The annoying part of AI coding tools
Every session, the agent forgets:
- your module boundaries
- the “every user belongs to an org” rule that lives in nobody’s file
- which service emits which event
- how you want features built (tests first? spec first? conventions?)
So you re-explain. It rewrites. Context drifts. Tokens burn.
Dumping a 40k-token AGENTS.md into every prompt is not a strategy. It’s a tax.
I wanted something simpler: write project knowledge once, in a format agents already understand, and install it into every tool I actually use.
That’s cast-skills.
npx cast-skills
What it is
cast-skills is a small CLI that bootstraps cross-tool AI skills for your machine (or project):
| Skill | Job |
|---|---|
using-project-skills |
Always-on router. When you start a feature/fix/refactor, it loads project context and routes work through the right module + workflow. |
creating-project-skills |
Bootstrapper (/skill-creator). Reads the repo, asks sharp questions, generates a project skill tree. |
explaining-changes |
After non-trivial work, offers a beginner-friendly explanation (ELI5 → runtime) as Markdown/HTML. |
Skills use the open SKILL.md format — progressive disclosure: the agent loads what it needs, not your entire wiki every turn.
cast-skills is built around Anthropic's Agent Skills approach: instead of loading one massive prompt every session, the agent progressively loads only the knowledge it needs.
One install. Six tools.
The wizard auto-detects what you have and drops skills in the right place:
| Tool | Where it goes | Format |
|---|---|---|
| Claude Code | ~/.claude/skills/ |
native SKILL.md
|
| GitHub Copilot | ~/.copilot/skills/ |
native |
| Gemini CLI | ~/.gemini/skills/ |
native |
| Codex | ~/.agents/skills/ |
native |
| Cursor |
.cursor/skills/ (project) |
converted .mdc
|
| Windsurf |
.windsurf/rules/ (project) |
converted rules |
Native tools get SKILL.md as-is. Cursor/Windsurf get an auto-converted format so you’re not maintaining six copies by hand.
npx cast-skills
# or
npm i -g cast-skills && cast-skills
Requires Node ≥ 18.
How a normal feature run feels
- You say: “add X / fix Y / refactor Z”
-
using-project-skillsloads domain + stack + conventions - It pulls the module-level skill for the code you’re touching
- Work is steered through a spec-driven path (spec → failing tests → implement) when that workflow exists
- After a real change,
explaining-changescan leave a human-readable write-up
The point isn’t another chatbot persona. It’s stable project memory that survives sessions and tools.
Why I didn’t make this Claude-only
Most of us don’t live in one harness.
Monday: Claude Code.
Tuesday: Cursor on a design-heavy UI.
Wednesday: Copilot in a boring enterprise repo.
If “project brain” only works in one tool, you still re-explain on Tuesday.
cast-skills optimizes for write once → run everywhere that speaks skills/rules.
What this is not
- Not a full coding agent (that’s a different product surface)
- Not magic RAG over your monorepo (skills are curated knowledge, not automatic embeddings)
- Not a replacement for good architecture — garbage conventions in → garbage guidance out
If your team has no conventions, the bootstrapper will still force a few useful questions. That’s a feature.
Try it in 30 seconds
cd your-repo
npx cast-skills
Pick the tools you use → let it install the three skills → start a real task and see if the agent stops inventing your architecture from scratch.
Repo: github.com/pedrocastanha/cast-skills
npm: npmjs.com/package/cast-skills
Feedback I actually want
If you try it, tell me:
- Which tool combo you use (e.g. Claude + Cursor)
- Whether the wizard put files in the right paths
- One rule your team has that should be a skill but isn’t obvious
Issues and PRs welcome — especially path fixes for tool updates (these CLIs move the goalposts often).
Top comments (0)