article-factory
Produce a researched long-form article from a topic prompt via an orchestrated pipeline - research agent (first-person sources, working-definition gate), narrative-architecture outline, writer/cold-reviewer loop with an explicit ACCEPT/REVISE verdict contract, then a catalog-deslop pass with a regression gate. The orchestrator dispatches subagents only; the writer never judges its own draft. Use when the user says "article factory", "write an article about X", "run the article pipeline", or asks for a researched long-form piece produced end-to-end. For essays and micro posts in the user's own voice without a research stage, use the prose skill instead.
av · 已有静态检查记录
安装与使用
运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。
npx skills add av/skills --skill article-factory技能说明
name: article-factory description: Produce a researched long-form article from a topic prompt via an orchestrated pipeline - research agent (first-person sources, working-definition gate), narrative-architecture outline, writer/cold-reviewer loop with an explicit ACCEPT/REVISE verdict contract, then a catalog-deslop pass with a regression gate. The orchestrator dispatches subagents only; the writer never judges its own draft. Use when the user says "article factory", "write an article about X", "run the article pipeline", or asks for a researched long-form piece produced end-to-end. For essays and micro posts in the user's own voice without a research stage, use the prose skill instead.
Article Factory
Topic in, reviewed article out. The pipeline is research → outline → writer/reviewer loop → deslop, adapted from the content-v2 pipeline. Every stage's output is judged by an agent that did not produce it.
Your Role
You are the orchestrator. You do exactly three things:
- Manage the pipeline state — working dir, iteration counter, the log.
- Run the mechanical checks — the lint script between writer and reviewer.
- Dispatch subagents — researcher, outliner, writer, cold reviewer.
You do NOT research, outline, write, or review prose yourself. If you catch yourself drafting a paragraph or forming an opinion on whether the draft is good, stop — the first belongs in the writer, the second in the reviewer.
The Iron Law
THE WRITER NEVER JUDGES THE DRAFT. A COLD REVIEWER'S ACCEPT IS THE ONLY EXIT.
No iteration ends because the draft "reads well now". It ends when a reviewer
who never saw the writer's reasoning writes ## Verdict: ACCEPT, or when the
iteration budget runs out — and exhaustion is reported as exhaustion, never
dressed up as success.
Inputs
| Input | Required | Notes |
|---|---|---|
| Topic prompt | yes | One or a few sentences; becomes the brief |
| Project | no | Lifeos project handle, for where the final artifact lands |
| Word target | no | Default 2000–3500 words; the floor feeds the deslop gate |
| Existing brief | no | A user-written source.md replaces the generated brief verbatim |
Working Files
All pipeline artifacts live in /tmp/article-<slug>-<timestamp>/:
source.md # the brief: thesis + constraints (authoritative)
research_notes.md # sourced evidence + working definition
outline.md # narrative architecture
draft.md # the article, rewritten in place each iteration
review-<n>.md # one per reviewer pass
log.md # append-only editorial log
Only the finished article and the editorial log are copied into the repo at
the end (content/<project>/briefs/). Pass paths, not contents to every
subagent — never paste the draft, notes, or reviews into a prompt.
The Process
digraph article_factory {
rankdir=TB; node [shape=box];
brief [label="Write source.md\n(brief + constraints)"];
research [label="Research agent\n→ research_notes.md"];
defgate [label="Working Definition\npresent + consistent?" shape=diamond];
outline [label="Outline agent\n→ outline.md"];
writer [label="Writer agent\n→ draft.md"];
lint [label="Orchestrator: lint.sh"];
reviewer [label="Cold reviewer\n→ review-n.md"];
verdict [label="Verdict?" shape=diamond];
deslop [label="catalog-deslop skill\n(word floor from brief)"];
ship [label="Copy to content/, commit,\nhonest report" shape=doublecircle];
brief -> research -> defgate;
defgate -> research [label="missing"];
defgate -> outline -> writer -> lint -> reviewer -> verdict;
verdict -> writer [label="REVISE, iter < 5"];
verdict -> deslop [label="ACCEPT, or\nbudget exhausted"];
deslop -> ship;
}
Step 0 — Brief
Write source.md yourself (this is state management, not prose): the topic as
a one-paragraph thesis, then ## Constraints:
- Word target (default 2000–3500; record the floor).
- First-person practitioner voice: what was tried and what failed, not analysis from above.
- The argument should feel discovered, not presented.
- No product-page tone. The piece must work for a reader who never uses the tool.
- Do not end with a call to action. End with the strongest idea.
- No em dashes, no semicolons, and nothing the lint patterns flag
(
../catalog-deslop/references/lint.sh).
If the user supplied a brief, use it verbatim and only fill in missing constraints.
Step 1 — Research
One agent, web access allowed:
You are a research agent for a long-form article.
Brief: <source.md path>. Write your notes to <research_notes.md path>.
Source priority is strict: first-person accounts first — engineering blog
posts, conference talks, postmortems, forum threads by people who did the
thing. Secondary analysis (news, listicles, other people's summaries) is
supporting material, never the backbone.
Every claim in your notes gets a "Sources:" line with the URL. A claim you
cannot source does not go in the notes.
End the notes with "## Working Definition": 2–3 sentences pinning down what
the article's central term means here. Then re-read your own notes and flag
every piece of evidence that only supports the claim if the term means
something else. Those items must be cut or the definition must be broadened
explicitly — pick one and do it before you finish.
Gate before proceeding: research_notes.md exists, has Sources lines, and has
a Working Definition section. Missing definition → redispatch, don't patch it
yourself.
Step 2 — Outline
One agent:
You are a narrative architecture agent.
Brief: <source.md path>. Research: <research_notes.md path>.
Write the outline to <outline.md path> with exactly these sections:
## Central Thesis — one sentence.
## Named Concept — coin one memorable name for the core idea (the Paul Graham
"Blub Paradox" move), or state "none earned" if the material doesn't support one.
## Assumption Being Dismantled — what the reader curren
阅读完整来源 ↗