An eval is a repeatable test that grades an AI output against a rubric. You define what "good" looks like in 5–8 lines, you collect a fixed set of inputs you care about (a "golden set" of 20–50 examples), you run the AI on each, and you score the outputs against the rubric. Run it weekly and you know whether your AI is getting better, worse, or drifting.

Where you'll hear it: Anthropic's Outcomes (rubric-grading at inference time, May 6). OpenAI's Evals platform (and PromptFoo, which OpenAI acquired in March — it's still MIT-licensed). Braintrust — the eval-first observability platform that raised $80M Series B in February. Inspect AI — the UK government's open-source eval framework now used by Anthropic and DeepMind on frontier-safety work. And every weekly meeting where someone says "the AI used to be good at this and now it isn't" — what they mean is we don't have an eval, so we noticed too late.

Pick the AI task you re-run most often during a normal week — the weekly client status update, the candidate-rejection note, the meeting prep brief, the cold-email rewrite. Open a fresh doc and write a five-line rubric for it. Each line is a yes/no criterion: Does the output name the client's specific request? Yes / No. Does it cite at least one data point? Yes / No. Is the closing question specific (not "thoughts?")? Yes / No. Is the tone match within one register of the client's last email? Yes / No. Is the length under 180 words? Yes / No. Make every line concrete enough that a stranger would score it the same way you would. Now for one week, every time you run that task, copy the rubric under the AI output and score yes/no. You're not changing the prompt. You're keeping a log. By Friday you'll see one of two patterns — either the same line is failing every time (the prompt has a fixable hole) or different lines fail randomly (your golden set is too small and you need three more examples). Either way, you know exactly what to do next.

Why it works: The fundamental failure of one-shot AI use isn't bad prompts. It's invisible bad prompts — you noticed the output was wrong this week but you can't articulate which part. Five concrete yes/no lines force the failure to name itself. Once it has a name, you can fix the prompt with one edit instead of three rewrites. This is the same trick Anthropic's Outcomes feature applies at inference time — the rubric narrows the model's degrees of freedom. You're applying it manually for one task. The skill compounds: the second rubric takes half as long to write as the first.

Pick a role you've done or watched closely — recruiter, paralegal, real estate agent, freelance editor, account manager, school principal, junior PM, financial advisor. Build a single PDF or Notion page called something like "Rubric kit for [role] — grade the AI before you send it." Inside, put four things: (1) eight rubrics, one for each recurring AI deliverable in that role (a recruiter's outreach email, a paralegal's deposition summary, a real-estate agent's listing description); each rubric is 5–8 yes/no criteria, concrete not generic ("names the deadline" beats "is clear"); (2) two worked examples per rubric — one passing, one failing — with the exact line that failed marked; (3) a one-page "when to use LLM-as-judge vs. human" guide for that role specifically — paralegal work has different lines that need human judgment than marketing copy does; (4) a short setup script for running the same rubric weekly in a Google Doc, a Notion database, or a free PromptFoo config. Price at $39–$79 on Gumroad or Lemon Squeezy. Position it plain: "You don't need a smarter AI. You need to know whether the one you have is good enough to send."

Why it works: The market is awash in prompt packs and getting cheaper every month — most of them have the same Claude/ChatGPT prompts repackaged. Rubrics are the inverse of prompts in market position: they're what you reach for when you've already tried the prompts and the output still feels off. Almost no one's productizing them for non-developers. The buyer pays you because you've done the role-specific work of figuring out what "good" looks like for their job — which is exactly the skill stack from Section 03, packaged. Realistic revenue: $500–$2,500/month with no audience; $2,000–$6,000/month if you DM it to your existing LinkedIn network in that role. The timing window is roughly two quarters before this becomes a saturated category — exactly the way prompt packs did in 2024.

Braintrust

The managed-eval option for teams that don't want to host the harness themselves. Braintrust raised an $80M Series B in February at an $800M valuation, led by ICONIQ with a16z, Greylock, and Elad Gil — the customer roster includes Notion, Replit, Cloudflare, Ramp, Vercel, and BILL, which tells you who's already running this in production. The reason it belongs in this issue: Braintrust does for evals what Datadog did for server metrics. You instrument the agent call, you ship the rubric, you get a dashboard that tells you which prompt regressed when. Free tier is enough to run your first weekly eval. Worth a Tuesday afternoon to see if the trace view fits how you work.

PromptFoo

The first tool I install on any new machine for this work. PromptFoo is a YAML-config CLI — you write your prompts and your test cases in a config file, you run `promptfoo eval`, and you get side-by-side comparisons across GPT, Claude, Gemini, DeepSeek, and any local model. OpenAI acquired it in March 2026 and it stays MIT-licensed; the cited usage covers a quarter of the Fortune 500 plus internal use at both OpenAI and Anthropic. The reason to know about it even if you don't write code: a non-developer can run `promptfoo init` against an OpenAPI key and have a working eval in ten minutes. The first hard part is the rubric (which is the actual skill from Section 03). The second hard part is keeping the test set fresh. PromptFoo handles everything in between.

Inspect AI

The serious option for anyone whose evals need to hold up under regulatory scrutiny. Inspect AI was built by the UK AI Security Institute and is now used by Anthropic and Google DeepMind for frontier-safety evaluations. The library ships with 200+ pre-built evaluation tasks and runs against every major API plus local models. The reason it belongs in this issue's stack: the style of eval Inspect AI enforces — datasets, scorers, sandboxed execution, reproducible logs — is exactly what auditable AI work looks like in regulated industries. If you work in healthcare, legal, financial services, or anything else where someone might subpoena the eval methodology later, the choice between Braintrust and Inspect is roughly "managed/proprietary vs. self-hosted/audit-grade." The interoperability documentation is the polite way the EU is signaling that this is the eval framework public sector buyers will trust. ---