What each step reads
Five labels say how each input to an AI step is filtered, and why some steps see more than what you approved.
Every AI step in CareerBot gets a fixed set of inputs. Each input carries one of five labels, and the label decides what gets through. The labels are the reason a resume can't cite a fact you never approved, and the reason a step doesn't keep proposing what you already turned down.
- Approved onlyOnly what you approved.
- Your wordsYour stories and what you write to it, exactly as you wrote them.
- Not yet reviewedIncludes proposals you haven’t reviewed yet, passed so they aren’t repeated, never as facts.
- Rejected, to skipThings you turned down, passed only so they aren’t proposed again.
- PublicPublic information about a company or a role: its website, its job board, its postings.
Approved only
Only what you approved, and only while it still stands. A fact counts while at least one of its sources stands: if you reject or delete the story a fact came from, and it has no other source, it stops counting everywhere at once. A project's facts count only while the project itself is approved.
This is the only input that writing ever builds on. Resumes, cover letters, answers, outreach and follow-ups read your approved roles, facts, projects, context, insights, skills and directions, and nothing proposed or rejected. When a reply cites anything else, the citation is dropped before it's saved.
Your words
Your stories, exactly as you wrote them. Reading steps need them as written, because a fact has to quote the story it came from. When Read a story proposes a fact whose quote can't be found in the story being read, the fact is flagged for you to look at.
Most steps also get your other stories, but for background only: they help the AI understand a passage, never as a source. Your goals story is never passed as background, and neither is a story you rejected.
Read a story
You choose Read on a story.
Reads
- Your wordsThe story being read
- Approved onlyYour roles and facts
- Not yet reviewedRoles and facts already waiting in Review, so it doesn’t repeat them
- Your wordsYour other stories, for background only
- Approved onlyYour directions and limits
- Rejected, to skipFacts you rejected, with your reasons
May
Propose roles, facts and context, each with the words it came from.
May not
Take a fact from any story but this one, or invent a customer, number, result or skill.
Two other inputs count as your words:
- A GitHub repository. Read a project reads the repository as it is, the way a story is read: it's the evidence, so each fact must cite the files it rests on, and a fact citing no file that was read is flagged. Secret-looking files, dependencies, build output, lockfiles and binaries are never fetched, and anything that looks like a password, key or token inside what is read is replaced with [redacted] before it goes to the AI.
- What you write when you ask about a role. Ask about this role reads your messages as the question and its context, never as a claim about you. If you tell it something new about yourself, that message becomes a note in your own words, read like any story, and what it holds arrives in Review as proposals.
Not yet reviewed
Includes proposals you haven't reviewed yet. Steps that get them are told they're waiting for review: not true, and not to build on. They're passed for three reasons, each a deliberate choice:
- So nothing is proposed twice. Read a story and Read a project get what is already waiting, so reading a second story doesn't fill Review with the same fact again.
- Because duplicates are usually new. Find duplicates reads facts waiting for review alongside approved ones, since a duplicate is most often a new proposal repeating an approved fact. Its only output is a flag you resolve.
- So waiting items can be tidied. Gather skills reads skills, tools and certifications waiting for review, so it can rename, regroup or offer them as merges instead of proposing a second copy.
Read the changes gets every fact the story is a source of, approved and waiting, because those are the facts a change to the story can affect. It may suggest updates only to those facts.
Set-aside items are never sent.
Rejected, to skip
Things you rejected, with the reasons you gave, passed only so they aren't proposed again. Reading steps also skip, in code, a proposal that repeats a rejected fact in the same words. Gather skills never brings back a rejected skill in any spelling.
You can look past your rejections once: Read again on a story or project has Including what I rejected, which leaves out that source's own rejections for that read only.
One step reads a rejected fact for a different reason. When you reject a fact that a resume or letter line cites, Update lines reads its words as what to take out of the line, never as a source.
Public
Public information about a company or a role: a company's name and website, its public pages and job board, a job posting. Screen companies reads only a company's name and website. Fill in companies reads its site and public job board. Rank roles reads the full posting.
Inputs with no label
A few steps read your own settings or a single thing you wrote, which aren't filtered because they're not record content:
- Filters from wording reads one limit's wording and nothing else, never your goals story.
- Writing steps read the length you set for a resume, and your name from your profile to sign letters and messages.
- Check against facts reads the line you put in your own words, beside the approved facts it rests on.