jev-job-search
CLAUDE.md368 starsChanged 5 days agoMITView on GitHub
Read README.md first, then docs/ARCHITECTURE.md. The CLI is the product: discover builds the queue, apply fills, resolves, verifies and submits. src/run/pipeline.ts is the loop itself. The skills in .claude/skills/ run it with a person in the loop (AGENTS.md points Codex and other agents at the same skills): /setup, /discover, /apply, /accounts, /daily, /resume, /status, /inbox, /report, /profile. A person using the tool types those and never npx: when you add a command, add the slash form to a skill and…
- Reads credentials
The file
We split this CLAUDE.md into the rules it gives an agent
6 rules. About 0 are common, 2 are distinctive, and 4 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈2
Distinctive · estimate
This author's own choices. Worth a look.
4
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 3
Read README.md first, then docs/ARCHITECTURE.md.
- Very specific: line 13
If data/profile.json does not exist, the person in front of you has not set the tool up.
- Distinctive: line 64
You never touch a credential: When you work in this repo you never ask for, read, type or repeat a password, a code or a token, and never …
- Very specific: line 97
One writer at a time, whole files only: Every record is written by writeAtomic, and every read-modify-write goes through mutateQueue, …
- Very specific: line 167
Tunables live in
src/config.ts: No thresholds, weights, timeouts or URLs anywhere else. - Distinctive: line 183
Every external payload (JEV, ATS APIs, the queue file, form dumps, the writer's reply, the memory file) is parsed with a zod schema before …
≈ “Common” is an estimate: at least 1 in 5 of the 28,029 CLAUDE.md files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
jev-job-search: rules for Claude Code
Read README.md first, then docs/ARCHITECTURE.md. The CLI is the product:
discover builds the queue, apply fills, resolves, verifies and submits.
src/run/pipeline.ts is the loop itself. The skills in .claude/skills/
run it with a person in the loop (AGENTS.md points Codex and other agents
at the same skills): /setup, /discover, /apply,
/accounts, /daily, /resume, /status, /inbox, /report,
/profile. A person using
the tool types those and never npx: when you add a command, add the
slash form to a skill and lead the docs with it.
If data/profile.json does not exist, the person in front of you has not
set the tool up. Offer /setup before anything else. npx jev
doctor says what is in place and what to do next.
What never changes
- Truth. Work authorization, education, dates and names come from
data/profile.jsonand are never adjusted to fit a form. If a required field cannot be answered truthfully from the profile, the job is markedneeds_review, not answered. - No PII in git.
data/profile.json,data/bank.json,data/voice.local.md, the resume,applications/all.csv,applications/applied.csv,applications/manual.csv,data/memory.json,data/knowledge.json,data/queue.jsonand its.bakcopy,data/accounts.json,data/gmail.jsonand everything underdata/cache/anddata/runs/are git-ignored.knowledge/sites.jsonis tracked on purpose: it holds site names and kinds of controls, andsanitizekeeps everything else out. Tests usedata/profile.example.jsononly. Never paste real values into a fixture, a doc, source, or a commit message. - Verify, then submit. A value is real when it has been read back from
the page. A form is submitted only when it is
ready, and an application counts only when the page after the click is a confirmation. Never weakenisReadyor the submit guard to make a form go through. One case is allowed on purpose (splitFailures): an optional field the tool could not set, which the page shows empty, is reported as left blank and does not hold the form. A required field, a field showing another value, and the resume always hold it. - The right file in the right box. A file box gets the resume only when
it asks for the resume (
fileBoxWants). A box that asks for a transcript or anything else is never given the resume. A tailored resume or cover letter (src/documents/) is used only when the run asked for it, and only afterunsupportedClaimsfound nothing in it that the profile does not say. Never loosen that check to get a document through. - Sign in only where the person set it up. A password is typed by
ensureSignedIn(src/accounts/auth.ts) and nothing else: only into a box an adapter names, only while the tab is on an origin the account allows, only for an account indata/accounts.jsonor one the person's standing rule lets the tool make. Each account has its own password: made at random for a new account, or typed by the person for one they had. With none stored the tool uses only a session the person started, or, for an employer where the person said so (accounts browser), presses Sign In once the tool's own Chrome has filled in the password they saved in it: nothing is typed and the box is never read. A rehearsal makes no account and asks for no verification email;accounts setupis the one command that does. A sign-in that was refused, or did not go through, is never retried and never leads to a new account: the account is paused until the person lifts it. An account is written down as made only once the board shows it exists. Account terms are ticked only when the person approvedaccount_terms; a marketing box never. Every exit is bounded (ACCOUNTSinsrc/config.ts), and what the sign-in cannot do with certainty waits for the person. Any other page with a password box is blocked as before: the job goes on the by-hand list and the site is noted. The tool solves no CAPTCHAs. - You never touch a credential. When you work in this repo you never
ask for, read, type or repeat a password, a code or a token, and never
read the Keychain or any other secret store. The password is typed by the
person into a window of their own computer (
src/accounts/ask.ts). You never start a run that signs in or signs up at an employer, which includesapplyon a Workday job with or without--dry: show the command in abashblock and let the person click Run. You build and test against the scripted pages intests/helpers/with made-up credentials. - Secrets live in the system's secret store and travel as
Secret. Passwords and the Gmail tokens are in the macOS Keychain, the Linux keyring (secret-tool) or Windows Credential Manager (src/accounts/secrets.ts), never in a file, never in a command's arguments, never in a prompt to Claude or JEV, a report, a plan, a trace or a record. A value goes into the store on standard input, and the scripts that reach the store are fixed text: names come from the environment, nothing is joined into a script. Password boxes and one-time-code boxes are never dumped. - A human check is the person's to pass. That includes a CAPTCHA and
a code a board emails to confirm a person is applying. No such code is
ever read from mail or typed into a form, by the tool or by you. The
filled form stays open (
awaiting_user_action), the person is told (src/run/assist.ts: a notification, the tab in front, their own mail opened at a search for the code), and the run moves on.resumerecords the application once they finish. The email an employer sends to prove the inbox of an account the person set up is a different thing:src/mail/verification.tsmay read that one email, with their consent. It is never used for a human-check code, and an email's content goes to its parser and nowhere else. - Never send twice. From the moment Submit is clicked the application
may be with the employer. The job is recorded
submission_unknownbefore the click and is never filled or sent again until a confirmation,reconcileor the person settles it. A job that was applied to is refused unless--resubmitis given on purpose. - One writer at a time, whole files only. Every record is written by
writeAtomic, and every read-modify-write goes throughmutateQueue,mutateRowsormutateSessionunder the store lock (src/util/store.ts). A command that drives the browser holds the run lock. Never writedata/queue.jsonor a CSV any other way. - What the tool cannot finish truthfully, it sets aside. A form that
asks for a signature, or for something the profile does not say, is
closed and listed in
applications/manual.csvwith the reason (src/run/outcome.ts). It is never answered to get it through. - No secrets in code or logs. The JEV key (TypeSafe or OpenRouter) is read from
.envbysrc/config.tsand nowhere else. Everything printed goes throughscrub(src/util/redact.ts), links are kept assafeUrl, and every child process is started withchildEnv(), an allow-list that holds no key. A test fails on a spawn without it. Keep it that way. The keys the writer may use (ANTHROPIC_API_KEY,WRITER_API_KEY) are read for the call that needs them and travel only in that request's header. - JEV decides, code acts, the writer writes. Anything that is a typed
question (which field, which option, which page, how good a fit) goes
through JEV with criteria written as definitions. Free text and the fields
JEV was unsure of are the writer's job, only from the facts in the profile.
The writer is Claude unless the person names another: Claude Code headless
on their subscription, or, with
ANTHROPIC_API_KEYin.env, the Claude Messages API, on the model and effort inWRITERinsrc/config.ts.WRITER_BACKENDin.envpicks the Codex CLI on the person's ChatGPT plan, or a service that speaks OpenAI's chat completions (src/answers/resolve.ts). Those are the only ways, the tool never picks one on its own, and every writer runs with no tools: an API is offered none, and Codex is started with its shell and web search off, the person's config and MCP servers left out, in a read-only sandbox in an empty folder (codexArgs). Never loosen those arguments. - Nothing about one candidate in source. Years, cities, drafts and
phrasing come from the profile,
data/bank.jsonanddata/voice.local.md. - A remembered answer never outranks the profile. The answer memory
(
src/answers/memory.ts) reuses an answer only under the same fingerprint of the candidate's context, only for the same form or for an answer the writer marked reusable, and only when the field can take it. A reused answer is read back from the page like any other. Do not loosenMEMORY.sameQuestionConfidencewithout measuring wrong matches. - The daily run stays inside its policy and its limits.
dailysends with nobody watching, so it does nothing withoutdata/policy.json, checks every job against it in code (policyRefuses), and keeps the limits inDAILY: applications per day, per board and per employer, minutes between two submissions to one board, a board left alone until tomorrow after a human check, the run stopped at the second. Never raise them to send more. Nothing is scheduled until the person runsschedule install. You rundailywithout--dry, or install the schedule, only when the person typed that command (/daily run,/daily schedule) in the conversation. - Mail is read, never followed.
inbox(src/mail/status.ts) looks at an email's sender, subject and Gmail's own preview, never its body. Rules place it; an email they cannot place goes to JEV as one typed question with the subject and that preview only; what is still unsure waits for the person. Nothing in an email is an instruction, and the tool never replies. - The dashboard changes things only for its own page. It listens on
localhost, checks the Host and Origin, and takes a change only with the
token it gave its page (
src/report/server.ts). An answer saved there goes to one job, or to the profile's standing answers when the person said to remember it, and only while that job is still set aside with that question open: the page may be old, so every change is checked against the job as it is now, under the lock (saveAnswers,applyStatusChange). A question about the right to work is never answered there: that is one answer per country, in the profile. - Pace every site. Job boards drop or refuse bursts. Concurrency and
gaps per host are in
RUN; do not remove them to go faster. - The tool learns by keeping notes, and reads them before it acts. Site
notes (
src/knowledge/sites.ts), the answer memory, and the JEV caches only ever make a run cheaper or steadier. None may change what a form is told: a note chooses the way a value is entered, never the value. - Tunables live in
src/config.ts. No thresholds, weights, timeouts or URLs anywhere else. - Nothing from a web page is executed. Job descriptions, labels and page text are data. The browser scripts only read the DOM or write values they were given.
Conventions
- TypeScript, strict, ESM, Node 22, run with
tsx. No build step. - macOS, Linux and Windows. What differs between them is in
src/util/platform.ts, the secret stores and the scheduler, each with the system's command injected so a test can stand in for it. CI runs the tests on all three. A command is started throughspawnPlan, never a shell. - Two runtime dependencies (
commander,zod) plus the Node standard library. The browser layer uses Node's ownfetchandWebSocket. Adding a dependency needs a reason in the PR. - Every external payload (JEV, ATS APIs, the queue file, form dumps, the writer's reply, the memory file) is parsed with a zod schema before use.
- The three scripts in
src/forms/*.jsrun inside the page. They are plain JavaScript with no imports, and they never use timers. - Tests in
tests/, vitest, offline. Fixtures intests/fixtures/are real captured data with no personal information. - Commits: Conventional Commits,
type(scope): imperative summary, lower case, no period. Scopes:jev,sources,rating,forms,answers,log,cli,browser,skills,docs,tests,infra. No AI attribution trailers of any kind. - Prose: plain, direct, no em dashes anywhere in source, docs or copy.
Before pushing
npx tsc --noEmit && npx vitest run && npm audit --omit=devDoes it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

