Skip to content

docs(decisions): 0237 — agent brevity binds messages, not board artifacts (#4694) - #4703

Open
usirin wants to merge 1 commit into
mainfrom
umut/adr-0237-agent-output-scope
Open

docs(decisions): 0237 — agent brevity binds messages, not board artifacts (#4694)#4703
usirin wants to merge 1 commit into
mainfrom
umut/adr-0237-agent-output-scope

Conversation

@usirin

@usirin usirin commented Aug 1, 2026

Copy link
Copy Markdown
Member

Records the founder ruling on agent output style, closing the three questions #4694 was typed type:decision to settle.

Adds one file: .decisions/0237-agent-brevity-binds-messages-not-artifacts.md. Purely additive — no index (there is none, ADR 0126/0129), no supersede, no amendment.

What the ADR rules

  1. "No AI jargon" means the AI-tell catalog, not coined labels generally. Scoped to claude-plugins/kampus-pipeline/skills/writing-clearly-and-concisely/references/ai-writing-tropes.md. House idioms stay exempt exactly as that skill already states — §CP, Fixes #N, fail-closed, the progress-comment headers, ADR frontmatter, .glossary/ defined terms. The literal reading would have contradicted a rule the review gates already enforce, and the pipeline runs on those labels.
  2. The brevity rule binds messages to humans only. Board artifacts — issue bodies, PR descriptions, review verdicts, epic ledgers, ADR and pattern bodies, commit messages — are carved out and stay governed by writing-clearly-and-concisely. Artifacts are pull, messages are push: a reader chooses to open an issue, a message arrives whether or not there is room for it. An over-broad reading would gut review verdicts and epic plans.
  3. CLAUDE.md states the message rule and points at the skill for the prose half. One source owns each half; no second copy of the prose rule.

Relationships stated in the ADR

Number claim

Took 0237, against a freshly fetched origin/main plus the in-flight reservation set:

  • Merged max on origin/main: 0236 (0236-eval-harness-gains-a-spawning-shell.md).
  • In-flight, from open ADR PRs via gh api REST: 0235, 0223 (exit 0 — the enumeration succeeded; a failed query would have been surfaced, not fallen back from).
  • max(union) + 1 = 0237. decisions-index validate passes locally.

Control-plane check

No control-plane surface is touched. The single file added is under .decisions/. Nothing under .claude/ or claude-plugins/** changes — the ADR only cites those paths. The downstream CLAUDE.md edit (also non-§CP) and the crew-def edit (which is §CP) are separate follow-ups.

Scope

Does not close #4694 or #4690, does not edit CLAUDE.md, does not touch the writing skill.

Vocabulary impact

Coins push surface / pull surface. Routed to .glossary/TERMS.md by report #4702 rather than added inline, so this PR stays purely additive.

Deviations

  • The vocabulary-impact term was routed by a filed report (glossary: add the push surface / pull surface pair coined by ADR 0237 #4702) instead of an inline .glossary/TERMS.md row. The adr skill permits either. Chosen so the PR stays a single added file, matching the skill's purely-additive shape.
  • The contradiction sweep's 8-entry shortlist was cleared by hand, with two citations added. sweep-shortlist.sh returned 8 adjacent live-accepted ADRs. Each was opened: 0099/0128 are glossary-maintenance mechanics, 0189 is crew-roster cardinality, 0229 is relay-vs-derive for scripts, 0225 rules what a verdict body carries for §CP classification (not how it reads), 0069/0126/0198 are unrelated domains. None re-decides a question this ADR settles, so no amendment was needed; 0053 and 0062 were cited in the body, and 0119 is cited as the tone-sibling.
  • No other deviation. The template, the required **What this decides:** line, title discipline, the bolded decision sentence, and off-disk-resolved cross-link slugs are all as the skill specifies.

Fixes nothing on its own — this is the recorded ruling for #4694; the implementation follow-ups remain open.

…acts (#4694)

Records the founder ruling on agent output style: the concise/bulleted/
jargon-free rule binds messages to a human; board artifacts stay complete
evidence under writing-clearly-and-concisely. "No AI jargon" is scoped to the
AI-tell catalog, so house idioms stay legal. CLAUDE.md states the message rule
and points at the skill for the artifact half.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
@github-actions

github-actions Bot commented Aug 1, 2026

Copy link
Copy Markdown

No preview deploy

  • No preview deploy for this PR — its diff touches no deploy-relevant path, so no preview stack was minted and e2e is not applicable. (737c785)

This was referenced Aug 9, 2026
usirin added a commit that referenced this pull request Aug 10, 2026
…me, never by backfill (#5041)

Repair round 1, addressing review-doc's FAIL on PR #5273.

F1 — the ADR recorded the opposite of what was ruled. The founder ruled directly
at 2026-08-10 ~07:43Z on #5041: LAZY cutover, no bulk backfill. That confirms the
founder-delegated ruling of 2026-08-10T00:15:28Z on the same issue, which the first
draft contradicted and cited nowhere. The decision is re-recorded as lazy: absence of
a `ready-for:` label stays an exclusion, the label is applied at triage time going
forward, and the already-triaged rows are not backfilled. The ruling history is now in
`## Context` as its own table, and backfill-first is written up as a rejected option
with the false-signal reason. Rewritten rather than amended: this is an unmerged draft,
and the landed record must read as one coherent decision.

F2 — renumbered 0253 -> 0254. PR #5271 opened 28 seconds earlier and owns 0253.
Re-enumerated at repair time: main carries 0250-0252, 0255, 0270 (no 0253/0254); open
PRs claim 0253 (#5271 and this one), 0256 (#5281), 0257 (#5282), 0258 (#5285), 0237
(#4703), 0235 (#4614). 0254 is free. The slug changed with the decision.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

CLAUDE.md: add an agent output-style convention — concise, bulleted, no AI jargon

1 participant