Agent Skills · A working notebook Index The pattern GitHub

Skills are an experiment. These are the notes.

Not a prompt library. Each of these is an attempt at a different shape a skill can have: a working loop the agent runs with you. Four so far, and each one turns on the middle beat.

1The agent acts. The part that already works.
2Something gets written down. The beat that always goes missing.
3You decide what it meant. Never the agent.
What is happening to you?
No. 1In use
YOU SPOKE THREE LINES · NO RULE LATER — READ TOGETHER, DECIDE WHERE EACH RULE LIVES

Divergences

Diverge · Record · Align

The agent writes a three-line note whenever the work changes direction. No announcement, no pause. Later you read the pile together and decide what becomes a rule, and where that rule should live.

For when the same correction keeps happening and nothing remembers it.
  • setup-divergences
  • write-divergence
  • align
Read the loop →
No. 2In use
PROPOSE · CUT · REFILL · “IS THIS IT?”

Narrowing

Propose · Cut · Guess

You can’t say what you want, so the agent puts a dozen concrete hypotheses in front of you and asks only the questions that would kill some of them. Every round it commits to a whole answer, is this it?, because your refusal is the fastest information in the session.

For a want you would recognise on sight but cannot describe.
  • narrow-me
  • narrow-then-grill
  • narrowing
Cut a board →
No. 3New
ATTENTION THE WALL · FILTERED TO WHAT NEEDS YOU

Attention

Filtered from the wall

An agent writes long because meaning takes a map of words to carry. What that wall cannot do is keep its parts apart, so every reply ends with one section collecting whatever needs you, and reading the rest becomes optional. One command turns it on where you are standing, another keeps it for the whole project.

For when the part you have to act on is buried in the middle.
  • attention
  • setup-attention
Find the four asks →
No. 4New
REQUIRE · WRITE · ONE LINE PER WORD

Prompting

Require · Write · Justify

A generation model takes every word literally, so the answer arrives as four sections: what the picture must achieve, the prompt itself, and a line for every single word in it. Paying a line per word makes padding cost something, and the reasoning finally has somewhere to sit other than the prompt.

For when a careful prompt keeps returning the wrong picture.
  • write-generation-prompt
Pay for a word →
The patternWhat is being tested

What each one is shaped like.

The interesting question is not what any one of these does. It is whether skills have shapes at all, meaning repeatable structures you could reach for again. Four attempts, built as unlike each other as they would go.

DivergencesNarrowingAttentionPrompting
RunsContinuously, in the backgroundIn one deliberate sittingAt the end of every replyEvery time a prompt is asked for
Triggered byThe agent, silentlyYou, by nameNothing — it is always onYour request, with no command in it
Its unitA record of a momentA hypothesis that can be wrongOne thing that needs youA single word
DirectionBackward, over what already happenedForward, over what has not happened yetOut of the wall, down to the end of itOutward, at a reader who takes it all literally
You supplyA correction you were making anywayA refusal, and the reason for itOne decision, in one lineThe picture that came back
Fails whenIt records everything, so nothing stands outIt asks a question that changes nothingEverything lands in itIt justifies phrases instead of words
Try it

One command.

Markdown files, not a program. It asks which skills you want and which agents to put them on. Claude Code, Codex, Cursor.

Run a set for a week. If it has not told you something you did not already know, delete it — the footprint is small enough that it leaves nothing behind.

  1. 1

    Fetch the skills

    npx skills@latest add franciscohermida/skills

    Add -g to install them for every project on the machine.

  2. 2

    Then, in a project

    /setup-divergences         watch this repo for turns
/narrow-me                 find a thing you can’t name
/attention                 collect what needs you at the end
“write me a prompt for…”   prompting has no command

    /attention lasts the session and /setup-attention writes it into AGENTS.md, as /setup-divergences does. Narrowing writes nothing until a session ends, and prompting starts on the request itself.