ฒโ๐ฟ๐ฒ ๐ฝ๐ฟ๐ผ๐๐ฑ ๐๐ผ ๐ฐ๐น๐ฎ๐ถ๐บ ๐๐ต๐ฒ ๐๐ถ๐๐น๐ฒ: ๐๐ฒ๐๐ ๐๐ธ๐ถ๐น๐น-๐ฐ๐ต๐ผ๐ผ๐๐ฒ๐ฟ ๐ผ๐๐ ๐๐ต๐ฒ๐ฟ๐ฒ ๐
โ
Now available as a standalone skill and in the stark AI Developer Plugin for Codex. Requires a TypeSafe API key; automatic use requires host integration.
โ
Code, installation and benchmarks:
https://stark-ai-de.github.io/agent-skills/skills/jev-capability-advisor/#benchmarks
โ
#AgentSkills #MCP #Codex #AIEngineering