Newsroom review
One draft, five editors who cannot see each other, one arbiter. The editors disagree where the draft is ambiguous; the arbiter turns that into one list a writer can fix in one pass.
Do not review the draft yourself first. Your read anchors everything after it. Build the packet, spawn the editors, and read only what they return.
Inputs
- The draft as a markdown file. Front matter
title:anddescription:if the channel needs them. brand-profile.md(template:templates/brand-profile.md; fill it withnewsroom-setup). Without one, run with the defaults and say in the review that the brand rules were not checked.- Optional: the format key (
guide,comparison,case,x-articleor one of yours), the author name, andsources.md(template:templates/sources.md).
Procedure
1. Mechanical checks (no judgement yet)
S=skills/newsroom-review/scripts
python3 $S/structure_lint.py draft.md --profile brand-profile.md --format guide > run-1/structure-lint.txt
python3 $S/claims_scan.py draft.md --sources sources.md --profile brand-profile.md > run-1/claims-scan.txt
Both exit 1 when they find something. That is expected; they feed the editors, they do not gate.
2. Build the blind briefs
python3 $S/review_run.py packet draft.md --profile brand-profile.md --format guide \
--author "Author Name" --out run-1 \
--attach structure=run-1/structure-lint.txt --attach claims=run-1/claims-scan.txt
This writes run-1/brief-<editor>.md for each of the five editors. Each brief has the editor's role
from references/editors.md, only the profile sections that lane needs, the draft with line
numbers, and the output contract (human verdict plus one json block).
3. Spawn five editors in ONE message
Send five subagent calls in a single message so they run in parallel and none sees another's output. Each gets the same instruction:
Read exactly one file:
run-1/brief-<editor>.md. Do not read any other file. Follow it and write your verdict torun-1/verdict-<editor>.md. Reply with one line.
Rules:
- Fresh subagents only. Never the agent that wrote the draft, never one that has seen another verdict.
- Read-only on the draft. Editors name locations and changes; they do not edit.
- If the host cannot spawn subagents, run the five briefs in five separate turns, clear your notes between them, and mark the final review single-context, degraded. The panel is weaker this way because later editors remember earlier ones.
4. Merge
python3 $S/review_run.py merge run-1
The script reads the five json blocks and:
- checks each quoted phrase against the draft; a quote that is not there is demoted from blocking;
- groups findings that point at the same words or the same missing thing;
- lists conflicts, where two editors want different fixes for one item;
- proposes a verdict (exit 0 SHIP, 1 FIX or REWRITE):
| Verdict | Rule |
|---|---|
| REWRITE | format editor reports a mismatch, or 3+ blocking items from structure and claims |
| FIX | any blocking item, any score under 6, or 5+ recommended items |
| SHIP | none of the above |
Output: run-1/merged.md (for the chief editor) and run-1/merged.json (for the next round).
5. Chief editor
Spawn one more fresh subagent with references/chief-editor.md, run-1/merged.md, the five
verdict files and the draft. It confirms each blocking item against the draft, resolves conflicts
in the written order of precedence, may move the verdict one step with a reason, and writes
run-1/review.md in the format of references/verdict-format.md.
6. Act on the verdict
- SHIP: hand off to
article-producerstep 6 (publish package) or to the user. - FIX: apply the blocking items, then the cheap recommended ones. Save as
draft-v2.md, rerun steps 1 to 5 intorun-2, and merge with--prev run-1/merged.json. The merge then reports which items were resolved and which are still open. - REWRITE: redraft in the right format first (the format editor's moves). Start a new run;
do not pass
--prev, because line numbers no longer match.
Stop at round 5 (--max-rounds). The merge prints a stop notice; hand the draft and the last
review to a human. Five rounds without SHIP usually means the brief, not the draft, is wrong.
What the editors check
| Editor | Lane | Blocks on |
|---|---|---|
| Format | right format for the topic and the reader's intent | fit under 7 |
| Structure | 10-item checklist: title, description, summary, headings, template, hook, code and media, paragraphs, close, length | a rule written in the profile |
| Voice | the named author's profile, line by line | a Banned-list hit |
| ICP | share test, the reader's words, segment clarity, admits limits | a fixable line that fails the share test |
| Claims | every number, superlative and "studies show", classed SOURCED / NO-SOURCE / VAGUE / WRONG | a WRONG claim, an unsourced claim that carries the argument, a brand superlative without method |
Full role texts: references/editors.md. To add a sixth editor (legal, accessibility, SEO), add an
H2 to that file whose first word is the new key, add the key to EDITORS and SECTIONS in
scripts/review_run.py, and pass --editors.
Rules for you, the orchestrator
- Do not soften a verdict. If the merge says REWRITE, the user hears REWRITE first.
- Do not add findings of your own to the review. If you notice something, put it under "Noticed, not reviewed".
- Never invent a source to turn NO-SOURCE into SOURCED. Either the writer adds a real one, or the claim changes or goes.
- Report the run honestly: how many editors returned, whether the panel was degraded, the round number.
Files
scripts/review_run.py:packetandmerge.scripts/structure_lint.py: the countable half of the structure checklist.scripts/claims_scan.py: the claims table, with sources-file and conflicting-figure checks.scripts/newsroom_profile.py: reads limits from the profile (shared by the scripts).references/editors.md,references/chief-editor.md,references/verdict-format.md.- Worked example:
examples/northwind-ledger/(two rounds, real outputs).