agentleFS
nfeuernfeuer/project-foundation-kit/.claude/skills/eval-harness/SKILL.md

eval-harness

Skill0 starsChanged 4 months agoMITView on GitHub

Set up and run tiered, version-controlled evaluation fixtures for non-deterministic behavior (LLM calls, classifiers, extractors) — score against pass-gates and catch regressions in CI

77 other projects use a version of this.

Adopted by

Most recently taken up by

Used with

Projects with it also use brand-discovery, agent-sort and 10 more.

Other files people use alongside this one.

Files found in the same public projects as a version of this one, more often than they turn up anywhere else. Each project counts once, large collections count for less, and this file's own repository is left out.
  • ysyecustSkill

    brand-discovery

    Use when a brand needs to discover or articulate its identity through structured multi-session interviews. Covers purpose, positioning, audience, personality, voice, narrative, and founder-brand tension across 8 modules using laddering, 5 Whys, and projective techniques. Produces a resumable session with disk-persisted state and a master brandbook (90_SYNTHESIS.md).

    Also in 6 projects that use this

  • niloykumarbarmanSkill

    agent-sort

    Build an evidence-backed ECC install plan for a specific repo by sorting skills, commands, rules, hooks, and extras into DAILY vs LIBRARY buckets using parallel repo-aware review passes. Use when ECC should be trimmed to what a project actually needs instead of loading the full bundle.

    Also in 14 projects that use this

  • royaluniondesign-sysSkill

    configure-ecc

    Interactive installer for Everything Claude Code — guides users through selecting and installing skills and rules to user-level or project-level directories, verifies paths, and optionally optimizes installed files.

    Also in 12 projects that use this

  • junimnjwSkill

    ralphinho-rfc-pipeline

    RFC-driven multi-agent DAG execution pattern with quality gates, merge queues, and work unit orchestration.

    Also in 12 projects that use this

  • loulanyueSkill

    nuxt4-patterns

    Nuxt 4 app patterns for hydration safety, performance, route rules, lazy loading, and SSR-safe data fetching with useFetch and useAsyncData.

    Also in 10 projects that use this

  • royaluniondesign-sysSkill

    content-hash-cache-pattern

    Cache expensive file processing results using SHA-256 content hashes — path-independent, auto-invalidating, with service layer separation.

    Also in 16 projects that use this

The file

We split this skill into the rules it gives an agent

3 rules. 0 are common, 1 is distinctive, and 2 are very specific to this project.

  • 0

    Common

    Most of the 71 edited versions we read keep these.

  • 1

    Distinctive

    This author's own choices. Worth a look.

  • 2

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Use it when adding or changing any capability whose output is judged rather than asserted.

    line 22
  2. Very specific:

    Write the expected outputs for the cases you care about.

    line 47
  3. Very specific:

    Run tiers 1–2 (the gated ones) in CI on the not slow and not llm marker set using mocked tools, so the suite is fast and free.

    line 67

“Common” is measured: kept by at least 6 in 10 of the 71 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Eval Harness

Applies when capabilities.llm.enabled is true (else skip, report N/A).

Profile-driven. The fixture root directory is capabilities.llm.eval_dir from .claude/kit.yaml (default: fixtures).

Unit tests check deterministic code. This skill covers the parts that don't have one right answer — anything backed by a model, a classifier, or a fuzzy extractor. The pattern: version-controlled fixtures, tiered by difficulty, scored against a pass-gate, run offline with mocked tools, and gated in CI.

Use it when adding or changing any capability whose output is judged rather than asserted. This skill authors, scores, and reports against fixtures; it never edits the prompt, model config, or code under test — those changes are a separate, explicit step.

Structure

Fixtures live at <eval_dir>/<task_type>/tier<N>_<name>.json, one file per tier (where eval_dir = capabilities.llm.eval_dir from kit.yaml, default: fixtures). See templates/eval_fixture.example.json for the shape. Each file declares a pass_gate (fraction of cases that must pass) and a list of cases, each with an input, an expected output, and optional tool_mocks for offline runs.

Tiers

  • Tier 1 — baseline. Simple, unambiguous. pass_gate: 0.90. A regression here blocks the merge.
  • Tier 2 — nuance. Domain-specific judgment calls.
  • Tier 3 — complexity. Compound / multi-part inputs.
  • Tier 4 — adversarial. Contradictions, buried tasks, edge cases. pass_gate: null — diagnostic only, tracked but not gated, so a hard-case failure doesn't block shipping while a baseline regression does.

Workflow

1. Author fixtures before (or with) the capability

Write the expected outputs for the cases you care about. For anything that calls a tool (calendar, email, search), add tool_mocks keyed by tool fingerprint so the eval runs deterministically offline — no live API calls, no spend, no flakiness.

2. Score with schema validation first, then correctness

Validate each output against its JSON schema (schemas/<task_type>.json) — a malformed output is an automatic fail. Then compare against expected. For fuzzy fields (free text), score with an LLM-judge rather than exact match: send the prompt + output to a current model ID selected via the project's model aliases (e.g. claude-sonnet-5; never a retired claude-3-* ID) and take its quality_score (0.0–1.0), flagging anything below threshold.

3. Compute the tier result

passed / total >= pass_gate  →  tier PASS

Report per-tier and per-case so a regression points at the exact case.

4. Gate in CI

Run tiers 1–2 (the gated ones) in CI on the not slow and not llm marker set using mocked tools, so the suite is fast and free. Run the full set (including live-model tiers) on a schedule or a label, not on every push.

When you change a prompt or model

Re-run the harness before and after. A prompt tweak that helps one tier can regress another — the fixtures are how you find out before your users do. Record notable score deltas in docs/followups.md if you accept a tradeoff.

Output

## Eval Run — <task_type>
| Tier | Cases | Passed | Gate | Result |
|------|-------|--------|------|--------|
| 1 baseline | 20 | 19 | 0.90 | PASS |
| 2 nuance   | 15 | 11 | 0.80 | FAIL |
Failing cases: <ids + why>
Verdict: PASS / REGRESSION — <detail>

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.