A job-search system you run with an AI assistant in your terminal. Local transcription of your interviews, resumes tailored per posting, a funnel that doesn't flatter you, and a method for working out why you're actually losing.
Built during a real job search after a layoff. Every opinionated thing in here is opinionated because something went wrong first.
There is no app and no sign-up. It's markdown, YAML and a few Python scripts. An AI coding assistant reads the files and runs the commands — the assistant is the interface.
Two ways in. Pick the one that matches what you have.
Everything runs on your own GPU. No account, no API key, no monthly fee. You get a chat window in your browser that interviews you about your career and writes the document the rest of the system builds from.
You need: Windows 10 or 11, an NVIDIA card (8GB or better), about 25GB free, and an hour where you can talk about your work history.
Install Python 3.12 first — not 3.13. The chat window doesn't support 3.13 yet, and 3.13 is what python.org's big green button gives you. Get it here: python.org/downloads/release/python-3128 — scroll to the bottom, take Windows installer (64-bit), and on the very first screen tick "Add python.exe to PATH".
Then download this repository (green Code button → Download ZIP), unzip it anywhere, and double-click these in order:
| Double-click | What happens | |
|---|---|---|
| 1 | Setup Adapter.bat | Installs everything. 20–40 minutes, mostly downloading. Lots of text scrolls past — that's normal. |
| 2 | Start Adapter.bat | Opens your chat window in a browser. Leave the black window open. |
| 3 | (in the browser) | Make a local account — it exists only on this PC. Then paste in the system prompt, as local/README.md shows. |
| 4 | (in the browser) | Say hello and talk about your career. Stop and come back whenever you like. |
| 5 | (in the browser) | Download the chat: click its name at the top → Download → Plain text. |
| 6 | Build My Profile.bat | Turns the conversation into source/master_profile.md, then prints a list of things to check. |
| 7 | — | Read that list and fix the file. This step is not optional; see below. |
If anything at all goes wrong — double-click Share My Log.bat.
It writes one plain-text file to your Desktop with your Windows version, graphics card, Python versions, which packages installed, and every error the scripts hit. Send that file. It contains nothing from your interview and nothing from your profile — the scripts never write your career history to a log. Open it and read it first if you want to; it's plain text on purpose.
You can run Share My Log.bat even if setup failed immediately — it checks the
machine fresh rather than relying on a log that was never written.
A model small enough to run on a 12GB card is noticeably weaker than a frontier one. In testing it kept every real story and invented nothing of substance — but it turned "we lost 14,000 gallons" into "saved 14,000 gallons", and it dropped a "don't claim I know Python" correction entirely. Both now get caught and printed for you to check.
So
Build My Profile.batends with a checklist: every job title it wrote, every number set beside the sentence you actually said it in, and everything you told it you had not done. Work through it. The whole point of this system is not claiming things you can't defend in a room.
git clone <your-fork> adapter && cd adapter
./setup.sh
claudeThen say: "I'm new here. Walk me through setting this up."
setup.sh detects your hardware — Apple Silicon, NVIDIA, or CPU — and
installs the matching transcription backend. The assistant takes it from
there: it runs the career interview, works out which job titles your record
actually supports, and builds your first tailored resume before you've typed a
command.
Not using Claude Code? Paste BOOTSTRAP.md into whatever assistant you do use. Full instructions in SETUP.md.
No resume and no idea what to apply for is the expected starting point, not a problem to solve first. The interview is the source document; everything generates from it.
The job market is getting harder, and the parts of it that are getting harder fastest are the parts where a person used to read your application.
Most job-search tooling optimises for volume — more applications, faster. That is the wrong variable. This optimises for knowing what's actually happening: which framing gets you screened out, which channels produce real conversations, what was genuinely said in the room versus what you remember.
1. Record and transcribe your interviews, locally.
Memory reshapes a conversation within hours. It inverts numbers, hardens maybes into commitments, and quietly rewrites what you were told. Real cases from this system's own use:
- A hiring manager said a licensing arrangement saved about a million a year. By the next day it had been remembered as costing a million.
- A posting described the company's strategy as "buy over build." The executive, in person, argued the exact opposite at length.
- A recruiter named the specific question that two previous candidates had failed.
All three would have been lost or inverted inside a day. Transcription runs on-device — nothing is uploaded, which is what makes it defensible for a recording with another person's voice in it.
2. Derive your framing from your own rejections.
Two postings can describe the same job and screen you completely differently, because the title decides which pile you land in. This system's author found a single phrase — "X+ years as a Software Engineer" — behind a run of consecutive rejections, while customer-facing titles at comparable companies advanced to late rounds.
You can measure your own version of that. FRAMING.md is the
method; role_framing.yaml is where it gets scored.
3. A funnel that tells you the truth.
This dashboard reported a hit rate roughly double the truth until three bugs were found: dead processes counted as conversions, an authoring convention misread as "never submitted," and no staleness rule at all. Underneath the corrected number was a second one that mattered more — applications sent cold converted at a small fraction of the rate of anything that arrived through a person.
Seeing that changed the whole strategy. A tracker that flatters you makes you do more of what isn't working.
4. Honesty flags in your profile.
The failure mode of AI-written resumes is confident over-claiming — the
assistant infers a capability from an adjacent fact and writes it as
experience. master_profile.md uses explicit markers ([RESUME-SAFE],
[PERSONAL], [DEMO], [STALE], [DO NOT CLAIM]) so corrections survive,
and so the assistant asks instead of guessing.
Two rules it enforces: silence is not absence — a technology the profile doesn't mention means nothing about whether you know it, so ask. And adoption is not deployment — driving a tool's adoption is not the same as having shipped it.
setup.sh Hardware detection, install, first-run layout. (Mac/Linux)
CLAUDE.md Standing rules the assistant reads every session.
BOOTSTRAP.md Paste-in equivalent for assistants that aren't Claude Code.
Setup Adapter.bat Windows, no terminal. Installs everything.
Start Adapter.bat Windows. Opens the local chat window.
Build My Profile.bat Windows. Interview in, profile document out.
Share My Log.bat Windows. One diagnostic file, no personal content.
local/README.md Click-by-click guide for the fully-local Windows path.
local/SYSTEM_PROMPT.md The interviewer prompt you paste into the chat window.
local/extract_profile.py Builds the profile, then prints what to verify.
local/_log.ps1 Shared logging. Every step, every exit code.
docs/PLAYBOOK.md The method, in the order you do it. Start here.
docs/FRAMING.md How to derive your own title rules from your outcomes.
docs/JD_RECOVERY.md Pulling postings out of Workday, ADP, Greenhouse, Lever, Ashby.
templates/ Copy these into source/ and make them yours.
master_profile.md The hour-long career interview. Everything generates from it.
APPLICATIONS.md The tracker the dashboard parses.
role_framing.yaml Title scoring rules.
job_filters.yaml Freshness, location, comp floor, exclusions.
company_watchlist.yaml Companies to watch directly. Starts empty; you build it.
jd_registry.yaml Joins applications to postings and documents.
skills/ Assistant workflows.
onboard/ Career interview, title triangulation, first resume.
build-resume/ JD in, tailored ATS-safe PDF out.
mock-interview/ Questions from real transcripts; records and grades you.
transcribe/ Recording in, transcript in the conversation.
source/ The scripts.
job_scout.py Multi-source scan, scored against your framing.
pipeline.py Parses APPLICATIONS.md into a dashboard.
md_to_pdf.py Styled, ATS-safe PDF rendering.
transcribe.py Local GPU transcription, plain + timestamped.
mcp_server.py Optional read-only Gmail/Calendar tools.
Warm beats cold by a lot. Only a small fraction of cold applications produced a human conversation at all, while nearly every late-stage process came from a recruiter, a referral or inbound interest. If that holds for you, more applications is the wrong answer and more humans is the right one.
Apply where your longest experience is, not where your most recent interest is. These are often different places, and the market prices the former. It is possible to spend months losing late rounds in a field you've worked in for a couple of years, while a decade of deeper experience sits unused one title away. Pivoting to the titles that matched the longer record changed the response rate immediately.
Silence is an answer. An application with no reply after three weeks is a rejection, not an open question. Counting it as open is how a pipeline looks healthy while nothing is happening.
- Transcription has no speaker labels. A two-person call comes back as one stream. Usually inferable from content; sometimes genuinely ambiguous.
- Proper nouns get mangled. Company and product names especially — check them.
- LinkedIn is manual on purpose. No usable API, active anti-scraping, and the terms-of-service risk lands on your account. You browse and paste. PLAYBOOK.md explains why that's not as bad as it sounds.
- Transcription without a GPU is slow. Apple Silicon and NVIDIA are fast; CPU works but a one-hour interview takes 20–40 minutes.
- Nothing here beats a referral. This makes you faster and better informed. It does not replace knowing someone.
- The fully-local model is the weakest link in that path. At 14B it interviews well and fabricates little, but in testing it reversed the direction of an outcome (a loss written as a saving) and omitted a correction the user had explicitly asked it to record.
Build My Profile.batprints both classes of error for you to check, and you have to actually check them. With Claude Code or another frontier assistant, this isn't a concern. - Windows is the least-tested path. The Mac/Linux route has been used daily for months. The Windows scripts are newer; they log everything to
logs/precisely because they're the ones most likely to surprise you.
The scripts are deliberately small and hackable. If you adapt the framing method to a different field and it works, a note about what your phrase turned out to be would be genuinely interesting.
MIT — see LICENSE.