The workflow
From request to published page
Hover over a step to see the details. Move your cursor away to close them. Keyboard: Tab to a step, Esc to close. On a touchscreen, tap to open.
The task
Step 01 · Generator
Brief
A request like “we need something about X” becomes an editorial task. The brief records:
- who the content is for;
- which reader problem it solves;
- what the reader should learn or gain;
- why the business needs it;
- which metric will show whether it worked;
- how costly a mistake would be: this decides whether the piece takes the full route or the short one.
Working hypotheses are recorded separately. These are assumptions that sound plausible but have not yet been proven.
Tools and inputsA model for complex reasoning · a brief template · version history in Git
Step 02 · Generator
Review existing content
The agent checks what the company has already published on this topic: which search queries overlap, which terms are already defined, and whether any pages contradict each other. It then recommends one of three options:
- create a new piece of content;
- add to an existing one;
- merge several pages.
The editor makes the choice at checkpoint 1.
Without this check, the website fills up with duplicate articles, different definitions of the same term, and advice that contradicts itself. That makes it harder for search engines and AI assistants to identify which version is current and reliable.
Tools and inputsVector search · a list of published pages · a glossary of terms and entities
Checkpoint 1 of 4 · Editor
Is the idea approved?
Checkpoint 1 · Human
Is the idea approved?
The editor looks at the audience, the purpose of the content and the expected result. They choose between a new page, an addition and a merge, and confirm the risk level. Approving a topic does not mean approving claims that have not yet been checked.
The decision is recorded together with the name of the responsible person, the date and any approved exceptions.
RoutingYes → step 03 · No → step 01 · Topic rejected → archive, with the reason recorded
Reader questions
Step 03 · Generator
Research search demand and AI answers
Before any evidence is collected, the system defines the information need that the future content must meet. It looks at several things at once:
- real questions from users;
- keywords and search queries;
- traditional search results;
- answers from popular AI assistants;
- pages that AI uses most often as sources;
- how easy it is to find and extract information from the brand’s own website;
- key concepts, companies, products and terms linked to the topic.
Based on this data, the agent proposes the main search query, secondary queries, related topics and the user questions the content must answer. The editor and the expert approve this list at checkpoint 3, together with the structure.
The result is a map of information gaps — a list of questions that existing content answers poorly: incompletely, inaccurately, in contradictory ways, or not at all.
This list tells step 04 which claims need evidence.
Tools and inputsA search results API · test queries sent to AI assistants · on-site search history · user support requests · vector clustering of queries and documents · a model that analyzes and groups the results
Evidence
Step 04 · Generator
Gather evidence
The work has four parts:
- finding and downloading sources;
- interviewing company experts about what the documents do not cover;
- splitting the content into separate claims that can be checked;
- checking that all required fields are filled in.
Every claim gets a unique ID, an exact quote, a date and a defined scope. A study result for one market, for example, cannot be applied to every country without extra evidence.
An expert interview is a regular input, not an exception. The answer is saved as a source with the expert’s name, the date and a defined scope. The expert confirms the wording in writing.
The team also records where the information came from. A full page and its short description in search results are not the same thing: a search snippet helps you find the right source, but it cannot confirm facts, numbers or exact wording.
You can use a faster, lower-cost model to find sources. But it is better to use a stronger model to read them and turn them into separate claims.
Tools and inputsA lower-cost model for search · a stronger model for reading sources · an automated browser · a PDF parser · saved copies of sources with hashes · transcripts of expert interviews
If a claim has no evidence after two attempts, it is removed from the register or sent to the editor.
Step 05 · Checker
Independent verification
The checker downloads every source itself and checks three things:
- the quote is present in the source;
- it supports exactly the claim that was written down;
- the claim does not go beyond its stated scope.
It also looks for broken links, unsupported generalizations and mistakes in attribution — for example, when one person’s or company’s words are attributed to another.
For an expert interview, the checker compares the claim with the transcript and with the wording the expert confirmed.
If the checker cannot download pages, it can only judge whether the reasoning is consistent. In that mode it cannot confirm that the source contains the fact: the link may be real, but the required number may not be on the page.
Tools and inputsA model from a different provider · a required second download of the source · a verification result stored in the metadata
If a source was misinterpreted, the work goes back to step 04.
Checkpoint 2 of 4 · Code
Has the evidence passed the check?
Checkpoint 2 · Code
Has the evidence passed the check?
The orchestrator — the program that starts the steps and passes the work between them — reads the status of the check from the metadata. If the evidence is blocked, the next step simply does not start.
This is what separates a process from a set of prompts. Here the status BLOCKED is not a comment for the next agent. It is a hard stop.
Retries are capped. After two failed attempts, the claim isn’t sent back again: it is removed from the register, and the editor decides whether the content still works without it.
RoutingPASS → step 06 · No source → step 04 · An expert is needed → interview, then step 04 · Two failed attempts → the claim is removed, the editor decides
Content design
Step 06 · Generator
Build the content brief
The core of the specification is a matrix:
section → claim IDs → sources → check status
A section can enter the outline only when its claims have passed verification.
For every section, the team writes one key takeaway — the single idea the reader should remember. The system then checks whether the chosen claims really support it.
Search demand shows which questions are worth covering. The evidence defines what each section can claim.
If a question from step 03 has no verified claims, it goes back to step 04 or is dropped from the outline.
Tools and inputsA model for complex reasoning · the claim register · data about search and AI answers · the list of published pages
Checkpoint 3 of 4 · Editor & expert
Is the structure approved?
Checkpoint 3 · Human
Is the structure approved?
The editor and the expert approve the whole evidence matrix, not just a list of headings.
They also confirm the search queries and the questions from step 03.
If a section has no confirmed claims, it goes back to research or is deleted. The expert’s answer is also recorded as a source, with a name, a date and a defined scope. Otherwise, the fact-check will treat it as unconfirmed.
RoutingYes → step 07 · The structure needs work → step 06 · Evidence is missing → step 04
Production
Step 07 · Generator
Draft
The model writes the text using only the approved claim register. It is free to choose the structure, the rhythm and the wording, but it cannot change or add facts.
The draft exists in two layers:
- clean text for the reader;
- a working version where every factual sentence is linked to a claim ID.
With the second layer, the checker doesn’t have to guess which source a sentence is based on.
Tools and inputsA model tuned for prose · a tone of voice guide with no factual information in it · the claim register
Step 08 · Checker
Check facts and conclusions
This step has two separate checks.
- Fact-check: every factual sentence is matched to a specific claim by its ID.
- Conclusion check: the system decides whether the conclusion of a section really follows from the sources provided.
The second check finds a particularly dangerous type of mistake: every quote is correct on its own, but the cause-and-effect story built from them is not supported by any source.
Tools and inputsA model from a different provider than the model used at step 07 · matching by ID · an error classifier
If a fact has no claim ID or a conclusion does not follow from the sources, the draft goes back to step 07. After two returns, the editor steps in.
Getting ready to publish
Step 09 · Generator
Make answers easy to find and cite
The goal of this step is to help a reader, a search engine or an AI assistant find and quote the right answer quickly. The content is updated to include:
- a direct answer at the beginning;
- self-contained sections that make sense on their own;
- tables with consistent column types;
- an FAQ section;
- metadata;
- internal links to pages that exist.
After this work, the set of claim IDs must stay the same. A new fact has to pass step 04 first.
Tools and inputsA mid-tier language model · an up-to-date list of website pages
Step 10 · Generator
Edit for clarity and style
The model removes empty generalizations, promotional language and stock phrases common in AI-generated text. For every such phrase, it checks whether there is a proven fact behind it:
- if there is a fact, the vague wording is replaced with a specific one;
- if there is no fact, the phrase is deleted.
You cannot take a suitable fact from another part of the register just because it sounds convincing. This can change the meaning or scope of the claim.
Everything added at step 09 must survive the edit: the direct answer, the FAQ and the self-contained sections. Code checks this.
After this kind of editing, the text often becomes 15–30% shorter.
Tools and inputsA language model · an editorial guide · a check for stock phrases and repetitive wording · automatic checks of numbers, dates, negations, and attribution
Step 11 · Checker
Legal and product risks
Whether the text sounds like the company, and whether the company is allowed to say it, are two different questions. This is why style and risk are checked separately, with different people responsible for each.
Claims about the company are compared with the approved fact sheet. Wording about promised results, licenses, competitor comparisons and other risks is checked against the compliance rules.
If there are no such rules, the process stops. The model must not replace company or legal requirements with its own “careful” wording.
Tools and inputsA separate checker model · the company fact sheet · the compliance rules
Risky wording goes back to step 10. If the rules do not cover the case, the person responsible for compliance decides, not the model.
Step 12 · Checker
Check that the meaning is intact
The final review checks that editing has preserved the meaning:
- all the claims are still there;
- the scope of each claim has not expanded;
- limitations and attribution have not been lost;
- the conclusions still follow from the sources;
- the required sections from the specification are present;
- no new facts have appeared.
This check is needed because the text goes through several more rounds of editing after the fact-check. Without it, the version that was checked and the version that is published can be noticeably different.
The checker reviews the text from the start without seeing the results of earlier checks.
Tools and inputsAn independent checker · a separate checking script · automatic checks of the required conditions
If a claim was lost, widened or added, the text goes back to step 10.
Step 13 · Human
Editor’s read-through
An editor reads the whole text the way a reader would. The reports from steps 08, 11 and 12 are open beside it.
The editor looks at what models judge poorly:
- whether the text answers the question from the brief;
- whether the argument holds from start to finish;
- whether it sounds like the company, not like a model;
- whether any phrase would embarrass the company if quoted out of context.
The editor can change the wording. If an edit touches a factual sentence, the text goes back to step 12.
The editor’s name and the date are saved in the log.
Tools and inputsThe full text · the check reports · the brief · the editorial guide
Step 14 · Code
Technical checks and version freeze
Code, with no model involved, checks:
- whether external links open;
- whether internal pages exist;
- whether the markup matches the visible text;
- whether the metadata is filled in;
- whether any internal notes remain in the content;
- whether the preview looks correct;
- whether a rollback point has been created.
After the check, the version is frozen. Any later changes require a new check.
Tools and inputsA link checker · a markup validator · a preview renderer · creation of a hash and a rollback point
If a fix does not change the visible text, only this step runs again. If it does, the text goes back to step 12.
Checkpoint 4 of 4 · Person responsible for publishing
Do we publish this version?
Checkpoint 4 · Human
Do we publish this version?
The person responsible for publishing sees one package: the frozen text, the editor’s sign-off, the reports from every check and the list of accepted exceptions. They approve the whole package, not just the text.
A “no” always names the reason, and the reason sets the route: facts go back to step 07, wording to step 10, risks to step 11. After the fix, the text passes the later checks again and the version is frozen again.
The log records their name, the time and the version of the package.
RoutingYes → step 15 · No, facts → step 07 · No, wording → step 10 · No, risks → step 11 · Publication declined → archive, with the reason recorded
Step 15 · Code
Publish and verify the live page
After approval, the content goes to the content management system and is published with no extra changes. No generative models are used at this step, which ensures that the published version is exactly the version that passed all the checks.
After publishing, the system automatically opens the live page and compares it with the approved version of the document. This confirms that nothing changed during publishing: no blocks of text disappeared, the markup did not break, and no formatting errors appeared.
Tools and inputsThe CMS API · a second download of the published page · the saved approved version of the document · a rollback mechanism
If the live page differs from the approved version, the system rolls the publication back and creates a task.
After publication
Step 16 · Code
Monitor and update
The work does not end when the content is published. The system tracks:
- the metric defined in the brief;
- whether the content appears in answers from AI assistants;
- whether sources are still available and whether they changed;
- how long each product fact remains valid;
- the date of the next review;
- who is responsible for updates.
All links are downloaded again on a regular basis. If a source has disappeared, changed, or stopped supporting the claim, the system creates a task to check it.
An update does not start from scratch. The content goes back to the step the change affects:
- a source changed or disappeared → step 04, for that claim only;
- a product fact expired → step 04;
- the metric from the brief is not moving → step 01, with a new brief.
From there the content follows the route as usual. The system marks related posts and translations for a new check.
Tools and inputsA scheduler · repeated downloading of links · measurements of how often the page is cited · logs of visits by AI bots
Parallel branch · separate output
Content for social media
A post is not just a shorter article. It has a new hook, a different context and shorter wording. During this kind of rewriting, it is easy to lose an important limitation or add a claim that was not in the original content. This is why a post is checked as a separate output.
Translations and versions for other markets follow the same branch. Only the base is different: a glossary and the legal limits of the market instead of a platform playbook.
Runs on the approved document only.
Publishing waits until step 15 is live and verified.
-
BaseHuman Platform rules
Base · Human
Platform rules
Every channel needs its own playbook: format requirements, platform conventions, access rights, legal limits, metrics and a review date.
There is no need to research the format of LinkedIn, Telegram or any other platform again for every article. The playbook is kept as a separate working document.
-
S1Generator Adaptation
Step S1 · Generator
Adaptation
The format comes from the platform playbook. The content uses only claims from the approved document.
Unapproved content is not published.
-
S2Checker Review the social posts
Step S2 · Checker
Review the social posts
Every fact is linked to a claim ID again. The checker looks at:
- whether a new fact appeared in the headline, the hook or the caption;
- whether important limitations survived the shortening;
- whether the meaning of the original claims changed.
If a new fact appeared or a limitation was lost, the post goes back to S1.
-
S3Human Editor’s approval
Step S3 · Human
Editor’s approval
An editor approves every post before it goes out. The name and the date are saved in the log, as at the checkpoints of the main route.
A “no” sends the post back to S1, with the reason recorded.
-
S4Code Publishing
Step S4 · Code
Publishing
The content is published through an API without any calls to generative models. Publishing starts only after the main page has gone live and passed its post-publication comparison.
When the article is updated, the system automatically marks related content for a new check.