feat(claude): split output contract into terse mid-turn and self-contained wrap-up

[jinyeow]

Why?

The measurable gap. The concise output style specified the final message of a turn in exactly
one unspecified phrase — "tight bullets" — under a rule that already demanded "be extremely concise"
for every message. Nothing required the wrap-up to be understandable on its own. This ships four
named slots (What changed / What this means / Caveat / Verify) plus a mechanical trigger for
when they apply, so the final message of a multi-file or multi-stage turn can no longer be a bullet
list of identifiers borrowed from a subagent report ("B4", "Slice 2", "the MI-020 failure shape") —
which is what it had actually degenerated into.

Concrete maintenance consequence. The wrap-up is the only part of a turn that gets re-read later.
When it is unreadable without scrolling back through the tool calls, the turn has to be re-derived
from scratch — either by scrolling, or by asking for a re-explanation, which is the request that
prompted this change. A fixed header sequence also makes an omission visible: a missing named slot
reads as a gap, where a missing sentence does not.

Size. 3 files, +19/−8 (claude/AGENTS.md +12/−3, claude/output-styles/concise.md +13/−4,
claude/README.md +2/−1). No code, no tests — this is prompt text on both the advisory and the
enforced path.

Cost, honestly. The output contract is now roughly ten dense rules where it was three. Every rule
added raises the chance the whole block is partially dropped under heavy context load, which is the
failure mode this work keeps hitting. None of this text has been exercised on a real transcript yet —
it has been reviewed five times and used zero times. If wrap-ups read as mechanical after a week of
real use, the first thing to cut is the precedence rule.

Change

Split the contract by position in the turn rather than softening terseness globally:

• Mid-turn stays extremely concise — one signpost line per batch of related actions, no
  per-call narration.
• Final wrap-up is self-contained, in full sentences, written for a junior-to-mid engineer
  who saw none of the tool calls.

Adds the fixed header sequence, the trigger (two or more files changed, or two or more named stages
each producing a deliverable), a shorthand-expansion rule (plain description first, label in brackets
after), and an explicit "no length cap, but every line earns its place" rule.

Both the advisory copy (claude/AGENTS.md, also installed as ~/.codex/AGENTS.md, so Codex CLI
reads it and never sees the output style) and the enforced copy (claude/output-styles/concise.md,
injected at system-prompt level) are updated in sync, per the sync note in the latter's frontmatter.

Review

Five passes: Codex MCP (three rounds), an Opus red-team subagent, a Fable subagent, and a cold-read
pass with no history to counter anchoring. Passes 4–5 were mostly correcting wording introduced in
passes 1–3, which is where returns flattened. The final Codex pass returned zero findings.

Notable rejections of my own drafting: an anti-padding brake ("if the wrap-up is longer than the diff,
it is padding" — meaningless on read-only turns, and wrong in the routine small-diff-after-long-
investigation case) and a precedence rule that wrapped skill output shapes rather than deferring to
them (a findings-only review contract stops being findings-only once wrapped).

The sharpest catch came from the red-teamer and had survived two prior passes: the below-trigger rule
read absolutely ("a direct 1–2 line answer, no headers"), which forbade the Verify command on
single-file changes — the most common turn shape in this repo.

Deferred

A blocked turn — several things asked, most done, one stuck — has no good slot for the blocker; the
current rule puts it in the last bullet under What changed. Two fixes were considered (an optional
fifth Blocked on header, or a rule that the blocker leads the wrap-up) and both were deliberately
declined for now. Reasoning is recorded in the project brain. Revisit with real examples.

Not live on merge

The live config is the C:\Users\jin-y\dotfiles clone, so this takes effect only after that clone is
pulled — and output styles load at session start, so it needs a fresh session after that.