A queue of generated drafts becomes difficult to manage when “done” can mean researched, written, reviewed or actually published. Make those meanings explicit before adding automation. The following design is an illustrative model for a solo editor, not a tested workflow engine.
Define the states
Use candidate, researched, drafted, reviewed, published and needs_update. A candidate has a reader question. A researched item has sources and unresolved points. A drafted item has text linked to those notes. Reviewed means a human has accepted the final wording and asset choices. Published means a public address has been checked. Needs_update records an issue discovered later.
Write transition requirements
| From | To | Required evidence |
|---|---|---|
| candidate | researched | Source notes and unknowns |
| researched | drafted | Draft with claim references |
| drafted | reviewed | Human review decision |
| reviewed | published | Verified public URL |
| published | needs_update | Specific issue and affected claim |
A model finishing its response does not satisfy the reviewed-to-published transition. Neither does clicking a publish button. The public page is a separate result to verify.
Keep a transition log
An event record can contain article_id, previous_state, next_state, actor, timestamp and reason. Preserve the reason when a draft moves backward. If review sends an article back because an availability claim lacks support, the next research brief should include that exact question.
For a hypothetical item called article-017, the log might say: drafted to researched; reason: public access not established; owner: editor. That record tells the next step what is missing. A generic “try again” instruction does not.
Handle failure deliberately
Keep publication failure separate from content rejection. A site timeout does not establish that the text is unsuitable. Check the public result before attempting a second submission, so an uncertain response does not create duplicate posts. If the site imposes a limit, keep the item reviewed and record the blocker.
Start with a table or plain text file. Add machinery only when the actual queue needs it. A state model clarifies what each label means, but it cannot establish that the underlying source or editorial decision is correct.
Disclosure: This article and its illustrative cover were generated with AI. The state model is an example, not a tested implementation.
Top comments (0)