נתתם לקלוד את קורות החיים שלכם.
עכשיו תנו לו לעבוד בשבילכם.

I use career-ops. Build me one daily scan command on top of the existing scanner (scan.mjs + portals.yml). Show me the design before writing code.

Requirements:
1. One entry point: node daily-scan.mjs, with --dry-run that writes nothing under data/.
2. Sources: the existing portals.yml ATS scan, plus these extra sources: [list the job boards / searches you want]. Every source is read-only and reports complete / partial / failed. One failing source never stops the others.
3. One writer: only a final triage step may append to data/pipeline.md and data/scan-history.tsv, under a file lock, with a backup and an atomic write.
4. Triage against data/applications.md: drop duplicates (same URL, or same company + role), drop companies where I'm in an active process, and respect a cooldown after a rejection ([60] days).
5. Print one short report and also save it to output/daily-scan/.
6. My settings go in a config file in the user layer. Put new code in new files; avoid editing upstream files like scan.mjs, and tell me if you must. Add tests with fixtures that never touch my live data.

Add a read-only WhatsApp source to my career-ops daily scan, for job-referral groups I am a member of. Show me the design before writing code.

Hard rules:
1. Read only. Never send, react, forward, or click links inside WhatsApp.
2. Only the groups I list by exact name in a private config file in the user layer. Ignore every other chat.
3. Use the Claude in Chrome browser tools in my own logged-in WhatsApp Web tab. No separate login, no stored credentials.
4. Read only messages since the last complete read (first run: last 24 hours).
5. Extract job leads only: company, title, location, link, post time. Never store message authors, phone numbers, or other members' text.
6. If the read is incomplete (not logged in, a group missing, history didn't load back to the window start), exclude WhatsApp from this run entirely, say so at the top of the report, and don't advance the "last read" marker.
7. A lead with a hidden company or no link is never guessed: put it in a review list (data/unresolved-leads.tsv).
8. Leads go through the same triage as every other source (duplicates, cooldown, filters).

I explicitly authorize career-ops to submit applications for me, in the pipeline workflow only. I understand upstream career-ops never submits on its own and that this is my choice and my responsibility. Write these rules to modes/_custom.md only, which pipeline mode already loads. Do not edit modes/pipeline.md, modes/_shared.md or any other system file, because updates overwrite them. Show me the plan first.

Rules:
1. Start in REHEARSAL mode: fill every form and STOP before Submit, until I write "rehearsal off".
2. Only jobs already in my pipeline that score at or above my apply threshold, best score first. Daily cap: [10]. Never scan for more jobs to hit the cap.
3. Right before each submit, re-check data/applications.md for this company. Skip if I already applied to this role, I'm in an active process there, or it's inside the cooldown.
4. Attach only the CV generated for this posting. Never reuse another one.
5. Answer only from cv.md, config/profile.yml and modes/_profile.md. If a required answer isn't there, don't guess: hand the job back to me.
6. STOP and hand back (never bypass) on: CAPTCHA, account creation or password, consent / EEO / self-identification boxes, e-signatures, or any judgment call. Respect each site's terms; if a site disallows automated applications, skip it.
7. Count a submission only after a confirmation page appears. Then run node set-status.mjs --report <report#> Applied.
8. A blocked job is logged with its exact reason and the run continues with the next one.
9. End with a report: submitted N/[10], blocked jobs and reasons, what's still pending.

Create a project skill at .claude/skills/reconcile/SKILL.md that reconciles my Gmail with data/applications.md. Read-only on Gmail.

1. Read the date in data/last-reconciliation.txt (if the file is missing, this is the first run: use the last 14 days). Search Gmail through the Gmail connector from one day BEFORE that date (overlap window). If the connector isn't available, stop and tell me.
2. Also read sent mail for the same window, to confirm which applications and replies actually went out.
3. Classify each relevant email: rejection / recruiter call / screening step / new posting.
4. Update matching tracker rows with date and source, using node set-status.mjs. Statuses only move forward (Evaluated → Applied → Responded → Interview → Offer). Rejected / Discarded / SKIP are terminal: if new evidence seems to reopen one, report it to me and don't change it.
5. NEVER send, reply, or create drafts.
6. Output a table of changes, then calls and interviews in the next 48 hours.
7. On success, write today's date to data/last-reconciliation.txt.

Change how my evaluations treat hard requirements. Write this to modes/_profile.md (user layer), not to system files:

1. In every evaluation, add a table "Hard requirement mismatches": each requirement the JD states as mandatory that I don't meet (degree, years, language, clearance, location, work authorization), with the JD's wording.
2. Add one line: "Confidence: High / Moderate / Low — N hard requirement mismatch(es)", based only on that table.
3. Never lower the score because of a mismatch, and never auto-skip a role because of one. A strong role with one mismatch stays a strong role with lower confidence.
4. A degree requirement I don't meet always goes in the table and lowers Confidence, never the Score.
5. Add the confidence to the report's machine summary too (only the report carries it; scripts and the dashboard don't read it).

Calibrate career-ops for the Israeli job market. Edit user-layer files only (portals.yml, modes/_profile.md, config/profile.yml), never system files.

1. Add these Israeli companies and their careers pages to portals.yml: [company list]. Then run node verify-portals.mjs and fix or disable entries that don't answer.
2. Location policy: I can commute to [cities / area]; hybrid or remote is fine for [...]; score anything else lower.
3. Comp targets: [range] ILS per month gross (mention Keren Hishtalmut and other benefits when relevant).
4. Framing rules for my CV and answers: how to describe my military service: [your rule]; titles I may and may not use: [...]; for Israel-based roles, don't sell timezone or visa.
5. Many Israeli JDs list a degree as boilerplate: treat it as a confidence note, never as an automatic skip.
Show me the diff before saving.

Harden my career-ops tracker. Put house rules in modes/_custom.md. Prefer new scripts over editing upstream ones (set-status.mjs, update-system.mjs); if an upstream change is unavoidable, show me the diff and add a test first.

1. Statuses only move forward (Evaluated → Applied → Responded → Interview → Offer). Rejected / Discarded / SKIP are terminal.
2. Add a check I can run after any batch: compare data/applications.md with a snapshot of the tracker (career-ops keeps no tracker backup, so save a snapshot at the end of each run) and list every row whose status moved backwards.
3. Re-apply cooldown: [60] days after a rejection with no human contact, [90] days after a rejection that followed a real conversation. Scan and apply steps must respect it.
4. Before any application, re-check the tracker for the same company + role and stop if it's already there.