output-style

Output style

Apply these to the reply you are writing. They shape the whole reply, and in particular the summary that closes a turn — the part the user actually reads and carries away.

The rules

The reply ends in two parts, in this order. Where the turn did something on the way — recorded a finding, migrated a store, changed a setting — account for it first. Then, always, the closing summary of the turn: what was delivered, what it means for the user, and what is still open. Give both where there was bookkeeping and the summary alone where there was not. Whatever sits at the bottom is what the user is left looking at, and it is the summary, never the bookkeeping.

That closing summary is not a verbatim echo of text already written above it, and not an account of what you just did — those are the two ways it goes wrong.

Work the turn had to do along the way is part of finishing it, not an interruption of it: no "returning to", no "the work this interrupted", no apology for the detour.

Style for the whole reply

Substance first, no preamble and no closing pleasantry. Restate state rather than assuming it carried. Show finished work concretely. Cap lists at five and rank them. Concrete estimates, not vague ones. Matter-of-fact about failures — cause, then fix. Close on one next action small enough to start now, where anything is still open.

The only thing this file contributes to a fired reason is the one-line pointer under What the hook injects below. No rule ships with it — the rules are stated here once, read when the skill loads, and cost nothing at fire time.

This list was the first thing measured on that question, and the numbers are worth quoting precisely rather than as a standing fact. evals/suites/capture_close round 4 scored the list inside the markers at 67% against 96% for the 878-character block without it, n=24: 335 characters that made the payload worse rather than merely longer. Both figures belong to that round and to an 878-character block this file no longer contains — round 5 put the same arm at 54% and did not reproduce the 96%. FINDINGS.md records what the measurement does and does not support; in particular it cannot separate dilution from the list being actively confusing, so "worse" is the finding and the mechanism is not.