job-scout

name	job-scout
description	Locate currently-open jobs at Finn's performance×technology nexus (DevRel/dev advocate, AI video & content production, production-tech/AV-tech, live events & stage) in NYC, NYC-office, or remote-OK-to-NYC. Loads when the user wants to find/scout/locate current job listings, check "what's hiring", look for new openings, or update nyc-job-targets.md. Searches live sources, confirms roles are actively hiring via public ATS APIs, dedupes against existing targets, ranks by genuine fit, and appends new entries to nyc-job-targets.md. Hands blocked boards to the render-job-url skill; does not write resume copy or build PDFs.
tools	Bash, Read, Edit, WebSearch, WebFetch

Job Scout (performance × tech, NYC)

Find the next batch of live job listings that fit Finn's unusual nexus: real tech credibility (Python, NixOS homelab, LLM automation, networking/security) plus deep performance/production (musical theatre, voiceover, improv, on-camera, A/V editing, audio mastering, lighting, stage management). Locate, verify, dedupe, rank, append.

When to load

Load when the user says any of: "find/scout/locate jobs", "what's hiring", "any new openings", "look for [role family] roles", "update job targets". Do NOT load for editing resume copy, building PDFs, or fetching a single already-known URL (that's plain WebFetch, or render-job-url if blocked).

Start of every run

1. Read profile.md (this dir) — the fit rubric and the per-family search queries.
2. Read boards.md (this dir) — where to look and the live-check ATS endpoints.
3. Read ../../../nyc-job-targets.md — the existing targets, for dedupe and numbering. (Skill dir is .claude/skills/job-scout/; the file is at the repo root.)

Confirm before proceeding: you have read profile.md (queries + differentiators), boards.md (ATS endpoints + discovery sources), and nyc-job-targets.md (existing entries + numbering). If any are missing, stop and read them now.

The loop

1. Search with tavily_search (available via MCP; check your tool list for the exact name, typically mcp__websearch__tavily_search) and WebSearch, using the query strings from profile.md. Other MCP search tools available: tavily_extract, tavily_map, tavily_crawl, tavily_research. Bias hard to recency: last 30–45 days, "actively hiring", "open roles". Run the four role-family queries; add the infra fallback set only if asked for straight-engineering targets.

2. Identify the ATS platform for each hit. Extract the slug from the careers URL: Greenhouse follows job-boards.greenhouse.io/<slug>/, Ashby follows jobs.ashbyhq.com/<slug>, Lever follows jobs.lever.co/<slug>, SmartRecruiters follows jobs.smartrecruiters.com/<slug>. If the company uses Workday, iCIMS, Taleo, or a custom form (no public JSON API), skip live-JSON verification and go straight to step 3 with a soft-lead marker. 2b. Confirm live for supported-ATS hits only. Hit the public ATS JSON endpoint for the company (see boards.md) and check the role is actually present and open. A posting that 404s, is closed, or is absent from the board's live JSON is a dead lead — drop it, do not append it. For companies already in the targets file, the ATS slug is in their Source: line.

3. Fetch detail for survivors with WebFetch. On empty body / Cloudflare challenge / 403 / 429 / login wall, hand the URL to the render-job-url skill. Never reimplement fetching here, and never scrape LinkedIn/Indeed/Workday directly — discover via search, then render-job-url.

4. Dedupe against nyc-job-targets.md:

   • Match on internal requisition ID first (Greenhouse: id, Lever: id, Ashby: id, SmartRecruiters: id) — the most reliable key.
   • Fall back: company + normalized role title + source domain.
   • Normalize titles by lowercasing and stripping parentheticals before comparison (e.g., "Developer Advocate, Events & Social (NYC)" → "developer advocate events social").
   • Skip anything already listed, including APPLIED, No longer accepting, honorable-mentions, and "also monitor" entries.
   • If an existing entry's status is not marked APPLIED, REJECTED, or No longer accepting, and it's >60 days old, re-verify it against the ATS endpoint before treating it as a dedupe match. Stale open entries may be closed.
   • When a scout run re-verifies an existing entry, append a status note to that entry's Source: line: _(verified still open YYYY-MM-DD)_ or _(confirmed closed YYYY-MM-DD)_. Do not delete closed entries — mark them No longer accepting and add the confirmation date. This keeps the file's staleness visible without rewriting history.

5. Rank each new role with the rubric in profile.md: count genuine differentiator matches, reward postings that invite non-exact-match applicants, and be honest about seniority gaps. Drop anything that would need fabricated experience to fit.

   Before writing any "Why it fits" line, re-read the relevant sections of metadata/work-experience.toml, metadata/tech-projects.toml, metadata/tech-skills.toml, and any variant-specific TOML you plan to reference. Each claimed differentiator must map to a concrete entry in these files. If you cannot point to a specific entry, do not make the claim.

   Before finalizing any entry for an AI-video or creative-production role, verify no AI-video tool experience is claimed (Runway, Pika, CapCut AI — per profile.md). If the posting asks for these, flag it as a Gap, don't fabricate.

   After writing the entry, re-check each "Why it fits" claim against the TOML files one more time. If a claim doesn't correspond to a concrete entry, remove it. 5b. Verify citations before append. Re-read each "Why it fits" bullet and confirm it contains at least one TOML key citation in the form file.toml [section.key]. If any claimed differentiator lacks a citation, either add the citation (if a concrete TOML entry maps) or remove the claim (if none maps). Do not append an entry with uncited fit claims. This is the post-write check that enforces the pre-write re-read in step 5.

6. Append each surviving role to nyc-job-targets.md (see format below). 6a. Stamp the run. Update or insert a <!-- Last scouted: YYYY-MM-DD --> HTML comment at the very top of nyc-job-targets.md (above the first heading). If one already exists, replace the date. This is the only edit to an existing line allowed outside the append-only rule, because it is metadata about the run itself. 6b. Validate the Resume pointer. Read variants.toml. If the slug you chose is already registered as a variant, write the Resume: line normally. If not, append _(variant not yet created)_ to the Resume line so the user knows the .typ file needs to be built before nix build will include it.

Append format (match the existing file exactly)

Numbering rules by section type:

• Main numbered section (1–22): continue with ## 23., ## 24., etc.
• Lettered sub-sections (e.g. Anthropic A–C): continue with the next letter for that company. A new company gets its own # <Company> — <Theme> Targets header and fresh ## A. numbering.
• "Additional … Targets" batches: use the existing batch header convention (## Additional … Targets (<Month YYYY> research)) with sequential numbering continuing from the highest existing number.
• Honorable mentions / Also monitor: append bullets, not numbers or letters.

Each entry:

## N. <Company> — <Role> (<location>; <comp if known>) _(status if any)_

- <what the role is, in the posting's own framing — 1-2 bullets>
- Why it fits: <which real differentiators map on; selective keyword echo, never parroted. Each claimed differentiator MUST cite a concrete TOML entry in the form `file.toml [section.key]` — e.g., `work-experience.toml [jobs.virtual-choir]`. If you cannot cite a specific entry, do not make the claim.>
- Gap: <always include; write "None — early-career-appropriate" when the role is
  genuinely junior/mid-level. Otherwise describe the honest gap.>
- Resume: `resumes/<slug>.typ` → `<slug>-resume.pdf`
- Source: <verified live URL; ATS slug if useful> _(verified live YYYY-MM-DD)_

Pick <slug> as a short company stem (e.g. vercel, descript), matching the file's convention. The Resume: line is a pointer for the documented variant workflow — this skill does NOT create the .typ file or touch variants.toml.

Slug convention: use the company stem alone (e.g. vercel, descript) when the company has only one target role. Use <company>-<role> (e.g. vox-unexplainable, nyt-tech-specialist) when the company has multiple distinct targets. Match the existing file's pattern.

Em-dash scope: CLAUDE.md's "No em dashes, ever" rule applies to resume copy (.typ files, TOML descriptions, header quotes). Em dashes are allowed and expected in nyc-job-targets.md headings and entry body text to match existing file style.

Hard rules

• Never fabricate a listing, a detail, or a fit. If a role can't be confirmed live, either drop it or mark it _(soft lead — unverified)_ exactly as entry #22 does — never promote an unverified lead to a normal entry.
• Append-only to nyc-job-targets.md. Never rewrite, reorder, or delete existing entries. Use Edit to add; diff should show only additions.
• Honesty over coverage. Finn is early-career; flag every seniority stretch. The file's Anthropic/Vox entries are the model for honest gap-flagging.
• Absolute dates. Convert "this month" / "recently" to a concrete date (today's date is the reference) in anything written.
• Stay in lane. No resume copy, no PDF builds, no applying, no form-filling.

Reference

• profile.md — Finn's differentiators, per-family queries, ranking rubric.
• boards.md — sources + live-check ATS endpoints.
• ../render-job-url/SKILL.md — fallback fetcher for blocked boards.
• ../../../nyc-job-targets.md — the file this skill reads and appends to.
• ../../../CLAUDE.md "Writing Resume Content" — the voice rules any emitted copy follows.