- The gathering workflow turns a stream of sources into a short list with a rationale, the drafting workflow combines generation with evaluation against criteria, and the publishing workflow enforces a checklist.
- In each you leave the decision and the responsibility for content to a human.
- Start with one and measure the time saved after checking the result.
A content creator’s work consists of a few repeating phases: gathering material, writing a draft and publishing. Each has something AI does well and something that must stay with a human. Below are three concrete workflows in the setup described in Anthropic’s write-up on effective agents, that is with a fixed path instead of full autonomy. The choice of pattern is covered in the text Workflow of the week.
Workflow 1: gathering and selecting sources
The goal: turn a stream of sources into a short list with a rationale. The flow is a chain with routing. A trigger fetches new items from chosen channels, for example documentation, makers’ blogs and announcements. The first step summarizes each item in two sentences. The second classifies it by topic, which is routing to the right lists. The third rates relevance against explicit criteria: does it concern your audience, does it contain something new, does it have a primary source. At the output you get a list of five items with a rationale, and you decide what to write about. The rule for working with sources: check every claim in the text against the primary source, not against the model’s summary.
Workflow 2: a draft with an evaluator loop
The goal: get a draft that needs corrections, not rewriting. The pattern is an evaluator loop: one call writes a draft from your outline and sources, a second evaluates it against written criteria and returns remarks, and the first corrects it. Write the criteria down explicitly: agreement with sources, no claims without a basis, tone, length, structure. Anthropic’s write-up notes that this loop is effective when the criteria are clear and refining has measurable value. After two or three turns stop the loop and read the text by hand. Google’s documentation on generated content reminds you that models predict a likely sequence of words and can be wrong, so manual fact verification is necessary, also in the title, description and alternative texts.
| Pattern | Input | The human decides about | |
|---|---|---|---|
| 1. Gathering | A chain with routing | Sources from chosen channels | What deserves a text |
| 2. Draft | An evaluator loop | An outline, sources, criteria | Truth and tone |
| 3. Publishing | A chain with a checklist | A finished text | Sending and labels |
Workflow 3: a checklist before publishing
The goal: not to publish an error. The flow takes a finished text and runs successive checks. Does every claim have a source from the list? Is the title descriptive and not exaggerated? Does the description have a sensible length? Do images have alternative texts? Does the text have internal links to related materials? Does it require labelling the use of AI, according to the rules described in the articles Labelling AI content and AI-assisted content and Google’s rules? The result is a report with remarks, not automatic publishing: sending stays manual, because a published error cannot be taken back without a trace. Criteria for people-first content, on which you can base the checklist, are given by Google’s page on creating helpful content.
- 01GatheringSummary, classification, relevance rating.
- 02SelectionA human decides what to write about.
- 03DraftGeneration and evaluation in turn.
- 04VerificationBy hand, in primary sources.
- 05CheckA pre-publication list, a report of remarks.
- 06PublicationManual approval and labels.
Common mistakes
The first mistake is automating everything at once. Start with one workflow and add the next only when it works. The second is the lack of explicit criteria for evaluating a draft: without them the loop corrects what it itself deems better. The third is skipping verification because the text “sounds good”: a model can sound confident even when it is wrong. The fourth is blurring your own voice: if every text goes through the same template, they all sound alike. The fifth is no record of the process: in half a year you will not remember what criteria the loop had or where sources came from.
How to build and measure them
You can build the workflows in a tool such as n8n (n8n from scratch) or in a simple script. Start with one, preferably gathering, because it has the lowest risk. Measure the baseline: how long this stage takes you today. After two weeks compare, including the time to check the result. If the saving disappears in corrections, change the criteria or the pattern. Make sure the text still sounds like you and contains your analysis, because merely processing others’ sources adds no value, as Google’s criteria remind you. A good starting point for topics is audience questions, as in the article From customer questions to a post. You will find all materials in the Workflows of the week hub.
Last checked: 20 October 2026. Tools and platform rules change; the flows keep their sense.