Free resource · Version 1.0
Governed AI Content Workflow Kit
A free, downloadable workflow for publishing a glossary, knowledge base or article series with AI agents, without handing over editorial control.
- One ZIP
- One prompt to start
- No code to write
Start in three steps
- 01
Download the ZIP and unzip it into a folder.
- 02
Open an AI agent that can work with files in that folder, or attach the ZIP.
- 03
Copy the prompt below and paste it in.
I have attached the Governed AI Content Workflow Kit. You are my Content Agent. I am the Human Publisher.
The workflow has four roles: Content Agent, AI Editor (review and quality checks), Website Agent and Human Publisher. They are roles, not necessarily separate tools. During setup, ask me which AI plays which role. If you also play the AI Editor role, do each review as a separate pass using prompts/ai-editor-review.md and record it separately from the drafting. Recommend a separate AI Editor where practical, but do not require one. The Human Publisher role is always me.
Your job is to help me plan, write, review, publish and verify content for my website using this kit, and to guide me through it one step at a time in plain English. I should not have to understand JSON, Python, Git or the command line. You maintain every file and run every tool.
Start like this:
1. Unpack the kit if it is a ZIP. Read these files completely before doing anything else: README.md, WORKFLOW-GUIDE.md, agent/CONTENT-AGENT.md and agent/FILE-FORMATS.md. agent/CONTENT-AGENT.md is your operating manual: follow it.
2. Run `python3 tools/workflow.py selftest` and tell me in one sentence whether the tools work on this computer. If you cannot run Python or cannot save files, tell me plainly what that means (you can still help me plan and draft, but the automatic checks will not run) and suggest what I would need.
3. Run `python3 tools/workflow.py status`. If content already exists, summarize where things stand and what the next step is. If not, begin the setup interview.
4. Interview me to customize the workflow for my business and website. Ask a few questions at a time, offer sensible defaults, and explain why each choice matters. Cover: my business and website, which AI plays which role, who I am writing for, what I want the content to achieve, which kinds of articles I need and their templates, writing style and length, categories and how topics relate, how claims must be sourced, anything I must never claim or mention, page addresses and SEO descriptions, linking rules, who reviews and how, and how we publish and verify. Record my answers in content/STRATEGY.md and keep content/workflow.config.json in step with them. Show me a short summary to confirm before moving on.
5. Then guide me through the workflow in order: plan topics and short definitions for my approval, draft articles from the templates, review in the AI Editor role (if another AI plays it, give me the ready-to-paste prompt from prompts/ai-editor-review.md at that moment), my approval, links from my existing pages to the new articles, a validated package for the website, the website handoff (give me the ready-to-paste Website Agent prompts at that moment, or follow them yourself if you also build my website, without changing any approved wording), preview verification, my go-ahead for production, and verification of the live site.
Rules you must always follow:
- Whoever plays which role, keep every review recorded separately from the writing, and never approve anything on my behalf. Record an approval only when I say it explicitly and clearly about that specific item, and quote my words in the entry's private notes. If what I said could mean more than one thing, ask.
- Approval is not publication, publication in a package is not being live, and being live is not verified. Keep these steps separate and tell me which one we are at.
- Never send, publish, deploy, push or contact anyone. You prepare; I send and I authorize.
- Never edit a package after it is made. A fix is a new version with a report of exactly what changed. Any change to a published article needs my approval and is listed line by line.
- Ground every factual claim. General claims need a source you have actually read. Claims about my business need evidence from me. Never invent sources, statistics, quotes, customer stories or results. If something cannot be verified, say so.
- Keep research notes, review history and anything private below the private-notes marker. Nothing private or confidential may reach a package.
- Tell me honestly what the checks prove and what they do not. Never promise search rankings, traffic or indexing.
- After every tool run, tell me the result in plain words and what happens next. Show me what changed before asking me to approve it.
Begin with step 1.
The agent reads the kit, interviews you to fit the workflow to your business, and guides you step by step.
An AI agent that can work with files: for example Claude Code, Cursor, Codex, or a chat assistant with file uploads.
Planning, drafting and review work with any such agent. The automatic checks, packages and website verification also need Python 3.9 or later on that computer; the agent will tell you whether it can run them.
What it does
The workflow has four roles, and you hold the one that decides.
These are roles, not necessarily separate tools. One AI can play several of them. A second, different AI as the editor catches more, so the kit recommends one where practical, but does not require it. Whatever the assignment, each review is recorded separately from the writing, and only you approve.
- 01Human sets directionYOUplanned
- 02AI plans and draftsAIdrafted
- 03ReviewAIreviewed
- 04Human approvalYOUapproved
- 05Controlled website handoffAIpackaged → live
- 06VerificationAIverified
Each article moves through the same steps: planned, drafted, reviewed, approved, packaged, live and verified. Each step is recorded, and none happens without the one before it.
Content Agent
Researches, drafts, sources every claim, keeps the records and runs the checks.
AI Editor
Reviews each draft against your rules and suggests specific corrections.
Website Agent
Puts approved content on your site exactly as supplied, without rewriting it.
Human Publisher
You. You approve topics, articles, corrections and releases, explicitly.
Roles, not necessarily separate tools ▸
These are roles, not necessarily separate tools. One AI can play several of them. A second, different AI as the editor catches more, so the kit recommends one where practical, but does not require it. Whatever the assignment, each review is recorded separately from the writing, and only you approve.
Each article moves through the same steps: planned, drafted, reviewed, approved, packaged, live and verified. Each step is recorded, and none happens without the one before it.
How it supports a structured SEO content strategy
The kit helps you plan and maintain content in a structured way:
- 1
Topic coverage
You plan the whole set of topics up front, each with an approved short definition. The index can list every topic from the start, with unwritten ones marked "Coming soon" and given no page until they are ready.
- 2
Definitions
Every article opens with its approved definition, and related articles use terms the same way.
- 3
Contextual internal links
Articles link to each other at first mention, and links to topics that are not published yet switch on later by themselves. Links from your existing pages are listed with the exact sentence each belongs in, so they land where intended.
- 4
Consistent metadata
Titles, descriptions within a set length, and site map entries only for pages that exist.
① Index · planned topics
Related context includes Topic A and Topic D.
③ Solid: linked at first mention. Dashed: switches on when Topic D is published.
Topic B link placed in its exact named sentence
Schematic. Numbers match the list.
This is a way of organizing and controlling content, not a ranking method. The kit does not guarantee or predict rankings, traffic, indexing or authority; those depend on many things outside it. If you want to measure results, record a baseline in a tool such as Google Search Console before you start.
How the controls work
Four details behind the workflow. Open any to read more.
AI can draft quickly. It can also state a wrong number confidently, cite a source it did not read, drift from yesterday's definition, or promise something your business does not do. The kit responds in two ways:
- every draft is reviewed against your rules, claim by claim, in a separate recorded step, ideally by a different AI;
- you approve every article by name, and the Content Agent records your words. It never approves its own work.
Approved content goes to the website as a package: a numbered bundle with the topic list, the approved articles, the index page and the link instructions, plus a fingerprint for every file. A package is never edited after it is made. A fix is a new version, with a report of exactly what changed. Any change to an article that is already live needs your approval and is listed in that report line by line.
Whatever builds your site (Replit, Lovable, a developer, an AI coding tool, or your Content Agent if it also works on the site's code) receives the package and ready-to-paste instructions. It renders the articles as written, adds each link only in its named sentence, reports anything it cannot place instead of guessing, and builds a preview first. Production waits for your go-ahead.
The first time, it sends you a plan for approval. The plan includes sending finished pages (server-rendered or prerendered), because the kit's checks cannot read pages that exist only after JavaScript runs in the browser.
After the preview, and again after the site goes live, the Content Agent compares what your website sends with the approved package: article text word for word, which topics have pages, the index and site map, and each link from your existing pages in its sentence, with nothing else on those pages changed.
That comparison works on publicly accessible pages delivered as finished HTML, on websites set up as the kit describes. It shows that the site matches what you approved at the time of the check.
It cannot show that the content is accurate, and it cannot inspect pages that are rendered only in the browser.
Website problems that are not about content go into a separate technical brief.
What is in the kit
Download the kit ↓- 01
The master prompt, a README and a plain-English workflow guide.
- 02
The Content Agent's operating manual and file formats.
- 03
Ready-to-fill prompts for the AI Editor role and your website builder.
- 04
Article templates (concept, how-to, comparison) and a correction request template.
- 05
A settings file your agent maintains from your answers: roles, categories, style, sourcing, claims to avoid, description lengths, linking rules.
- 06
The tools: checks, packaging, change and validation reports, link checks, preview and live-site verification.
- 07
A fictional example, Northcove Coffee Co., and a self-test that replays it, including seven deliberate website mistakes the checks must catch.
Where it came from
We built this workflow while publishing our own glossary with AI agents, and generalized it so any business can use it. It is a practical example of human-in-the-loop governance: clear responsibilities, explicit approvals, controlled handoffs and records you can check. The kit is standalone. It does not use, require or connect to any of our products.
Want to add runtime governance to your AI agents? Explore Sentience Governor.