Side A
Four ways AI fails.
Four fixes.
Each rule in Liner Notes answers one of these.
-
A1
Making Things Up
feat. the Evidence Ledger
The AI wants your resume to win. So it writes
ran A/B tests
when you never did, and it sounds great.A1 · The fix
No source, no sentence.
Every claim needs proof: a document, a date, or your own words. Tailoring can change emphasis and order, never claims.
Any phrase you've ever cut gets checked again before every export. A fixed mistake stays fixed.
-
A2
Agreeing With Myself
feat. Rate It First
Show the AI your strengths, then ask what the job needs. Surprise: it needs your strengths. Your gaps become "minor."
A2 · The fix
Rate the job before the evidence.
Every requirement gets its weight from the posting alone, before the AI sees your experience. Then it can't move the goalposts.
Every big gap gets the interview question it will trigger, and your honest answer.
-
A3
Moving Too Fast
feat. the Three Stops
You said "submit." Then a new page popped up with a box you'd never seen. The AI ticked it anyway, because you said go.
A3 · The fix
It stops. You click.
Go or no go, approve the resume, and Submit: three stops, every time. A new page after "submit" means it asks again.
Some companies ask people not to use AI on applications. It checks for that first, and stops.
-
A4
Forgetting Everything
feat. Memory With Reasons
You told it on Monday. By Thursday it's a new session, and it makes the same mistake with total confidence.
A4 · The fix
Keep the rule and the reason.
Every lesson is saved with why it exists and how to apply it. A rule without its reason gets broken the first time it's inconvenient.
Side B
Run it yourself.
The job-search kit, in the order you'd play it. Seven tracks, each with its prompt and a sample run.
What you need
- Claude Code, on a Claude plan
- Claude in Chrome, for the forms
- One folder for your search
- Your documents: resume, LinkedIn as a PDF, offer letters, old job descriptions, and certificates
The house rules
CLAUDE.md · read at the start of every session
# House rules 1. No source, no sentence. Every claim about me traces to a row in CLAIMS.md: a file, a link, a date, or my own words. Tailoring changes emphasis and order, never claims. Anything I cut stays cut. 2. Rate the job first. Weigh each requirement from the posting alone, before you read my evidence. 3. Three stops, every time: go or no go after the fit read, my approval of the resume, and I click Submit. You never click Submit. If a new page appears after I say "submit," ask again. 4. Check every posting and every form for an AI-use policy. If a company asks people not to use AI, stop and tell me. 5. Never guess. A blank in answers.md stays blank until I fill it. 6. Nothing sends without me. Draft messages; never send them. 7. Keep the rule and the reason. When I correct you, add the rule to this file with why it exists.
-
Everything you've actually done, each line with its proof. It reads your documents first, shows you where they disagree, then asks about the rest. Everything after this can only use what's here.
PromptBuild my evidence file. Read every document in my sources folder, and show me where they disagree before you ask anything else. Then make CLAIMS.md: one row per claim I make, with its proof next to it (a file, a link, a date, or my own words). A job description shows what the role was, not what I did, so ask me before you claim anything from one. If a claim has no proof, don't guess. Ask me about it, one question at a time, and cut anything I can't back up.
Have readyA sources folder: your resume, LinkedIn as a PDF, offer letters, old job descriptions, certificates, and reviews.
-
The questions every application asks, answered once, plus the ones your field adds, like references and past supervisors. Anything blank means it stops and asks you, never guesses.
PromptMake my answers file. Ask me, one at a time, the questions applications always ask: where I'll work, work authorization, sponsorship, start date, salary range, and my links. Then add the ones forms in my field ask, like references and past supervisors. Save them in answers.md. If I skip one, leave it blank and ask me each time it comes up. Never fill a blank with a guess.
Have readyYour salary range and where you're willing to work.
-
A daily digest of new openings, so you stop refreshing job boards. Claude suggests the boards that fit your field, you pick, and you can add your own or drop one any time.
PromptSet up my daily job digest from digest/watchlist.md. I'm looking for [titles], in [where you want to work]. Show me the boards in the watchlist, suggest which to switch on for my field and why, and wait for my pick. Add these companies: [companies], and find each one's job board. Then write a small script that reads the watchlist fresh every morning at 9 and lists only new openings that match. When I tell you to add or drop a company, a board, or a title, change the watchlist.
Have readyThe job titles you want, and ten companies you'd be glad to work for.
-
A fit read, a tailored resume, and the form filled in, with three stops.
PromptHere's a job posting: [link]. Check it for an AI-use policy first. Then weigh each requirement from the posting alone, before you read CLAIMS.md. Tell me: apply or skip, my two biggest gaps, and the interview question each gap will trigger. Then stop and wait for my go.
Have readyOne posting link.
You've already heard this one.
It's the demo at the top of the page: same run, same three stops.
-
Every application in one list, plus a watch on each posting. When one comes down, you'll know. That can be days before any email.
PromptAdd this application to applications.md: company, role, date, link, and status. Every morning, check whether each posting I've applied to is still up. When one comes down, tell me and mark it likely closed.
Have readyNothing. Track 04 adds the row.
-
Make something small for that one team, then ask one person for their input on it. Never for a call.
PromptI'd love to work at [company]. Research their product and the team I'd join, and find one problem I could actually help with. Help me make one small thing about it in an evening. Then draft a short note to one person there asking for their input on it. Don't ask for a call, and don't send anything.
Have readyA role you'd really want, and an evening.
-
One page per company and per round, built from what you sent and what's true.
PromptI have a [round] interview at [company]. Make a one-page prep sheet from what I actually sent them and CLAIMS.md: the questions they'll likely ask, my honest answer to each, and three questions I should ask them. Don't put anything in my answers that isn't in my evidence.
Have readyThe interview invite.
Where it all lands
One page for you.
Plain files underneath.
Once your documents are in, Claude builds you a private page on your computer. It asks for what's missing, keeps your answers, and shows where everything stands. Underneath, it's one folder of plain files you can open and fix yourself.
FAQ
Fair questions.
The short ones are here. Anything else gets answered from the notes on this page, or not at all.
-
No. It reads the posting, tailors your resume, and fills in the form, then it stops. You make three calls: go or no go, approve the resume, and Submit. It never clicks Submit.
-
It's built not to. Every claim needs a row in your evidence file with its proof: a file, a link, a date, or your own words. Phrases you've cut get checked again before every export, and you approve the resume before it goes anywhere.
-
Everything that shows your work: your resume, your LinkedIn saved as a PDF, offer letters, old job descriptions, certificates, and reviews. It reads all of them before it asks you anything, and it starts by showing you where they disagree.
-
No. Claude builds you a Career Record: a private page on your computer that asks for what's missing, keeps your answers, and shows where every application stands. The files stay underneath.
-
They show what the role was, not what you did in it. So they help it ask better questions, and a claim still needs your yes.
-
It checks every posting and every form for an AI-use policy first. If a company asks people not to use AI, it stops and tells you.
-
The prompts and house rules are plain English, so they'll paste into any AI assistant. The rest, like filling in forms and the Career Record, is built and tested on Claude only.
-
No. You talk to it in plain English, and every prompt on this page is ready to paste. Setting up Claude Code happens once.
-
The kit is free: download it from GitHub, or copy the prompts and house rules from this page. Claude Code needs a Claude plan.
-
In one folder on your computer. Claude reads them while it works, the same as anything else you share with it, and nothing goes to an employer until you click Submit.
-
Yes. The steps are the same for any job. Forms ask different things by field (a city job may want references and past supervisors up front), so your answers file covers those too. The daily digest watches the companies and boards you pick, so add your city or state jobs board to your watchlist.
Deep cuts
Five more skills.
Free, for any project.
Liner Notes runs on skills like these. These five are public.
claims-ledgerKeeps every claim in a resume, portfolio, or bio traceable to evidence.
handoff-specTurns a designed screen into a handoff developers and QA can build from without asking.
safe-bulk-actionDesigns bulk deletes and cleanups so nobody loses what they meant to keep.
a11y-auditRuns a WCAG audit on a page, a prototype, or a design, in the same format every time.
pegboardKeeps a library of tools and sources, and checks it before picking one.
Credits
Who played what.
- Daniel Glaze
- Design, direction, and every rule.
- Claude Code
- Typing, form-filling, and being told "no."
- Review agents
- Three of them. They report. They never touch the work.
- You
- Every final call.
Want to run it?
It's free.