Start from a sentence — or your own data

Stop measuring prompts.
Start finding them.

Eval tools tell you your prompt scores 72%. findmyprompt hands you the one that scores 91% — at half the cost. Give it a labeled dataset and a budget; it searches the prompts and model configs you'd never test by hand.

Start a search →See how it works

Searches across every model

OpenAI Anthropic Google Meta Mistral DeepSeek xAI Amazon

1Start — or resume — your problem

Classify each incoming support ticket into one of six categories, from the subject and body.▌

Input / output shape

input  { subject: text, body: text }

output  { category: enum(6) }

You never write a prompt. findmyprompt generates and optimizes the prompts itself — your description is the goal, not the prompt.

Budget$ 20stops when spent, resumes free

Find configurations →

2Configurations we found

● searching

$/1k reqs ↑

$1.70

$0.80

84%

87%

90%

gpt-4o

mistral-large/ 1

llama-8b

gpt-4o-mini/ 3

gemini-flash

llama-70b/ 2

claude-haiku/ 1

claude-sonnet

accuracy →

frontier (ringed)dominated — faded

● searching — evaluating configurations…

fig. 1 — one problem, solved eight ways · illustrative

Each pill is a prompt we wrote, run on that model — click one to inspect its full config in the loop below. Illustrative.

A loop, not a one-shot

Finding the config is step one. Staying optimal is the loop.

The config you just found doesn't ship once and rot. It enters a cycle — and because we search only the differences, past work replays free. Staying on the frontier is cheap, not a restart.

Nothing ships silently — every promotion waits for your approval. The search is a one-time spend; the runtime saving is forever.

Cross-model by design

We have no model to sell you.

A model lab will never optimize you onto a competitor's model. We mix every pigment on the palette and keep whichever wins on your data — even a cheaper one.

−61%

inference cost

+19 pts

accuracy

3 hrs

search pays back

$ fmp search --describe "classify tickets" --budget 20
→ writing prompts · searching 9 models · racing on cost…
✓ best: claude-haiku · 91% · $0.0004/call

Two costs, honestly: the search is a one-time investment (billing-exact, capped, resumable); the runtime saving is forever.

The primary search — MCTS × genetic

MCTS searches the space; genetics evolve the prompt.

The primary search is a hybrid. MCTS grows a tree over the whole configuration space — model, parameters, structure — and UCB sends each new rollout down the branch that's winning, starving the rest, so the budget lands where it pays off. At the config it settles on, a genetic search breeds and mutates the prompts.

Within each rollout, successive halving keeps it cheap — a quick probe first, full evaluation only for the survivors — so a fixed budget covers a huge prompt × model × config space.

Better than hand-written

The prompt you'd write vs the one we found.

Same task — your hand-written prompt and the one findmyprompt wrote, both scored on your data.

you wrote

Classify this ticket:
{{subject}} {{body}}

72%$0.80 / 1k · gpt-4 class

we wrote+19 pts · 2.1× cheaper

You triage support tickets. Read the subject and body, weigh intent over keywords, and return exactly one of: billing, bug, account, feature, refund, other. When ambiguous, prefer the user's primary ask.

91%$0.38 / 1k · a smaller model

No black-box scoring

Every score comes with a receipt.

The judge is built from your own labels — then shows exactly how well it agrees with you on held-out examples.

Built from YOUR labels — your notion of “good” is the ground truth.
Held-out agreement, sample size, and every disagreement, in the open.
No confidence-interval theater — honest numbers, even at small N.

JUDGE AGREEMENT
held-out audit · 48 items
agreement94%
bootstrap band±3 pts
sample size48
DISAGREEMENTS (7)
#12 billingyou 1 / judge 0
#27 refundyou 0 / judge 1
#41 bugyou 1 / judge 0
… +4 shown in app
signed by YOUR labels

You bring the problem. We bring the rest.

The data, the prompts, the model, and a config you can paste into production.