A recipe for one job

Think of it when you explain the same multi-step process to the agent for the third time.

In everyday life

It is like a recipe card in a restaurant kitchen. The title on the rack tells the cook which dish the card is for, and the cook takes the card down only when that dish is ordered.

One case

A food delivery app team wants every bug fixed the same way: first a test that fails on the bug, then the fix. Over two weeks the agent fixes 8 bugs. A developer types the steps for 5 of them and forgets for 3. On those 3 the agent guesses a fix with no test, and 2 of the 3 bugs come back within a week.

The steps were followed only when someone typed them

The case, cell by cell1 cell = 1 bug

Test first, as the team wants

yesyesyesyesyesyesyesyes

Test first, as the agent did

yesyesnoyesyesnonoyes

Bugs that came back

Fixes with no test

Bugs #3, #6 and #7 were the ones where nobody typed the steps.

Every case

A process you type from memory changes with your memory. A skill writes it down once, so the agent reads the same steps on every run. It costs little until it is needed: the agent keeps only its descriptionIn a skill, the one or two lines the agent can see before it loads the skill. The agent loads the full steps when the description matches the job. in view and reads the steps when the job matches. So the description decides when the skill loads on its own, and with a vague one the steps may never load.

The word

Engineers call this

skillA written set of steps for one kind of job, such as fixing a bug or opening a pull request. Usually the agent keeps only a short description in view and loads the full steps when that job comes up; some skills load only when you call them by name.

Turn any process you repeat into a skill. End each step with a clear sign that it is done, and write the description so it names the job that should trigger it.

Where this idea leads

is loaded only when needed, unlikeStanding orders
can make the agent start withOne red commandHabits, While you build
can end every job withProve it worksHabits, While you build
should be checked withA test for the agent itselfSteering, Check it

Skills that encode it

writing-for-agents Matt PocockA skill’s description is a context pointer: its wording, not its target, decides when the agent reaches the steps.
reflect pstackRoutes a session’s lessons into edits to existing skills, and tunes a description that did not trigger when it should have.
principle-build-the-lever pstackWhen helpers share a job, write the recipe and its verification contract as one skill they all read.