agentleFS
melissa-pereira-deelmelissa-pereira-deel/tiny-model-lab/.claude/skills/tiny-model-triage/SKILL.md

tiny-model-triage

Skill0 starsChanged 56 days agoMITView on GitHub

Decide whether a narrow task should be a tiny embedded model, classical ML, a deterministic rule, or a large-model API call. Use this whenever anyone proposes putting a model inside a product, mentions on-device or in-browser inference, says "we could train a small model for this", or asks whether an AI feature is worth building — even if they have already decided and just want help executing.

The file

We split this skill into the rules it gives an agent

3 rules. About 0 are common, 3 are distinctive, and 0 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈3

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 0

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Start at the bottom and stop at the first rung that works.

    line 12
  2. Distinctive:

    If none of the four apply, use an API.

    line 32
  3. Distinctive:

    Every experiment names a specific existing thing it must beat.

    line 36

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Tiny-model triage

Most tasks that sound like tiny models are not. Work the ladder from the bottom.

The ladder

Start at the bottom and stop at the first rung that works.

RungUse whenCost to try
Deterministic rule / parser / lookupOutput is fully specified by the input under known rulesminutes
Classical ML (logistic regression, trees)Tabular or hand-featurisable, closed output spaceminutes
Tiny neural model (KB–few MB)Needs learned representation, closed output, cheap labelsdays
Large model APIOpen-ended generation, world knowledge, reasoningminutes

The ladder is asymmetric on purpose: rungs 1, 2, and 4 cost minutes to try. Rung 3 costs days. Earn it.

The four legitimate reasons to go tiny

A tiny model has to win on a dimension the user can feel:

  1. Latency — the network round-trip alone breaks the interaction. An API call cannot land under 100ms; local inference can.
  2. Privacy — data cannot leave the device, so the feature is only shippable locally. Health, finance, journaling, unreleased work.
  3. Offline — it has to work on a plane, a subway, a construction site, rural Goiás.
  4. Cost at frequency — you want to run it continuously (every keystroke, every frame), which is only sane at ~zero marginal cost.

If none of the four apply, use an API. "More elegant" is not a reason.

The baseline requirement

Every experiment names a specific existing thing it must beat. Not "improve accuracy" — a named tool with a measured number. If you cannot name one, the task is not scoped. This is the single highest-value constraint in this repo.

Expected output

A filled harness/templates/experiment.yaml, or a written refusal naming which rung of the ladder to use instead. Never "it depends".

Worked contrast

Good candidate — highlighting code in a browser editor without shipping a grammar per language. Closed output space (token classes), free labels from an existing highlighter, wins on size and generality, needs to run per-keystroke. This is gpu-lexer, ~27.5KB.

Bad candidate — "summarise the user's document". Open-ended output, needs world knowledge, no closed label set, and a 2-second response is acceptable. Use an API, or a built-in browser model if privacy demands local.

Trap — "classify support tickets into 8 categories". Sounds like a tiny model. Is almost always logistic regression over embeddings, trains in seconds, ships as a matrix. Rung 2.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.