Metis

Metis is a coding skill for LLMs that improves engineering judgment during implementation, refactoring, and code review.

What it pushes models toward

• plain-data domain models over behavior-heavy object hierarchies
• top-down API design and call-site-first thinking
• explicit control flow
• boundary assertions instead of implicit assumptions
• behavior-focused testing
• early performance thinking
• sparse, useful comments instead of code narration
• tiered code review: sequential lens passes by default, sub-agent fan-out for large diffs

These are defaults, not laws. Metis is opinionated, but it should improve the local codebase, not fight it.

Measured results

Metis is eval-driven: every non-trivial rule was validated or discovered through isolated baseline-vs-Metis runs (Codex gpt-5.5, xhigh reasoning, clean sandboxed home, 60+ runs). Highlights from the 2026-07 eval round:

• Code review, seeded ground truth (evals/cases-v4/): PR-review cases with known planted defects, scored on objective recall and precision. Metis found ~60% more defect value than baseline on the tuned case (11.3 vs 7.0 avg, max 15.5) and +53% on a held-out case in a different domain (12.7 vs 8.3, max 19.5), winning or tying every rep.
• Zero false positives in 39 consecutive review runs across all conditions — the verify-before-reporting rule holds under pressure.
• Five real unplanted bugs discovered by Metis conditions across the two review cases (e.g. refund ledger rows indistinguishable from charges; an event marked processed before its transaction commits). Baseline found one.
• Design cases (evals/cases-v3/): Metis beat or tied baseline in 9/9 case-reps; its one reproducible defect (applying batch elements before whole-batch classification) became design principle 10, which flipped that failure from 1/3 correct to 3/3 on reruns.
• Changes that failed their A/B (e.g. a mandatory per-lens ledger) were rejected, not shipped.

Methodology, rubrics, and per-run scoring live in evals/; raw run output stays git-excluded.

Install

Install from GitHub with:

npx skills add https://github.com/gitRasheed/metis-skill --skill metis -g

Install only for a specific agent if you want:

npx skills add https://github.com/gitRasheed/metis-skill --skill metis -g -a codex
npx skills add https://github.com/gitRasheed/metis-skill --skill metis -g -a claude-code

If symlinks are blocked on Windows, add --copy.

npx skills add https://github.com/gitRasheed/metis-skill --skill metis -g -a codex --copy

Use

Metis is meant to auto-load for relevant coding work.

You can also invoke it explicitly with $metis or /metis, depending on the agent UI.

Example prompts:

Use /metis to refactor this module.
Use $metis to review this design.
Use /metis to implement this feature.

What is in this repo

• skills/metis/ is the main skill and the canonical copy
• .claude/skills/metis/ mirrors it for Claude Code with adjusted reference paths; when editing the skill, regenerate it with sed 's|\references/|`../../../skills/metis/references/|g' skills/metis/SKILL.md > .claude/skills/metis/SKILL.md`
• PORTABLE_PROMPT.md is the plain markdown version for tools that do not support skills directly
• evals/ contains isolated baseline-vs-Metis eval cases and a local runner; raw eval runs stay git-excluded