Changing a generated blueprint manually or by asking the agent, reviewing what changed, and what the version history does and does not record.
A generated blueprint is a first draft, and there are two ways to change it: edit it yourself, or ask the CAB agent to.Both routes end in the same place: a review screen where you accept or discard each change before anything is written.
Share CAB gives a blueprint to people who do not have a BlueGenAI account. A stakeholder, a client, or a colleague on the business side can read the whole document from a link, without signing up and without being added to the project.This is what makes a blueprint usable as a review artifact. A specification is only worth generating if the people who need to agree with it can read it, and most of them are not platform users.
A shared blueprint needs no account and no approval; anyone holding the link can read it. Share it only where you would be content for the whole document to travel.You may turn off the sharing link at any time
Note that this is a different mechanism from inviting someone to the project. An invitation gives a person a role and a sign-in; a share link gives one document to anyone who has the URL. See Collaboration Features.
Edit CAB turns the blueprint into a form. After a few seconds the panel title gains an Editing pill and the toolbar becomes Review with AI, Save Changes, and a close control.
In edit mode, individual values become editable in place. There is no formatting toolbar because there is no formatting.
The editor knows the blueprint’s structure. You are editing fields, not a document. Click any value to change it. Every list ends with a labelled control for adding an item: + Add acceptance criteria, + Add field, + Add role. Hovering an item offers controls to insert a sibling or remove it.
A removed item stays visible, struck through, until you save.
Removals are pending until you save. A removed item stays on screen struck through, so you can see what you are about to lose.
Adding a role, reviewing it, and saving. The single manual edit produces one 'your edit' change and fourteen AI suggestions, and saving takes the blueprint to version 5.
Notice the count. One added role produced fourteen further suggestions. Adding a role means deciding what it can reach, which screens mention it, and which stories belong to it, and the review works those out rather than leaving you to find them. That is the review earning its cost; see Reviewing changes below.
The BlueGenAI CAB Agent tab beside CAB Configuration takes requests in plain language. Its composer reads Ask anything about your CAB, or request changes....
The agent tab opens completely blank: no heading, no examples, nothing describing what it can do. It works, but the product does not tell you what to ask it.
Requests that work:
Ask for
Example
A new item in a named section
Add a business rule requiring a completed W-9 before an application can be submitted
A change to something specific
Change the applicant role so it cannot approve its own submission
A correction across the blueprint
The system is called LEAD, not Lead Grants; fix it wherever it appears
An answer rather than a change
Which screens does the reviewer role reach?
One request can change several sections. A single rule request produced six changes across user stories, the data model, three screens, and the rules section; the agent works out what else the change implies rather than editing only what you named.
Both routes land here. Hand edits go through it when you select Review with AI; agent changes go through it automatically.The review looks at your edits, not only at its own suggestions. It reads what you changed, checks it against the rest of the blueprint, and can propose corrections to your wording or to anything your change has knock-on effects for. Running it is a second opinion on your work as much as a way of getting the agent to do some.
The review screen, part by part. It shows one change at a time.
Each change is labelled by where it came from:
Label
Meaning
YOUR EDIT
You changed it by hand
AI ADDED
The agent added something new
AI SUGGESTED
The agent changed existing content
Modifications show a WAS row and a NOW row; additions show what is being added. Agent changes carry a one-line reason, your own do not. Selecting a change scrolls the blueprint to it and outlines it, so you see it in context.
Reviewing with AI is optional and it costs. You can save hand edits without it, though you then lose the consistency check. When you do run it, the cost is added to the project’s generation cost.
The same review, before and after two decisions. Deciding anything relabels the buttons and starts a tally.
Accept or discard each change with the controls on its card. Nothing is applied until you use the toolbar.
Undecided changes count as accepted. Saving keeps everything you accepted and everything you never looked at. Only changes you explicitly discarded are dropped. On a review with 20 changes, working through the first five and saving keeps the other 15.
Discard All drops everything and writes nothing. It does not create a version.
Saving creates a version. Open the clock button to see them.Each entry is a timestamp: no author, no description, no summary. The current version sits at the top, and selecting an older one previews it without restoring.
Restoring deletes every version newer than the one you restore. This is not an undo stack you can move up and down; going back three versions destroys the two in between, permanently. The confirmation says so, and the confirm button is styled like an ordinary action rather than a destructive one.
The restore confirmation. The second sentence is the one that matters.
This distinction matters if anyone ever asks you who changed what.While a change is under review, attribution is complete. Every proposed change is marked as yours or the agent’s. Each accept or reject is recorded against a named person with their email address and the time they decided. The agent’s suggestions move through a tracked lifecycle from queued to resolved.Once a change is applied, none of that travels with the blueprint. The document has no edited marker, no record of who edited it, and no per-section history. Reading a finished blueprint tells you what it says, not which parts a person rewrote.Version history records when, not who; versions are named by timestamp, and the save operation is not given a user identity.So the review process is auditable; the finished document is not self-describing. If your organisation needs a change record, capture it during review rather than expecting to reconstruct it afterwards.
Only one person can edit a blueprint at a time. The platform holds a lock while a review is running or awaiting a decision, and checks that the version you started from is still current before applying anything.