Who Edits an AI Writing Staff? 191 Articles, 15 Iron Laws, One Stamp
Key Takeaways — Executive & AI Summary
- A one-person Chinese publishing operation archived 191 finished HTML articles — 72 in articles/, 105 in daily/, 14 in archive/ — between June 15 and Aug 27, 2026, produced by a staff of 17 agent persona files and signed by four emotion-typed pen names, whose bylines appear in 30 archived files (archive recounted 2026-08-27).
- Quality is codified as 15 iron laws enforced as 7 machine hard scans plus 4 AI judgments; a passing gate writes a .gate.json stamp and the push script refuses to publish without it, delivering to the platform's official draft-box API (workflow standard + gate stamps, 2026-08).
- Every caught flaw lands in quality-lessons.jsonl — 102 records since July 18 (88 first-round, 11 second-round, 3 third-round failures), the newest written Aug 27 17:46:40 — and any error repeated three times is promoted into a permanent hard-scan rule; today a missing-counter-view failure was cured and re-stamped 21 seconds later.
Episode 9 of One Man One Legion — the second stop in the workshops arc, after the image workshop. Episode 19 told how the auto-publishing cron jobs — six of them WeChat-article slots on the old schedule — went dark on July 23. This is what the factory looks like without a clock.
A daily publishing operation run by AI forces one question on you: who reads the drafts before real readers do? One person cannot read everything a legion writes and still have a day. This workshop’s answer is the most disciplined thing on the fleet: nobody reads every word, and nothing ships unchecked anyway. Trust wasn’t asked of the models. It was replaced by machinery.
A staff meeting of seventeen files
The editorial staff lives in one directory — 17 persona files. The editor-in-chief is Chen Rui; his file gives him a resume (fifteen years at a major Chinese investigative weekly, now founder of a studio) and a standing demand for every pitch: where is the hook? Around him: a writer, a researcher, a trend analyst, a hook creator, a compliance reviewer, a designer, a formatter, five video writers. The most telling file is the writer’s “de-AI” variant, whose rules forbid the machine’s virtues — no perfect first-second-last structure, no evenly balanced both-sides, no tidy uniform lists. That persona’s entire job is to write like a person whose mind occasionally wanders.
Emotions get pen names
Articles don’t ship under one byline. An emotion-typing scheme maps each piece to a persona: type A signs as Lu Shi, B as Shen Jianwei, C as Lin Shu, D as Qin Yin. Their bylines — Shen Jianwei on industry teardowns, Lu Shi on surviving the AI era — appear in 30 archived files. A fifth persona, Zhixing Xiaoya, exists only as configuration so far: her own app config for US-stock analysis, but zero signatures in the archive. Configured is not shipped, a distinction this series keeps relearning.
What the floor produced
The archive holds 191 finished HTML articles — 72 in articles/, 105 in daily/, 14 in archive/ — from the first daily piece on June 15 (the topics pool opened June 12) through three articles dated today, Aug 27, one of them on the Gates AI warning. The cron contraction left no gap on this shelf. The factory just stopped caring what time it was.
Fifteen iron laws, eleven enforced gates
Quality is codified as 15 iron laws in the workflow standard, enforced in two layers. Seven are machine hard scans, run by a gate script before anything moves:
| # | Machine check |
|---|---|
| M1 | at least 3 source domains, deduped |
| M2 | banned and high-risk words |
| M3 | quotable-line density |
| M4 | pain word in the first 15 title characters |
| M5 | at least 3 images on the platform CDN |
| M6 | legal safety: disclaimer plus counter-view |
| M7 | stripped length over 500 characters |
Four more are AI judgments the writer must answer before push: an information gap in the first 100 characters, argument-ammo matching the topic cluster, conversational temperature, an unresolved question at the end. And one law outsources truth itself: high-risk claims — scores, names, specific numbers — must be confirmed by an AI judge, a local model cross-checking against web search. A failed fact-check doesn’t warn. It forbids the push.
The stamp that unlocks the push
A passing gate writes a .gate.json stamp — passed, round, 7 machine checks, 7 passed — and the push script refuses to run without that stamp. The gate is not advice; it is the physical key. The final leg is the platform’s official draft-box API: the machine delivers into a human’s draft box and stops there. After the push it verifies delivered length and the returned media ID before logging the topic as done.
A memory that writes back
Every caught flaw lands in a lessons file — 102 records since July 18, the newest written today at 17:46:40. The rounds read 88 first-round failures, 11 second-round, 3 third-round. And the loop has a ratchet: any error that repeats three times gets promoted into a permanent hard-scan rule. The factory’s constitution grows out of its own scars.
Today’s catch, live from the log: a story about champions paying their own way was flagged for missing a counter-view — a legal-safety failure, iron law 14 — at 17:46:40. Its passing stamp is timestamped 17:47:01. Twenty-one seconds from caught to cured — a turnaround only machine speed explains: the draft was patched and the gate re-ran, with no human in that loop.
That speed is the tell. Episode 19 took the clock away because publishing on a schedule, unattended, produced output nobody asked for. What survived is the mirror image: machines that check relentlessly, one human alone holding the trigger, and a JSONL memory that got a new line while this episode was being written.
The QC gates get their own episode next in this arc. Back to the fleet audit.
FAQ — Direct Answers
- Does a human read these articles before they reach readers?
- Not every word — and that is the design. Machines run the checks that can be mechanized (sources, banned words, quotable-line density, title pain words, image counts, legal safety, length), AI judgments cover the rest, and the push script physically cannot run without a passing gate stamp. What the human holds is the only decision machines don't get: the draft box and the trigger.
- Who exactly are the writers?
- Seventeen persona files: an editor-in-chief with a fictional fifteen-year investigative-press resume, a researcher, a trend analyst, a hook creator, a compliance reviewer, a designer, a formatter, five video writers — and a 'de-AI' writer variant whose rules forbid perfect structure, balanced both-sides arguments, and tidy uniform lists. Four pen names sign by emotion type, verified in 30 archived files; a fifth exists only as config with zero shipped signatures.
- What happens when an article fails a gate?
- The failure is logged to quality-lessons.jsonl with its rule and detail, the draft is rewritten, and the gate re-runs — today one counter-view failure went from caught to re-stamped in 21 seconds. The ratchet: any error that shows up three times gets promoted into the machine hard-scan script, so the same class of mistake becomes structurally impossible.