I built a brief gate checklist so expensive models stop “fixing prose” while the same mistakes keep coming back. I wanted a reusable list after watching the same classes of miss return: template leftovers, hedged facts, and a Rank Math focus keyword left unset. This post is the two lists I actually run — before a writer starts, and before a WordPress draft ships.
It sits in Layer 4 of my Cursor workflow. The editor/writer split is already covered in AI writing pipeline; here I only deepen the gate that makes that split stick. The checklist encodes the loop already described in AGENTS.md and the style guide.
Why fixing prose teaches the wrong lesson
Rewriting a bad paragraph in the same expensive chat feels productive. The next draft often makes the same class of mistake.
The harness only improves if the failure becomes a brief sentence or a style-guide line. On fail I change the brief or config/style-guide.md, then re-delegate. I do not hand-polish the article body in the editor model.
Two checklists work as a pair: what the brief must contain before the writer runs, and what the editor checks before a WordPress draft. That contract is the point of the sections below. The brief is the product; the draft is a derivative.
Brief gate checklist (before the writer runs)
Nothing starts until the brief is complete enough that a cheaper writer model can draft without inventing. For manual and methodology posts on this blog, the editor writes or sharpens drafts/manual/<slug>.brief.md first. The writer then drafts Markdown from that brief only.
Checklist A — the brief must include:
- Audience + intent (who searches, why)
- Exact title + focus keyword + planned slug
- Structure with required beats (H2 map)
- Fact sheet / verified claims (numbers need source + as-of when relevant)
- Hard constraints (what the writer must not invent)
- Pointer to
config/style-guide.md - For methodology track: artifact excerpts + a human-supplied failure case
The methodology extras matter. Artifact claims about rules, skills, or configs must come from real files the brief already cites. The human material contract means writer delegation waits until trigger, failure texture, and refuse-to-claim answers are pasted into the brief.
I refuse to invent experience details the brief does not supply. No time-saved percentages. No claim that any gate catches everything. No invented Rank Math scores for a specific post. No team-process claims that are not on the sheet.
I treat incompleteness here as the real bug. Gate fails that look like “bad prose” are often brief incompleteness. When Checklist A is thin, the writer fills gaps with inventiveness — and the expensive model ends up cleaning that mess later.
Gate checklist (before WordPress draft)
After the writer drafts Markdown from the brief only, I gate against the brief and config/style-guide.md before anything goes to WordPress. This half of the brief gate checklist is editorial judgment first, scripts second.
Checklist B — editorial (all tracks):
- Facts match the brief (no invented paths, features, or metrics)
- Style-guide readability (short paragraphs, lists for 3+ items, bold budget)
- Focus keyword in title, slug plan, first paragraph, and sparse exact uses
- At least one internal link (from post #2 onward) and at least one outbound official or primary source
- Image placeholders present if the brief required them
Automated helpers vs methodology
This repo has automated helpers. make gate SLUG=... runs seo_lint, a non-experience critic, and dedup. That path is tuned for evergreen synth.
Methodology and experience posts do not need the non-experience critic or the synth table/FAQ shape. For those I run dedup plus editorial checklist B. Dedup checks topic collision against the drafts and published registry.
Forcing the evergreen critic onto a first-person methodology post produces false fails. Automated gate scripts are track-specific. Use the helper that matches the track, then finish with human eyes on Checklist B.
Publisher checks after the WP draft
The publisher step is editor work, not writer work. After make wp-publish, WordPress status stays draft until a human Publishes.
Publisher checklist:
- Rank Math focus keyword + meta description set
- Featured or hero alt includes the focus keyword when possible
- Expected Rank Math band about 75–85 for evergreen and methodology; do not chase 100 via stuffing
Body can be keyword-correct while Rank Math looks broken if the focus keyword was never set on the WP post. A collapse to about 20 usually means that. Set the focus keyword via Rank Math, or the publish script’s updateMeta path, after draft upload.
That is Failure 3 in short form: SEO metadata is a publisher item, not a prose rewrite. Evergreen synth gets one short beat only: collect → facts → brief → writer → make gate → WP draft. The full Makefile tutorial is out of scope here.
When the gate fails: fix the brief, re-delegate
When the gate fails, I do not rewrite the body in the editor model. I fix the brief or the style guide, then send the writer again. That is the AGENTS.md contract for this blog: the editor stays on brief, gate, and publish steps; the writer stays on prose from the sheet.
Failure 1 — template leftover
An early evergreen path once produced a scaffold draft with leftover wrong-persona and FAQ junk from a template. The fix was not polishing paragraphs in the expensive model.
The brief was sharpened and the writer re-ran until the draft matched the assignment. The miss looked like bad prose. The real gap was an incomplete brief. Style-guide voice for synth drafts says the same thing: leftover topic-agnostic junk means improve the brief and re-delegate, not hand-edit template prose.
Failure 2 — hedged official facts
Before harness v5 (2026-07-20), comparison drafts hedged official prices with lines like “commonly reported around $X”. After the style-guide rule landed, briefs must supply verified numbers with as-of dates, and writers state them plainly.
Four posts were re-briefed and rewritten once that rule existed. A repeated gate miss became one enforceable style-guide line the next brief must satisfy. The editor contract is clear: if the fact sheet lacks a number the article needs, fix the brief, not the prose.
Those scars often freeze into sticky agent rules. I keep that encoding story in Cursor agent rules that stick. Official Cursor docs on rules are the product reference when I wire the same idea into the editor.
A short worked loop (this blog)
Here is the qualitative loop I run for a methodology post like this one.
- Editor writes or sharpens
drafts/manual/<slug>.brief.md - Writer subagent drafts Markdown from the brief only
- Editor gates against the brief and
config/style-guide.md - On miss → add a brief or style-guide line → re-delegate the writer
- Images sidecar +
make wp-publish→ human Publish in WP Admin
The same shape shows up when a miss is a style-guide gap rather than a missing fact. I add one enforceable line to config/style-guide.md, point the next brief at it, and let the writer produce a clean draft instead of patching the old one by hand.
I do not invent timestamps, word counts, or Rank Math scores for a specific post. The intentional target band is about 75–85. A collapse to about 20 still means the focus keyword was never set on the post.
A checklist does not replace taste. It catches repeatable misses. Overlong briefs become unread walls — the same failure mode as overlong rules. Keep the brief short enough to read; put the force into hard constraints and the fact sheet.
Wrap-up
A brief gate checklist is how Layer 4 stays monotonic: each miss makes the next assignment stricter. The expensive model stays on brief and gate work. The writer stays on prose from facts already on the sheet.
If you are mapping this back to the full harness, start from my Cursor workflow. Next methodology candidates on my list are skill authoring and style-guide evolution — no publish-date promises, just the next places the same loop will tighten.