Write your first rule

This is the whole loop in five minutes: edit a source file, recompile, confirm the rule reached the artifact an AI reads.

This is the shape knob — changing who the version of you is. The simple move is one line in a file; the steps below (recompile, confirm it landed) are the deeper knob, for when you want proof it stuck rather than hope.

1. Open the behavioral file

context/who.md holds behavioral rules — how the mind responds. Open it in your fork.

2. Add a rule as trigger → action

Write rules as a detectable moment plus what to do, never an aspiration. “Be concise” has no activation point; a trigger does.

## Output Behavior

<!-- constraint: soft -->
- When a reply is drifting past three sentences for a yes/no question — stop and
  lead with the answer.

Why this shape works: the model evaluates the when at generation time and steers. Aspirations (“be helpful”) compete with task pressure and lose. Background: Soft / routing / verifiable.

The <!-- constraint: soft --> comment tells the compiler this is soft guidance, so it routes into the injected CONTEXT.md. Routing and checkable rules get the routing / verifiable annotations and go to ENFORCEMENT.md instead.

3. Recompile

./scripts/compile-context.sh

The build prints a word count and a per-section position report — so you can see whether your new rule pushed others deeper into the weak-attention middle. If the total is over budget, remove before adding: Why context rots.

4. Confirm it landed

Open the generated CONTEXT.md and find your line. This is the artifact an AI actually receives — if your rule is there, it ships next session.

grep -n "lead with the answer" CONTEXT.md

That grep is the point, not a formality: the change isn’t done until the receiver — the compiled artifact — shows it. (Know it’s working takes this one step further: proving the rule changed the AI’s behavior, not just that the text exists.)

What you just did

You wrote software in a sentence, compiled it for how a model reads, and verified it reached the model. Everything else in MemMini is that loop at larger scale.

Next

• Write rules that bind — make rules that survive across model families
• What is MemMini — the layers and the capture loop