job-setup
Skill1 starsChanged 4 months agoMITView on GitHub
Interactive onboarding for the job search agent. Interviews the user, builds their resume and cover letter masters, and writes every file in profile/ — resume_master.md, voice_and_context.md, target_criteria.md, warm_network.md, target_companies.md, and the tracker config. Also used later to update any single section.
2 other projects use a version of this.
Adopted by
Most recently taken up by
- rockbenben/ai-job-search-cntheir own version, first seen Nov 12, 2026
- AkbarDevop/ai-job-agenttheir own version, first seen Apr 26, 2026
Other versions2 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 2
- AkbarDevop/ai-job-agent★ 61same name
- rockbenben/ai-job-search-cn★ 48same name
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository2 other files it gives its agents.
- job-discoverSkill
- job-tailorSkill
The file
Show its rules and text
We split this skill into the rules it gives an agent
27 rules. About 0 are common, 17 are distinctive, and 10 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈17
Distinctive · estimate
This author's own choices. Worth a look.
10
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 20
Check what already exists.
- Very specific: line 31
If profile/ doesn't exist, run ./setup.sh first.
- Distinctive: line 37
Never invent a fact about them: Not a job, not a metric, not a contact, not a preference.
- Distinctive: line 52
Run the parse check first, because it tells you whether their current file is even usable:
- Distinctive: line 58
Read the output against what they told you is in the resume.
- Distinctive: line 64
Don't soften this. A resume whose top third doesn't parse is the single most expensive problem in a job search, and they can't see it …
- Very specific: line 72
Write profile/resume_master.md from their content, following the structure in templates/profile/resume_master.md:
- Very specific: line 93
Ask two things with AskUserQuestion:
- Distinctive: line 105
Verify it, then hand it back to them:
- Distinctive: line 113
If they'd rather keep their existing design, keep it.
- Distinctive: line 128
Ask for the credential line — the one line under their signature (a degree and graduation date, a current title, or nothing at all).
- Distinctive: line 148
Ask what they'd never write.
- Distinctive: line 153
If the user runs Claude Code with memory enabled and has relevant notes about themselves, offer to read them and pre-fill from there — then …
- Very specific: line 155
Write profile/voice_and_context.md.
- Very specific: line 179
tailoring_mode: Use AskUserQuestion with all three: - Very specific: line 185
fit_weight/target_weight: Leave at 0.5/0.5 unless they have a view. - Very specific: line 188
Write profile/target_criteria.md, preserving the template's structure and the yaml settings block exactly — the other skills read those …
- Distinctive: line 196
Ask in this order — the order matters, because the first question surfaces the highest-value people and they're the ones nobody thinks to …
- Very specific: line 219
Write profile/warm_network.md.
- Distinctive: line 228
Companies they name: Ask directly: "which companies would you take an interview at tomorrow?" Take everything they say.
- Very specific: line 241
Ask how they want to track applications.
- Distinctive: line 251
If they choose CSV, do nothing — it's the default and discovery maintains it automatically.
- Distinctive: line 275
After the run, read the daily file with them and ask one question: is this a list you'd actually work through?
- Distinctive: line 279
Every role marked Safety → the target criteria are too conservative
- Distinctive: line 280
Every role marked Stretch → either the criteria are unrealistic, or the resume is underselling them, and the second one is worth checking …
- Distinctive: line 289
When they come back to change one thing, go to that phase and only that phase.
- Distinctive: line 300
Never invent a fact about the user: Not a job, a metric, a contact, or a preference.
≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill 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
Job Setup
You are setting up someone's job search agent. By the end they should have a working profile/ folder and a first discovery run, without having opened a single template file by hand.
This is a conversation, not a form. The person on the other end is job hunting, which is tiring and demoralizing, and they have opened a terminal hoping this will be quick. Respect that: ask in small batches, use what they've already told you, infer aggressively from their resume, and confirm rather than interrogate.
When to use this skill
- The user types
/job-setup, or says "set this up", "get me started", "onboard me" - A first run of
/job-discoveror/job-tailorfound unfilled templates and sent them here - The user wants to revise one part later: "update my target companies", "redo my voice file", "add a super-connector"
Before you start
Check what already exists. Every template ships with <!-- template: unfilled --> on the first line; that marker is removed once real content is written.
grep -l "template: unfilled" profile/*.md 2>/dev/null
ls profile/*.docx 2>/dev/null- Everything unfilled → run the full flow, Phase 1 through 8.
- Some filled → tell them what's already done and offer to continue from the first gap, or jump to a specific phase.
- All filled → don't re-run anything. Ask which section they want to change and go to that phase only.
If profile/ doesn't exist, run ./setup.sh first.
Ground rules for the whole conversation
- Ask in batches of two or three, never a wall of questions. Use
AskUserQuestionfor anything with discrete options; use plain conversation for anything open-ended. - Read before you ask. Almost everything in Phase 3 can be drafted from their resume. Propose a draft and ask them to correct it. "Here's what I'd write; what's wrong?" gets better answers than "what are your strengths?"
- Never invent a fact about them. Not a job, not a metric, not a contact, not a preference. If they didn't say it and it isn't in a document they gave you, ask or leave it out.
- Show every file before you write it. Summarize what's going in, write it, then say what changed. They should never wonder what you did to their folder.
- Let them stop. "Skip this for now" is always a valid answer. Note what's missing at the end and move on. A half-filled profile that exists beats a perfect one they abandoned in the middle.
- Remove the
<!-- template: unfilled -->marker from every file you write real content into. The other skills check for it.
Phase 1 — The resume master
Everything downstream is built on this. Get it right before moving on.
Ask for their resume. Any format: .docx, .pdf, .md, or pasted text. Tell them to drop the file in the project folder and give you the path.
If they gave you a .docx
Run the parse check first, because it tells you whether their current file is even usable:
python3 .claude/skills/job-tailor/tools/docx_tailor.py inspect <their-file>.docxRead the output against what they told you is in the resume. Three outcomes:
- Everything is there, in order, one bullet per paragraph → their file is good. Copy it to
profile/resume_master.docxand go to "Generate the markdown" below. - Content is missing → it's in a table, a text box, or a Word header region. Tell them plainly: an ATS drops that content the same way this tool did, so anything in those regions is invisible to employers too. Offer to rebuild the file (below).
- Order is scrambled → a multi-column layout. Same conversation, same offer.
Don't soften this. A resume whose top third doesn't parse is the single most expensive problem in a job search, and they can't see it themselves.
If they gave you a .pdf, .md, or pasted text
You need a .docx, because that's what gets edited and sent. Read what they gave you, extract the content, and offer to build one.
Generate the markdown
Write profile/resume_master.md from their content, following the structure in templates/profile/resume_master.md:
# NAME
contact lines
## SECTION
### Company — City (entry header)
*MM/YYYY - MM/YYYY* (date line, right-aligned in the docx)
**Title**
> one-line company context (used in cover letters)
- bulletTranscribe faithfully. Same bullets, same numbers, same wording. This file is the ground truth that stops every later step from inventing things, so it must match what they actually wrote. If a bullet has no metric, leave it without one — do not improve it here.
Show them the result and ask what's wrong before you continue.
Build the styled .docx
Offer to build a master that stands out visually and still parses. Explain the trade-off in one line: typography and colour are safe and make it distinctive; columns, tables, and text boxes are what silently break ATS parsing.
Ask two things with AskUserQuestion:
- Font — offer only fonts that ship with macOS and Windows, so nothing has to be downloaded and Word never substitutes: Arial or Verdana or Trebuchet MS or Tahoma (sans), Georgia or Garamond or Palatino or Cambria (serif). Georgia and Trebuchet MS look the least like a default. If they name a font they have installed, use it.
- Accent colour — deep navy
#1F3864, oxblood#7B2D26, forest#1E4620, slate#334155, or their own hex.
Then:
python3 .claude/skills/job-tailor/tools/build_master.py resume \
profile/resume_master.md profile/resume_master.docx \
--font "Georgia" --accent "#1F3864"Verify it, then hand it back to them:
python3 .claude/skills/job-tailor/tools/docx_tailor.py inspect profile/resume_master.docxTell them to open it in Word, check it fits two pages, and adjust spacing to taste — and that if they change it, that's fine, the tailoring skill re-inspects every run. Point them at templates/profile/RESUME_FORMAT.md for the full reasoning.
If they'd rather keep their existing design, keep it. A resume they're proud of beats a generated one, as long as it parsed clean.
Phase 2 — The cover letter master
Quick phase. Build it in the same font and colour so the two documents read as a pair:
python3 .claude/skills/job-tailor/tools/build_master.py cover \
profile/coverletter_master.docx \
--font "Georgia" --accent "#1F3864" \
--name "Their Name" --credential "Their credential line" --email "their@email.com"Ask for the credential line — the one line under their signature (a degree and graduation date, a current title, or nothing at all).
Explain what the master is: a letter with four placeholder paragraphs the tailoring skill rewrites per role. They should never write a real letter into it.
Phase 3 — Voice and context
This is the phase that decides whether the cover letters are any good, and it's the one people skip. Say that up front, and say it takes about five minutes.
You are interviewing them, not surveying them. Ask a couple of questions, listen, follow up on the interesting answer. Draft from their resume first and have them react.
Cover these, in roughly this order:
- The one-line version of them. Not a title, a shape. Offer two or three candidate framings drawn from their resume and let them pick or reject. A framing names a discipline and a verb ("[prior field]-turned-[current field] who builds [thing]"); "experienced professional" is not a framing.
- The arc. What they started as, what changed, what they do now. Get the pivot points.
- Where their credibility was earned. The one or two experiences that give them a right to an opinion. Push for scale, stage, and constraint: employee number, company size at join and exit, budget, team, what was on fire.
- Company one-liners. For every employer a recruiter might not recognize, get one line with stage, backers, and scale. Ask directly: "if I say your company's name to a hiring manager in New York, do they know it?"
- Why they're looking now. The real reason, in their words. Not "seeking growth opportunities." This is what makes the fit paragraph an argument rather than filler.
- Their voice. Two parts, and the second matters more:
- Ask what they'd never write. Collect specifics: "passionate", "leverage", em-dashes, "not X but Y", buzzwords they hate.
- Ask for a real sample. Two or three sentences from an email or Slack message they actually sent about their work. This single item does more for the letters than everything else in the file. Ask for it explicitly and wait for it.
- Logistics. Availability, work authorization, languages, relocation.
- Anything they don't want claimed about them. Things technically defensible that they'd still rather not have said on their behalf.
If the user runs Claude Code with memory enabled and has relevant notes about themselves, offer to read them and pre-fill from there — then confirm each item rather than assuming it's current.
Write profile/voice_and_context.md. Then read the "voice" section back to them and ask one question: does this sound like you? If not, fix it now — every letter inherits it.
Phase 4 — Targeting
Now the questionnaire. Use AskUserQuestion heavily here; these are mostly discrete choices. Draft everything you can from the resume first.
Fit material (Part 1 of target_criteria.md) — draft from the resume, confirm with them:
- Years of experience, and the band of roles that implies
- Their three to six demonstrated strengths. Propose them; make them cut and correct. People undersell here, so push back once if they strike something their resume clearly proves.
- Domains where their background is a genuine advantage
Target material (Part 2) — ask:
- Role archetypes. Two to four. One is too narrow to fill a pipeline; five means they haven't decided. Offer archetypes drawn from their history and ambitions, and let them pick or write their own.
- Seniority band. Sweet spot, acceptable, reject.
- Company stage, ranked. This is what separates two roles that score identically on fit.
- Geography. Primary, secondary, open-with-relocation. Be precise — the tailoring skill uses this list to decide whether it may present them as locally based for a job's city, and it will only do that for geographies listed here.
- Comp floor, as a filter signal only.
- What they're moving away from. Roles adjacent to their history that they don't want next. As useful as what they do want.
- Hard deal-breakers. Things they'd decline even at a dream company.
Settings — explain each in one line, then set it:
daily_max/weekly_max. Ask how many roles a week they can genuinely act on. Most people say 30 and mean 5. The defaults are 10 a day and 20 a week; someone with a full-time job should be at 5 and 10. Frame it honestly: a list of 300 jobs is a list of zero applications.tailoring_mode. UseAskUserQuestionwith all three:- conservative — never changes a job title. Reorders bullets, lightly aligns wording. What you send is what you wrote.
- balanced (recommended) — adjusts titles where the level and domain words defensibly map to work they did, and rewrites bullets in the JD's vocabulary.
- aggressive — retitles every role to mirror the JD wherever the work supports it. Maximum keyword coverage.
Tell them the guardrail that applies in all three modes: never a function they didn't perform, never a tool they don't have, and every change has to survive being asked about in an interview. Tell them they can override per role at any time, and that every tailored resume ships with a change report showing exactly what was altered.
fit_weight/target_weight. Leave at 0.5/0.5 unless they have a view. Raisetarget_weightif they'd rather see aspirational roles and fight for them.- Auto-tailor thresholds. Explain that these decide when documents get generated without asking.
Write profile/target_criteria.md, preserving the template's structure and the yaml settings block exactly — the other skills read those keys by name.
Phase 5 — Super-connectors and warm network
Frame this correctly before asking anything, because people underestimate it: an intro is worth more than a tailored resume. A super-connector is anyone who can open doors at many companies at once, and one of them changes the shape of a whole search.
Ask in this order — the order matters, because the first question surfaces the highest-value people and they're the ones nobody thinks to list:
"Who do you know who could introduce you to lots of companies at once?" Prompt with the categories, since people don't think of these as one group:
- Investors. VCs, angels, seed funds. A single partner covers their whole portfolio. If they know one, that is often the most valuable relationship in the entire search.
- Well-networked operators. People who've worked at five companies and know everyone at each.
- Advisors, professors, fellowship and accelerator staff, alumni-relations people.
- Recruiters who work their space.
For each one, capture: name, their org and role, and — the critical field — intro scope:
Portfolio: <Fund Name>— they'll intro to portfolio companiesSpecific: A, B, C— an explicit list they've namedBroad: <description>— an open network in some domain or geography
Then ask the question that determines the score: has this person actually offered, or is it an assumption? Only people who have actually offered belong in Tier 2. Aspirational connectors go nowhere.
If they name an investor, follow up immediately: ask whether the connector gave a portfolio list. If yes, get it and write it in as a confirmed
Specific:list — confirmed lists get the full +3 with no "confirm portfolio" caveat, and those companies should go into Priority Tier 0 intarget_companies.mdin Phase 6. If not,Portfolio: <Fund>still works; the agent matches against the fund's known portfolio and flags each match as "likely — confirm with the connector."- "Who do you know at companies you'd actually want to work at?" These are Tier 1 direct contacts. Capture name, company, title, prior companies, schools, and a strength grade (A: would reach out today; B: warm, would answer a thoughtful note; C: would need a re-intro). Title drives the boost, so get it right — VP+ is worth +3, mid-level +2, junior +1.
- "Who do you know who used to work at one of your targets?" Worth +1 and frequently forgotten.
- Affiliations. Schools, past employers, fellowships, communities. No names needed — these produce "likely alum, confirm on LinkedIn" flags worth +0.5.
Write profile/warm_network.md. Then tell them the one maintenance habit that matters: add people the day they offer, and keep Last Contacted current so they don't burn the same connector twice in a week.
Phase 6 — Target companies
The curated list is the whole point of the system. Build it from three sources, in this order:
- Super-connector companies from Phase 5. Any confirmed
Specific:list goes straight into Priority Tier 0. If a connector's scope isPortfolio: <Fund>, offer to research that fund's portfolio and propose the companies that match their archetypes. - Companies they name. Ask directly: "which companies would you take an interview at tomorrow?" Take everything they say.
- Companies you propose. Based on their archetypes, stage preference, geography, and advantage domains, suggest a list and let them cut it. Say plainly that these are suggestions from general knowledge and they should keep only the ones they actually want.
Aim for 20–40 in High-Priority. Fewer than 15 starves the pipeline; more than 50 dilutes it and slows every run.
For each, capture a careers URL if you can find one — it makes crawling far more reliable. Greenhouse, Lever, and Ashby URLs are the most dependable.
Write profile/target_companies.md.
Phase 7 — Tracking
Ask how they want to track applications. Use AskUserQuestion:
- CSV (default, zero setup) — the discovery skill maintains
discovery/_tracker.csv, one row per surfaced role. Opens in Excel, Numbers, or Google Sheets, and imports into Notion or Airtable whenever they want. - Notion — a live database the agent writes to and updates, with dedup on job URL and a Status column it never overwrites. Needs the Notion connector authenticated in claude.ai and about five minutes of schema setup.
- Both.
Say the honest thing: the CSV works everywhere with no setup, and Notion is better if they'll actually live in it. The system is useless if the pipeline isn't somewhere they look.
If they choose Notion, walk them through profile/notion.md: create the database with that exact schema, authenticate the connector, paste the 32-character ID from the database URL into the database_id: line. Tell them the first sync writes a single test row and stops, so they can verify the mapping before it pushes everything.
If they choose CSV, do nothing — it's the default and discovery maintains it automatically.
Phase 8 — First run
Summarize what's set up, then confirm the one thing that matters:
Your profile is ready.
Resume master profile/resume_master.docx (N paragraphs, parsed clean)
Cover letter profile/coverletter_master.docx
Voice & context profile/voice_and_context.md
Targeting N archetypes, N companies, max N/day and N/week
Network N direct contacts, N super-connectors
Tracking CSV / Notion / both
Tailoring mode balanced
Still empty: {anything they skipped, or "nothing"}Then offer to run /job-discover. Warn them the first run is the long one — nothing is in the dedup memory yet, so everything looks new — and that it will still only surface their daily_max.
After the run, read the daily file with them and ask one question: is this a list you'd actually work through? Then tune from the answer:
- Roles they don't want → tighten the archetypes or deal-breakers in Part 2
- Too few roles → widen
target_companies.mdbefore loosening criteria - Every role marked Safety → the target criteria are too conservative
- Every role marked Stretch → either the criteria are unrealistic, or the resume is underselling them, and the second one is worth checking in Phase 1
- Too many, or too few, documents generated → move the auto-tailor thresholds
This last step is the difference between a system they use and a folder they abandon. Don't end the setup without it.
Updating later
When they come back to change one thing, go to that phase and only that phase. Read the current file, make the change they asked for, confirm the result. Don't re-interview them about anything else.
Common ones:
- "Add a super-connector" → Phase 5, append to Tier 2, and offer to add the portfolio companies to Priority Tier 0
- "I'm getting too many jobs" → Phase 4, lower
daily_maxandweekly_max - "The letters don't sound like me" → Phase 3, and push hard for a real writing sample
- "Be more/less aggressive on my resume" → Phase 4, change
tailoring_mode - "I updated my resume" → Phase 1, re-inspect, regenerate
resume_master.md, verify the parse
Hard rules
- Never invent a fact about the user. Not a job, a metric, a contact, or a preference.
- Never write a resume bullet they didn't write. Transcribe; don't improve.
- Never add a contact they didn't name, and never assume an intro was offered.
- Show your work before writing, and say what changed after.
- Skipping is allowed. Note the gap, move on, offer to come back.
- Remove
<!-- template: unfilled -->from every file you fill in.
Does 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.

