write-evals

name	write-evals
description	Use when the user wants to generate behavioral evals, LLM-judge tests, or eval cases for a skill, agent, prompt, or feature, write a failing eval before implementing, or scaffold an eval harness in a repo that has none

Write Evals

An eval is a failing test for one observable behavior: a fixture, a prompt, and machine-checkable assertions, with an optional LLM judge for behaviors substrings can't express. The invariant this skill protects: write the eval red first — a case that can't fail proves nothing. If the repo has no harness, scaffold one before writing cases.

When to Use

Do not use this skill for ordinary unit/integration tests with no model in the loop — use the project's normal test tooling. Do not use it to implement the feature; this skill writes the eval and hands off.

Process

1. Locate or scaffold the harness. Search the repo for an existing eval runner and spec dir (*.evals.json, evals.json, a run-evals script, a test:evals package script). If one exists, read a sample case and mirror its schema — never introduce a second eval format. If none exists, scaffold from run-evals.template.ts: copy it to evals/run-evals.ts, add "test:evals": "bun evals/run-evals.ts" to package.json, and create the specs dir. The harness needs Bun and the claude CLI; for a non-Claude system-under-test set EVAL_AGENT_CMD.
2. Pin one behavior. State the target behavior in a single sentence (expected_output). One observable behavior per case; if you're testing two things, write two cases.
3. Build a minimal fixture. Seed only the files the behavior needs into fixture.files. Small and real beats a mirror of the repo.
4. Write deterministic assertions first. Use required_files, required_file_substrings, required/forbidden_substrings, unchanged_files. Split a two-part requirement into two assertions so a half-answer fails. Add expectations (LLM-judged) only for behaviors substrings can't capture, e.g. "reads the repo before asking". When the whole behavior is "which skill/route did it pick", use kind:"routing" with a routing block (expect/forbid) instead — code-graded, judge-free, and scored by strict-majority over trials. See eval-spec.md for the full schema and a worked example.
5. Confirm it's red. Run only the new case (the AFK harness filters via AFK_EVAL_ID=<id>; the template via EVAL_ID=<id>) and verify it fails for the right reason — the behavior is absent, not the fixture or assertion malformed. Carve-out: coverage cases that lock in already-correct behavior — negative gate twins, edge-case classes, routing-volume cases — may be born green; you can't make passing behavior fail. State this in the PR/report and rely on review to catch dead assertions, rather than faking a red. Net-new behavior still goes red first.
6. Go green. Implement the behavior (or hand to the owning skill/agent), then re-run until deterministic assertions pass and the judge clears the threshold.
7. Read the judge samples. Open a judge*.json artifact for at least one judged case and confirm the <thinking>-then-JSON output parses and the verdicts are sane — a judge that misreads the transcript silently inverts the gate. Routing cases need no judge; check their per-trial route log instead.
8. Keep zero-token checks green. Make sure the spec JSON parses and any harness you scaffolded runs (bun <runner> with no specs should fail loudly, not crash).

Stop and Ask

STOP and ask the user when:

• The behavior to test, or which output is authoritative, can't be inferred from the repo or the request.
• The system-under-test isn't the claude CLI and how to invoke it (EVAL_AGENT_CMD, a binary, an HTTP call) is unknown.

Do not ask about facts discoverable by reading the repo's existing evals.

Red Flags

Thought	Reality
"I'll write the eval after the feature works."	Then it can't go red — you've stamped a regression, not tested. Write it red first.
"One required substring is enough."	A two-part requirement needs two assertions, or a half-answer passes.
"Judge everything."	Substring/file assertions are deterministic and free to re-run; reserve the LLM judge for what they can't express. A pure route check is a kind:"routing" case — no judge at all.
"This repo needs its own eval format."	Reuse the existing harness; a second format splits the suite.
"The fixture should mirror the real repo."	Minimal fixtures isolate the behavior and run fast.
"More cases is always better — add variations."	Add volume that mirrors the real request distribution (gate twins, edge classes, adversarial routes), not broad variation-spam that just re-tests the happy path.

Output

Report:

• The spec file path and the new case id(s).
• The exact run command (e.g. AFK_EVAL_ID=<id> bun run test:evals).
• Whether the case is currently red (pre-fix) or green, and — if scaffolded — the harness files created.

References

• run-evals.template.ts — self-contained Bun harness to scaffold when none exists
• eval-spec.md — eval case schema and worked red-first example