ai-eval-plan

AI Eval Plan Skill

You can't improve an AI feature you can't measure, and "it looks good in the demo" is not measurement.

This skill produces an evaluation plan that turns a fuzzy quality goal into a repeatable, gated test —

so a prompt change that quietly makes outputs worse can't ship.

Required Inputs

Ask for these only if they aren't already provided:

• The feature & task — what the model does and what "good output" means to a user.
• Failure modes that matter — what bad looks like (hallucination, wrong format, unsafe, off-tone, too slow).
• Available data — any real examples, logs, or labelled cases; or note there are none yet.
• Who judges quality — automated checks, an LLM judge, human raters, or a mix.
• The decision this gates — ship/no-ship, model selection, or prompt iteration.

Output Format

Eval Plan: [feature]

1. What we're measuring — the task, and a one-line definition of a good vs. bad response.

2. Eval dataset

• Cases: how many, where they come from (real logs > synthetic), and how they're split (smoke set vs. full set).
• Coverage: the slices/scenarios that must be represented (edge cases, adversarial, each major input type).
• Golden answers / references: present or not, and how they were created.

3. Metrics & rubric

• Per-dimension scores — define each dimension (e.g. correctness, grounding, format, safety, tone) on an explicit 1–5 rubric with anchor descriptions, not vibes.
• Automated checks — deterministic assertions first (valid JSON, contains required fields, no PII, latency budget).
• LLM-as-judge — the judge prompt, the rubric it applies, and how you guard against its bias (calibrate against human labels on a sample).
• Human eval — when it's required (safety, subjective quality) and the rater instructions.

4. Baselines — what each candidate is compared against (current prompt, previous model, a plain-prompt control).

5. The bar — the explicit threshold to ship (e.g. "≥4.2 avg correctness, 0 safety failures, p95 < 3s") and what happens if it's missed.

6. Regression gate — how this runs in CI on every change, and the score-drop threshold that blocks a merge.

Quality Checks

• [ ] Each metric has an explicit rubric with anchors — not just a name
• [ ] Deterministic/automated checks are used wherever possible before reaching for an LLM judge
• [ ] The LLM judge is calibrated against human labels on at least a sample
• [ ] The eval set includes adversarial and edge cases, not just happy-path examples
• [ ] There is a single, explicit numeric bar for the ship decision
• [ ] The plan specifies how it runs as a regression gate, not just a one-time check

Anti-Patterns

• [ ] Do not rely on a single overall score — a feature can pass on average while failing every safety case
• [ ] Do not trust an LLM judge you haven't calibrated against humans — it has its own blind spots and biases
• [ ] Do not eval only on happy-path inputs — the failures live in the edges and the adversarial cases
• [ ] Do not let the eval set leak into the prompt/few-shot examples — that's training on the test set
• [ ] Do not define the pass bar after seeing the scores — set the threshold before you run, or it means nothing

Based On

LLM evaluation practice — task-grounded rubrics, LLM-as-judge with human calibration, and regression-gated CI evals.