AI Prepares; A Person Publishes
An AI content pipeline can research, draft, format, and prepare images while you sleep. It is still not the publisher.
The person responsible for the website remains responsible for the final page. Name that person before the run starts and give them the authority to stop it.
Build Gates, Not One Giant Prompt
Treat the pipeline like a production line with inspection gates. Automation moves the work; each gate decides whether it can continue.
The final gate is human. A second AI pass can find problems, but it cannot approve its own output.
Keep Evidence Attached
A polished sentence is not proof. Material claims need a direct source, while project facts and unconfirmed points need clear labels.
How n-g.be Runs It
A scheduled agent starts nightly from a topic roadmap, a benchmark article, formatting rules, and templates. It researches, drafts, edits, and prepares branded desktop and mobile thumbnails. When the roadmap runs out, the schedule stops.
A person skims the pack, then a second AI pass edits it for the site format and checks the claims. The thumbnails are converted to WebP and the article is published. It is read again on the live page and revised months later with an updated date.
This is the small-site version of the release gate: a close read after publication, not before. A larger site should not copy that part.
Maintenance Is Part of Publishing
The workflow does not end when the page goes live. Read the rendered page, fix what the preview missed, and set the next review date.
For a material revision, show an honest updated date and keep the visible and structured dates consistent.
Steps
Guide
Name the Release Owner
Write down one person who owns the release decision. 'The content team' is too vague. The owner needs the context, time, and authority to reject the pack.
AI can prepare evidence. A second model can find formatting mistakes, unsupported claims, or awkward wording. Neither action transfers responsibility to the tool.
The NIST AI Risk Management Framework calls for clear roles, accountability, ongoing monitoring, and periodic review. The same practical logic fits a content pipeline.
For a small, low-traffic site, the owner can skim the complete pack before release and perform a closer read on the live page. A larger site needs a named person who approves before anything becomes public. High-risk subjects need stronger specialist review regardless of site size.
- 2
Lock the Input Pack
Read GuideDo not start with a blank 'write an article' prompt. Give the run a controlled input pack.
- approved topic roadmap
- one selected topic with a stable ID
- benchmark article from the same series
- editorial rules and the correct content template
- internal-link map
- source and evidence standard
- image references and required dimensions
- named human owner
Separate the roadmap from the operating state. The roadmap says what may be produced and in what order. The state records what is reserved, complete, or blocked.
If the roadmap, required template, or reservation state cannot be read, stop the run. Guessing at the source of truth is an untracked editorial decision.
Reserve One Approved Topic
Take the first eligible item from the approved queue. Write its exact ID, title, and start time into durable state, then read the state back.
Only after verification should research begin. This prevents two scheduled runs from quietly producing the same article.
If a run fails halfway through, leave the item reserved until a person decides whether to release or retry it. Do not let tomorrow's run repeat it automatically.
The agent may sharpen the angle inside the approved scope. It may not silently select another topic, rename it, or insert a fashionable idea into the queue.
Research Before Drafting
The research stage should produce a fact pack, not polished article copy. Keep four kinds of material separate.
- verified external facts with direct sources
- project facts supplied by the owner
- interpretations and recommendations
- claims that could not be confirmed
Plan the source queries before browsing and prefer primary sources. Keep every material claim traceable. If sources conflict, use the safer interpretation or show the disagreement.
Google's guidance on generative AI content says AI can help with research and structure, while automatically generated content must still focus on accuracy, quality, and relevance.
Do not optimise the pipeline for page count. Google's scaled content abuse policy covers large amounts of unoriginal, low-value content made mainly to manipulate rankings, regardless of how it was produced.
- 5
Draft, Format, and Build the Assets
Read GuideDraft from the approved strategy brief and fact pack. If the prose needs a new external claim, return it to research instead of inventing a source during drafting.
Convert the draft into the site's real code structure. Add the summary, description, search metadata, series relationship, internal links, banner paths, ordered steps, checks, and common issues required by the renderer.
Generate separate desktop and mobile thumbnails from one visual idea. Check both dimensions, integrate the brand symbol deliberately, and convert the final files to WebP before publication.
The complete pack contains the final article source, SEO title, meta description, slug, grounded internal links, desktop and mobile images, source list, and material uncertainty notes.
- 6
Use the Second AI Pass as a Checker
Read GuideRun a second AI pass against the final pack. Its job is to compare the output with the rules and evidence, not to offer general writing advice.
- verify the exact title, scope, template, and series relationship
- trace material claims to the fact pack and original sources
- test links against the approved link map
- check image dimensions, filenames, and visual angle
- find missing uncertainty notes
- flag invented examples or unsupported promises
- check readability and British spelling
Any material rewrite must return through the relevant checks. The second pass should report what changed. It must not approve its own work.
Give the Human a Real Release Gate
The owner reviews the whole pack, not only the draft. Show the final title, summary, metadata, links, images, sources, uncertainty notes, and rendered preview together.
The decision has three useful outcomes: approve, return with required changes, or block. Silence is not approval. A passing AI check is not approval.
Google's current guidance focuses on useful, reliable, people-first content rather than the tool used to produce it. The release gate asks the same practical question: does this page genuinely help its intended visitor?
Read the Live Page and Set the Next Review
Open the live URL on desktop and mobile. Read the page again in its real layout. Test the table of contents, internal and external links, banner crop, metadata, dates, and copyable script.
Fix what the preview missed. Record the live URL, owner, publication date, and next review date in the audit trail.
When the page is materially revised, use an honest updated date. Google recommends consistent visible and structured dates in its publication date guidance.
- 9
Use Behaviour Data When There Is No Blog Queue
Read GuideNot every site needs a blog calendar. A hub site can use the same controlled pipeline with a different trigger: observed behaviour.
GA4 events can measure page loads and link clicks. Clarity heatmaps aggregate clicks and scroll reach, while session recordings recreate interactions across a visit.
A pattern is a hypothesis, not an explanation. If almost nobody clicks below the hero cards, a person decides what that may mean. The team can prioritise navigation and hero cards, release a controlled change, then check the data again.
A nightly agent can also prepare a feature from an approved queue and open a pull request. It should never merge or deploy. A person reviews the code, preview, tests, and expected effect before release.
Be Aware
The agent chooses or rewrites the topic.
Keep the canonical roadmap separate from operating state and verify one exact reservation before production.
The second model becomes the approver.
Use it to produce findings for a named person. Record the human decision separately.
The reviewer sees only the prose.
Present the final code, metadata, links, images, sources, uncertainty notes, and rendered preview as one pack.
The pipeline rewards output volume.
Measure usefulness, correctness, and maintained quality instead of the number of pages generated.
The live article is never revisited.
Schedule the live-page check and later review during the same publishing run.
Human-Owned Content Pipeline Run
Copy / paste
Run one controlled content-production cycle.
Before production:
1. Read the canonical topic roadmap, durable state, benchmark article, editorial rules, article template, internal-link map, and image references.
2. Identify the first eligible approved item in exact roadmap order.
3. Reserve its exact ID and title in durable state, read the state back, and verify the reservation.
4. Confirm the named human release owner.
Stop before drafting if the roadmap, state, required template, benchmark, or release owner is missing. Do not select a fallback topic or invent missing rules.
Production:
1. Create a strategy brief without changing the approved title or scope.
2. Build a fact pack from primary sources. Separate verified facts, project facts, interpretation, recommendations, and unconfirmed claims.
3. Draft from the strategy brief and fact pack.
4. Convert the draft into the required site format.
5. Prepare SEO metadata, grounded internal links, desktop and mobile images, and material uncertainty notes.
6. Run a second AI pass against the rules and sources. Report every material edit.
7. Present the complete pack and rendered preview to the named human owner.
The second AI pass may check and edit. It may not approve, publish, merge, or deploy.
Release:
1. Record the human decision: approve, return, or block.
2. Publish only after approval.
3. Read the live page on desktop and mobile; test links, images, metadata, dates, and interactive elements.
4. Fix live issues and record the next review date.
5. Only after complete delivery, mark the exact roadmap item completed and append the audit entry.
If the run fails after reservation, leave the item reserved and report the exact recovery action. Never retry it automatically.
