Plate 1 · nibble · in build
Built on Kev · games first

Typed answers to your agent's multiple-choice questions

Which template? Which art style? nibble, a small decision model, picks from your options, says how sure it is, and aims to beat a big AI model on speed.

Watch it get built Read the plan

In buildIn build. Follow along.

In
situation + options, JSON
Out
pick · odds · time
Runs on
12 GB graphics card or CPU
Base
Kev, Apache-2.0
Fig. 1 · One question, one pickExample numbers
How nibble answers one question An example. The agent sends the prompt "A frog who fishes by lantern light" and the question "Which template?" with three options. nibble scores them and returns top-down at 81 percent, cozy at 12 percent and arena at 7 percent. A dashed line at 70 percent marks a rule: picks below it go to the big model. State A frog who fishes by lantern light. Question · choose one Which template? top-down cozy arena nibble scores every option Decision record · pick, odds, time top-down 81% cozy 12% arena 7% 70% rule 1 2 3 4 How nibble answers one question An example. For the prompt "A frog who fishes by lantern light" and the question "Which template?", nibble returns top-down at 81 percent, cozy at 12 percent and arena at 7 percent. A dashed line marks a 70 percent rule. State A frog who fishes by lantern light. Question · choose one Which template? Decision record top-down 81% cozy 12% arena 7% 70% rule 3
  1. Askthe situation and your options
  2. nibblescores every option
  3. The pickwith odds for every option
  4. Your ruleunder 70% sure, ask the big model
Fig. 1. Example numbers, not a measurement. nibble isn't built yet.

01 · The problem

Your agent asks a big model lots of small questions, like which template to start from or which art style fits. Each one waits on a written reply that never says how sure the model was.

02 · What you get

Benefits

2.1Pick

A pick, with no essay

nibble scores every option you give it and hands back a pick, with no paragraph to wait for.

Example A frog who fishes by lantern light. Which template?
top-down 81% · cozy 12% · arena 7% (example numbers)
2.2Confidence

Know when to double-check

Every pick comes with a calibrated confidence score, so you decide which ones go to a person or a bigger model.

ExampleSet a rule like under 70% sure, ask the big model. Easy picks stay fast, and only the unsure ones cost more.
2.3Train

Train a decision model on your agent logs

Turn choices logged in past runs into a small model of your own that runs on a home graphics card or a CPU.

ExampleTrain on your logged game builds, then run the result on one 12 GB graphics card.
03 · How it works

How it works

  1. Ask your question.Send the situation and your options as JSON (structured text programs read).
  2. Read the pick.You get the pick, the odds for every option, and how long it took.
  3. Train your own.Point the trainer at decisions logged from past runs, and get a smaller decider tuned to your task.
04 · See it work

See it work

Target demo. Not built yet.

You type "a frog who fishes by lantern light" into playdough, the game maker in this set. Before writing any code, playdough must choose a starter template and building blocks, like a walking character, a coin and night lighting.

It asks nibble, which hands back each pick with how sure it is and how long it took.

Then a big model gets the same questions, on prompts neither trained on. nibble passes if it is faster and picks at least as well.

Fig. 2 · The testTarget design · not yet built
The proving test playdough sends the same questions to nibble and to a big model, on prompts neither trained on. nibble passes if it is faster and picks at least as well. Held-out prompts · neither trained on Game maker playdough nibble Big AI model the same questions PASS? faster+ at leastas good The proving test playdough sends the same questions to nibble and to a big model. nibble passes if it is faster and picks at least as well. Game maker playdough nibble Big AI model same questions PASS? faster + at least as good
Fig. 2. The first demo. Every pick is timed.
05 · Who it's for

Who it's for

Yes For you if

  • Your agents make the same kinds of choice again and again.
  • You pay a big model to answer multiple-choice questions.
  • You want those choices to run on your own hardware.

No Not for you if

  • You need writing, chatting or summarising. nibble only picks from the options you give it.
  • Your questions are one-offs and a few seconds don't matter.
  • You need a finished tool today.
06 · How it fits

How it fits

nibble is the choosing step in a set of five tools David Hatley is building for making things with AI agents, starting with games.

Mise en Placedoes the prep for any project (research, assets, references, context), so you or your agents start on the work.
bonfireruns AI agents until the work passes your checks, even through a crash.
playdoughturns one sentence into a small game you can play from a link.
napkin (planned)
is a design canvas where you and an AI agent edit the same components.
07 · FAQ

FAQ

Why not just ask GPT or Claude?

You can, and that's the bar. nibble earns its place only if it is faster and at least as good on prompts it never trained on. That test is the first demo. It isn't an LLM router either: a router picks which model answers, and nibble is the model that answers.

What is Kev?

Kev is Jared Palmer's open family of small decision models (Apache-2.0), an open alternative to the closed Jev. nibble installs Kev as a dependency, credits it, and adds game questions, the big-model test and the trainer.

How fast is it?

We haven't measured nibble yet. The only published figure is Kev's, measured by Kev's team. Kev's README reports Kev-4B answering six questions about a short text in 18.1 ms of model time on an H100, a data-centre GPU. nibble targets a 12 GB home graphics card or a CPU, which will be slower. Its first demo times every pick.

Can I trust the "how sure" number?

It is calibrated, so a pick marked 80% sure should be right about 8 times in 10. It can drift on a new task, so nibble re-checks it on answers kept aside for testing.

Is it only for games?

For now. playdough needs games first, so games are all we're testing.

Can I try it today?

No. Nothing has shipped. Read the plan, and watch it get built.