Agent Skills · Field Manual · No. 2 Try it Install GitHub

Propose. Cut. Guess.

For the want you would recognise instantly
and cannot describe at all.

Nobody can answer “what should we build?” Everybody can answer “not that, and here’s why.” So the agent stops asking the first question and only ever asks the second.

Plate IThe instrument

Cut a board yourself.

Nine hypotheses, numbered H1 to H9: candidate answers concrete enough to be wrong. Hover an answer to see what it would kill before you spend it. Then commit, and watch the board refill.

seed  ·  “we need better onboarding” 9 standing · 0 cut
Round 1
Every option below kills something. A question where they all killed nothing would be a dead question, and never gets asked.

A real session runs five or six rounds, not three. Notice what happened after each cut: the board was refilled. Survivors breed finer versions of themselves, so a first-round hypothesis names a whole approach and a fifth-round one differs from its neighbour by one decision.

Plate IIWhy asking fails

“Could you be more specific?”

01 · The request “we need better onboarding”

True, urgent, and shapeless. There is a real answer behind it and no sentence in anyone’s head that holds it yet.

02 · The usual reply “what did you have in mind?”

This asks for the expensive thing. Precision, produced alone, from nothing. The answer comes back vague, and the vagueness gets blamed on the person rather than the question.

03 · What works instead “not the video. we don’t have the footage.”

Instant, specific, and free. A wrong proposal costs nothing to reject and carries a reason with it, and the reason is the part worth having.

So the skill forbids the question outright. The agent is never allowed to ask you to be more specific, because that is the job it was called for.

Plate IIIThe one rule

A question that changes nothing is never asked.

Before asking, the agent works out what each surviving hypothesis would answer. If they all answer alike, the question carries no information, and asking it costs you real thought for nothing. Pick one and watch the column.

— —
—

Dead questions are the ones that feel most productive to ask. They sound thorough, they take real effort to answer, and the board is exactly where it was.

Plate IVThe round

Four beats, repeated until it lands.

The narrowing round: propose, cut, respawn, guess Twelve hypotheses are proposed. A question strikes several out. New, finer ones drop into the gaps. A lens closes on one, which is stated aloud as a guess; a refusal loops back and recasts the whole board. 1 — PROPOSE eight to twelve hypotheses, concrete enough to be wrong 2 — CUT only questions that split, and the reason each one died 3 — RESPAWN the gaps refill with finer ones, bred beside survivors 4 — GUESS a whole answer, stated aloud, every round — sure or not “NOT QUITE” — AND THE BOARD IS RECAST
A refusal reshapes every row, because the reason behind it applies to all of them.
Plate VThe output

One file, and half of it is the graveyard.

The session ends when you confirm a guess, or when the survivors differ only in ways you say don’t matter. What gets written down is the winner, the dimensions you called free, and every hypothesis that died with the reason it died.

The graveyard is what recognises a rejected direction when it comes back months later wearing a different name.

narrowing-onboarding.md
WONsample data in every new account, seeded from their first import
R1H2 video — no footage, and no one to make it
R1H7 human call — cannot cost human minutes per account
R2H3 interactive tour — nothing to tour in an empty account
R2H9 setup wizard — they don’t yet know what to configure
FREEwhere it lives in the nav · whether it can be skipped
Plate VIWhere it sits

It runs before grilling.

First — narrowing

Searches a space

Nothing exists yet. Its unit is a hypothesis, and its failure is asking a question that changes nothing.

Then — grilling

Walks a tree

The thing exists and needs sharpening. Its unit is a decision, reached once its prerequisites are settled.

Grilling is Matt Pocock’s skill, installed separately; narrow-then-grill runs the two in order, handing the winner and the graveyard straight across.

Try it

One command, then one word.

Nothing to clone and nothing to build. These are markdown files, not a program, and there is no setup step, because narrowing writes into your project only when a session ends.

Best pointed at something genuinely unshaped. On a decision you have already made, it has nothing to cut and will say so.

  1. 1

    Fetch the skills

    npx skills@latest add franciscohermida/skills

    Asks which skills you want and which agents to put them on. Claude Code, Codex, Cursor. Add -g for every project on the machine.

  2. 2

    Point it at something shapeless

    /narrow-me we need better onboarding

    Or /narrow-then-grill to keep going once it lands and stress-test what survived.