optima

Files (1)

• SKILL.md

SKILL.md (first lines)

# Optima Skill

## Overview

Apply production AI/token optimization without sacrificing correctness.

## When to use

- User asks to save tokens, reduce cost, or stop verbose agent output
- Long sessions approaching context limits
- Setting up Cursor/Claude/Codex FinOps rules
- Designing API prompts for prompt caching

## Levels

`/optima off|lite|on|ultra`

- **off** — no diet
- **lite** — compress chat only
- **on** — default FinOps (answer-first + surgical context)
- **ultra** — telegraphic chat; never telegraph code/docs/tests/IDs

## Process

1. Confirm diet level and zone (sacred wins for code/docs).
2. Prefer grep → targeted read → batch tools.
3. Respect ignore boundaries.
4. For API work: stable prefix, cache breakpoints, lint for timestamps.
5. Recommend taxonomy techniques from `@optima/classify` when coaching.

## Buckets

| Bucket | Rule |
|--------|------|
| A Replies | Answer-first |
| B Docs/plans | Dense |
| C Tests | Key + edge; never skip money/auth/data-loss |
| D Code | YAGNI, idiomatic — not cryptic |
| E Context | Grep-before-read, line ranges, min turns |
| F Tools | Batch, stop early |
| G Sub-agents | Cheap bounded search; strong model verifies |

## Rationalizations to reject

| Excuse | Response |
|--------|----------|
| "I'll dump the whole repo to be safe" | Surgical retrieval is safer and cheaper |
| "Ultra mode on the PRD" | Premium/sacred content stays precise |
| "Skip auth tests to save tokens" | Forbidden — quality floor |

## Verification

- [ ] Code/docs remain precise
- [ ] No critical tests skipped
- [ ] Context from ignores avoided unless required
- [ ] Zone 3 did not pollute Zone 0

From julienlhk/optima, MIT.

More from julienlhk/optima