Compress

Two modes, best used together: write compressed, then re-pass the finished section. On-write gets the shape right — a fat draft trimmed later keeps its seams. After-write catches what only becomes visible once the whole thing exists: duplication across sections, a rule that outgrew its host, scaffolding nobody reads.

The five tests

Both modes run these. Every line here is read by a model mid-task.

• Behaviour — what does the model do differently because of this line? Nothing → it doesn't exist.
• Host — does a list, table row or sentence already cover the topic? Then the rule belongs inside it; a fresh section is the last resort.
• Home — is the fact already stated elsewhere in the repo? Reference it, never restate. Grep before assuming it isn't.
• Seam — can a reader tell which lines are new or touched? Visible seams mean an explaining voice or a block where an integration belonged.
• Deletion — remove the clause; does any agent behave differently? No → it stays removed.

Compress on write

1. Answer behaviour, host and home before typing — the host decides where the edit goes, so find it first.
2. Rules, not reasons. State the imperative; add the why only where an agent would otherwise break the rule, as one clause in the same sentence.
3. Clause > sentence > paragraph. Merge related rules with · or — rather than listing them.
4. Match the surrounding density — the voice of the lines above and below, not your natural one.
5. Cut on sight: preamble, "note that", restated rationale, examples where the rule is already unambiguous, hedges, any recap of what the text just said.
6. Seam and deletion test on the new lines, then save.

Compress after write

On a finished file or section, never mid-draft.

1. Read it whole first. Local edits without the full picture create the duplication you came to remove.
2. Map what it asserts — one line per rule. Two lines saying the same thing → merge into the stronger host. A rule with a home elsewhere → replace with a reference.
3. Deletion test every clause, top to bottom. Delete first, restore only what fails the test.
4. Rewrite the survivors into the fewest sentences that keep every rule. Restructure the host — shaving words while the structure stands is what leaves seams, and buys little.
5. Verify against the map: every rule still present, and report the size delta.