Imported from lantisprime/pi-extensions (
AGENTS.md). Install upstream withnpx skills add lantisprime/pi-extensions. Copyright stays with the author.
Agent instructions
TypeSafe skill
This repo ships the typesafe-ai skill at .agents/skills/typesafe-ai/.
Consult it when working on this project. It is third-party vendor guidance (MIT, © TypeSafe) for building with TypeSafe's System One models, including Jev — small units of AI intelligence that return typed answers and probabilities instead of generated text, so ordinary code can consume them.
It is most relevant when a task involves:
- a feature that needs programmable common sense or semantic understanding
- an LLM prompt-and-parse step that could become a structured decision
- routing, ranking, extraction, verification, or grading
- choosing or composing typed judgments:
Choice,Noul,Score
Docs are the source of truth
The skill gives direction only. The live docs carry current concepts, API
contracts, SDK usage, models, and limits — read them as part of the task, starting
from https://docs.typesafe.ai/llms.txt. Mintlify serves Markdown by appending
.md to a page path. Do not invent version-dependent API details; if live access
is unavailable, say so and prefer installed SDK types over guesses.
Keep API credentials server-side in any web app.
Scope
This is vendor guidance, not a repo utility. Apply it where semantic judgments genuinely help the task at hand — it does not override this repo's own conventions, and it is not a reason to introduce TypeSafe into work that does not call for it.
Skills layout
.agents/skills/— project skills loaded by pi and Codex (episodic-memory,typesafe-ai). Tracked in git.skills/— this repo's own agent discipline packs, symlinked into~/.pi/agent/skills/(e.g.skills/tasks/)..pi/skills/— gitignored; local-only installs.
See README.md for the full extension catalogue.