Topic list to checked drafts through a code QA gate

Topic list to checked drafts through a code QA gate

An n8n workflow that takes a list of topics and turns each one into a draft article. It researches the topic, writes an outline,
writes the draft, and then checks the draft with plain code before anything is filed. A draft that passes goes to your
publishing endpoint as a draft. A draft that fails goes back with the reasons.

This is a skeleton, not a ready service. Every outside step is an HTTP call to a REST API that you provide, and the file names
four endpoints it expects. Read "Before you start" first.

What is inside

• workflow.json: the workflow as n8n exports it, 18 nodes, one of them a sticky note. It holds no credential, and no HTTP node
  sends a key.
• README.md: these notes.

What we changed from the original copy

• The file had two sticky notes, a read-me and a setup list, both with promotional links. They are replaced by one short title
  note, and the setup is written here.
• Nothing else. There was no credential to remove and no code that needed repair.

Before you start

• An n8n instance.
• Your own REST API at one address, with four endpoints. The workflow does not call a model or a search service directly. Your
  API does.
• POST /search takes {"query": "<topic>", "recency_days": 90} and replies with {"results": [...]}.
• POST /llm/complete takes {"task": "outline", "input": {"topic", "sources"}} and replies with
  {"output": {"title", "sections"}, "usage": {"input_tokens", "output_tokens"}}. The same endpoint takes
  {"task": "draft", "input": {"topic", "title", "sections", "sources"}} and replies with {"output": {"title", "body"}, "usage": {...}}.
• POST /alerts takes {"channel", "text"}, plus an approval field on the approval message. Any reply is fine.
• POST /publish/doc takes {"title", "body", "status": "draft"} and replies with {"url"}.

Set it up

1. In n8n, create a workflow, open its menu and choose Import from File. Pick workflow.json.
2. In Config, replace the base value, https://api.your-company.example, with your API's address. Do the same in
   Config (errors).
3. In Config, set min_words, which is 120 as shipped, and auto_approve_in_test.
4. Activate the workflow. Copy the production URL of Webhook / weekly schedule: topics. Its path is content-run.
5. Send a POST request to that URL with a JSON body, for example {"topics": ["first topic", "second topic"]}. A string of
   topics separated by commas works too. Add run_id if you want your own run name.

How a run goes

1. Webhook / weekly schedule: topics receives the request. Config adds the settings, and One item per topic is a Code node
   that makes one item per topic. It stops with "No topics supplied" when the list is empty.
2. Loop topics takes one topic at a time. When none are left it sends everything to Run summary (respond).
3. Agent 1 · Research (search API) posts the topic to /search with a 90 day window.
4. Agent 2 · Outline (LLM) posts the topic and the search results to /llm/complete and gets a title and sections.
5. Agent 3 · Draft (LLM) posts the outline and the sources to the same endpoint and gets a title and a body.
6. Agent 4 · QA gate (deterministic checks) is a Code node. It marks the draft as failed when any of these is true:
   - the body is shorter than min_words
   - it has no Sources: line
   - it holds one of the phrases guaranteed, risk-free, 100% or best in the world
   - it lacks the headings ## Why it matters now or ## What to do this quarter
   - it has fewer than two web links

It also adds up the tokens from the outline and the draft, and estimates a cost.

1. QA passed? sends a passing draft to Human approval (Slack), which posts a message to /alerts in the channel
   #content-approvals. The next node, Publish as draft (Google Docs / CMS), posts the draft to /publish/doc at once.
2. A failing draft goes to Send back with reasons (no publish), which posts the reasons to /alerts in the channel
   #content-qa-failed.
3. Run log row writes one row for the topic: title, word count, pass or fail, issues, the published address if there is one,
   the estimated cost and the tokens. The loop then takes the next topic.
4. Run summary (respond) counts the topics, the drafts filed and the failures, adds the costs, and the webhook returns that
   result to the caller.

The error branch is separate: Error trigger runs Config (errors) and Alert ops, which posts "Content pipeline failed" with
the error message to /alerts in the channel #ops-alerts.

Change the look

• The checks: the Agent 4 · QA gate (deterministic checks) code holds the word count, the banned phrases, the two required
  headings and the link count. Change them to your own rules.
• The research window: recency_days in Agent 1 · Research (search API).
• The channels: #content-approvals, #content-qa-failed and #ops-alerts are plain text in three nodes.
• The cost: the estimate uses 3 and 15 dollars per million input and output tokens, written in the QA code. Put your model's rates there.

Good to know

• Nothing waits for a person. Human approval (Slack) only posts a message, then the draft is filed straight away as a draft.
  The approval field of that message reads auto-approved when auto_approve_in_test is on and awaiting human when it is
  off, and the flow behaves the same in both cases. For a real approval, add a Wait node after the message and resume it from your
  own tool.
• The name says weekly schedule, but the only trigger is the webhook. Add a Schedule Trigger if you want a weekly run.
• The webhook has no authentication, so anyone with the address can start a run, and every run spends your API. Turn on
  Header Auth in the webhook node. None of the HTTP nodes sends a key either. Add authentication in each one if your API needs it.
• The caller's request stays open until every topic is done, because the webhook answers with the last node. A long list can
  pass a time limit in front of n8n. Keep lists short, or answer at once and send the summary somewhere else.
• The summary's run_id comes out empty, because Run log row does not carry it. Add run_id: q.run_id to the object it returns.
• The error branch runs for automatic runs only. A run you start from the editor does not trigger it. n8n uses a workflow that
  contains the Error Trigger as its own error workflow by default.
• The QA gate checks form, not truth. It cannot tell whether a link is real or a claim is correct. Read the drafts.
• Your data: the topics, the sources and the drafts go to your own API and whatever service sits behind it.