bon-vet is a tool you invoke at decision moments. Instead of taking the first plan, verdict, or diagnosis an AI hands you, it has several competing answers written independently, puts them in front of a separate AI judge that doesn't know who wrote what, and shows you the winner — with the price told to you before a cent is spent, and a receipt after.
The whole idea in one picture. Everything else is detail.
When you ask an AI for a plan, you get one answer, delivered confidently. You usually have no way to know whether it's the best answer or just the first one. bon-vet's move is old and human: get independent opinions, hide the names, and let a neutral party score them against criteria you approved in advance.
Three things make it trustworthy rather than just fancy:
The judge is blind. Before judging, every draft is stripped of "I'm Claude" tells and shuffled, so the judge scores the writing, not the reputation.
The price comes first. Nothing is ever spent without showing you an honest estimate and getting your yes — twice, at two separate moments.
Everything leaves a receipt. Every run writes down what was compared, what it cost, and how decisive the result was. If the tool is ever wrong, the receipt is how you find out.
The pattern behind every good moment to use bon-vet: the decision is expensive to get wrong, and more than one reasonable answer exists. Some of these you'll feel; the last three are the ones worth training yourself to notice, because they don't feel like decision moments at the time.
"Here's what I want to do — get me competing plans and tell me which holds up." Several AI personas (a pragmatist, an adversary, a minimalist) each write a plan independently; the judge ranks them against criteria you approve first.
The signature move: answers come from two different AI families (Claude and GPT), and a judge from a third family referees. Use it when you'd otherwise be picking between vendors on vibes.
Ranks competing explanations of a bug — each one required to state its evidence, what would disprove it, and the next experiment to run. It deliberately stops short of writing fixes: wrong diagnosis, wasted fix.
You type the command with a short
description in quotes, e.g.
/bon-vet:vet plan "move my site's images to a CDN" — and then mostly
answer yes/no questions.
Every run follows the same arc. You are asked for permission at two clearly marked gates, and money can only move at the second one.
One or two sentences. The tool drafts judging criteria and shows you them — you approve or edit the criteria before anything else happens, and they're frozen from that point so nobody can move the goalposts mid-game.
Before any drafts are written you see the plan for the run: how many drafts, who writes them, and a first cost estimate for the judging. Nothing has been spent yet; saying no here costs nothing.
Each persona writes its answer independently — none of them see each other's work. Drafting rides your existing AI subscriptions, so this step is effectively free. Broken or empty drafts are caught and retried once; if fewer than two good drafts exist, the run stops honestly instead of pretending.
Now the real drafts exist, the price is recomputed from their actual size and shown as an honest range — typically well under a dollar — along with exactly where the text will be sent to be judged. This is the only moment spending can start, and only after your yes.
Names stripped, order shuffled, every pairing scored several times against each criterion. If anything goes wrong mid-judging — a rate limit, a bad response — the run stops and says so rather than quietly limping to a fake answer, and re-running resumes from where it stopped without re-buying finished work.
You get the winner, the full ranking, one honest line about what the runner-up did better, and the estimate-versus-actual cost. A receipt folder keeps the whole record.
| You'll see | What it means for you |
|---|---|
| Winner + ranking | The judge's pick and how everyone placed. Scores close together mean a near-thing; far apart means a clear call. |
| The dissent line | One sentence on where the runner-up beat the winner. This is the most useful line in the report — it often names the thing worth stealing from second place before you proceed. |
| "No clearly separated winner — indeterminate" | An honest tie. The tool refuses to fake confidence; it offers to judge more thoroughly (priced separately) or hand the final call to you. A tie between two good plans is information: either is probably fine. |
| "Data leaves machine: yes → …" | Plain disclosure of where your text goes to be judged (normally Google's Vertex service). If the work is too sensitive for that, say so — there's a private, on-machine judging option to set up instead. |
| Warnings in capital letters | Rare, and always mean "the judging quality itself is suspect." The tool would rather cancel a run than sell you a noise-based verdict. |
bon-vet is in its shake-down phase. After each run on this machine it will ask you two things: do you agree with the winner (and if not, which draft you'd have picked), and did anything feel off. Answering takes under a minute.
This isn't a survey for its own sake — your answers are the calibration data that decides when version 1.0 is actually earned: they check the judge against human judgment, tune the cost estimates, and surface friction. About a week after a run, it's also worth a one-line check-in on whether the winning plan actually held up in practice.
~/.bon-vet/dogfood). It is structurally unable to end up in the
public project, on GitHub, or anywhere else — that was a design requirement,
not an afterthought.
# from whatever project you're working in:
claude --plugin-dir ~/Developer/tooling/bon-vet/bon-vet:vet plan "what you're deciding"
/bon-vet:vet arb "the question two AIs disagree on"
/bon-vet:vet diagnose "the bug + what you've observed"cd ~/Developer/tooling/bon-vet
.venv/bin/python bin/setup.py doctorThe doctor prints a plain table of what's working and names the next step for anything that isn't. It never claims "ready" unless it truly is.
If getting it wrong costs hours and there's more than one reasonable answer — vet it. Otherwise, carry on.