build-profile

build-profile

Produce profile/master-profile.md from whatever the user already has, then fill gaps with a short, track-aware interview. Ingest first, ask last.

1. Gather what exists (don't ask the user to retype anything)

Collect any of these the user provides; actively offer the options:

• Resume / CV — see step 1a for PDFs (mandatory link-extractor first), then read the body with your runtime's file-read tool.
• LinkedIn PDF — read it. (Point them to LinkedIn → profile → Resources → Save to PDF if they don't have it.)
• AI-memory (proactively offer it). Ask once: "Have you used ChatGPT, Claude, or Gemini for your job search or career? Paste this prompt into the one you use most and bring back the result — I'll fold in recent interviews, feedback, and career details you'd otherwise have to retype." Then surface the prompt from ../../profile/IMPORT_MEMORIES_PROMPT.md (it pulls only job-search-relevant memories, not everything). On ingest, merge only new facts — never duplicate what the resume/profile already has; mark uncertain items [VERIFY].
• Links — record every link, then enrich the user's own ones (see §1b). LinkedIn stays record-only (ToS).

Read everything before asking anything.

1a. For every PDF input — run the link extractor FIRST

A typical agent file-read tool surfaces visible anchor text only (it shows the word "LinkedIn" but not the URL it's hyperlinked to). LaTeX/Word/Docs resumes routinely embed \href{url}{anchor} annotations whose URLs are invisible to a plain read. You will silently lose the user's LinkedIn / GitHub / portfolio URLs if you skip this step — which is why extract_links.py exists.

python3 skills/build-profile/scripts/extract_links.py <path-to-resume.pdf>

The script decompresses every FlateDecode stream and pulls every /URI annotation it can find. Stdout is JSON:

{
  "urls": ["https://...", ...],           // every URL found, in order
  "classified": {
    "linkedin": "https://linkedin.com/in/...",
    "github": "https://github.com/...",
    "other": { "<host-or-label>": "<url>", ... }
  }
}

The script also returns a status field — branch on it:

• ok → use the classified map directly to populate links: (canonical keys land in their named fields; unknowns land in links.other). Surface any links.other entries with one short question — "I found these other URLs in your resume: <list>. Keep any as portfolio links?" — so project pages (e.g. lambdatest.com/visual-ai-testing) can be promoted to portfolio if relevant.
• encrypted / scanned_image / unreadable → the script never crashes (exit 0). When status is one of these and the file-read tool also returns little/no body text, the PDF is unusable: tell the user what's wrong (use the script's note) and offer the fallback ladder — (a) paste the resume text, (b) supply a .txt/.md/.docx version, or (c) re-export from LinkedIn (profile → Resources → Save to PDF). Explicitly warn that embedded links are lost on scanned/encrypted PDFs and ask them to paste their LinkedIn / GitHub / portfolio URLs manually.

After the extractor, use your runtime's file-read tool for body text. For non-PDF inputs (markdown, pasted text, LinkedIn URLs), scan the text directly with regex for any https?://… matches.

1b. Enrich the user's own links (don't just record URLs)

A portfolio or personal site usually holds proof points the resume omits (projects, metrics, roles, skills). Mine the user's own links — portfolio, personal site, project pages, and GitHub beyond the API — with a graceful fallback ladder. Skip LinkedIn (record-only, ToS). Cap at the few most relevant links so onboarding stays fast.

For each link, stop at the first tier that yields usable content:

1. GitHub → WebFetch https://api.github.com/users/<user> + /repos?sort=stars (top repos, languages, bio). No key.
2. WebFetch the page → extract real proof points from the readable HTML.
3. WebSearch (when WebFetch returns an SPA shell / thin content) → search the person + domain/handle for indexed content (project write-ups, talks, press).
4. Playwright MCP (only if the session exposes it — optional, not required) → mcp__playwright__browser_navigate then mcp__playwright__browser_snapshot to render a JS-only page and read it. If Playwright isn't available, skip it and ask the user to paste a few highlights instead.

Rules: ingest only new facts not already in the resume/profile (no duplication); mark anything inferred (not explicitly stated by the user) as [VERIFY]; never fabricate. Summarize what you learned back to the user in one line per link ("Your portfolio shows X, Y — added as projects").

2. Know the track and region

• Track (ask if unclear): product · software · finance · quant · marketing · founders-office · hr · design · data · sales · operations · customer-success · content. Decides which fields matter and which links to push for — see reference/track-presets.md.
• Region: infer the applicant's region using this ladder, stopping at the first that resolves:
  1. Location string → region code (New York→US, Bengaluru→IN).
  2. Contact clues if location is missing/ambiguous (e.g. "Remote") → a +91 phone or ₹/CTC/LPA mentions → IN; a +1 phone or $/state abbreviations → US.
  3. target.regions if already known from the conversation.
  4. Otherwise ask one short question: "Which country are you based in, and where are you hunting?" Codes must match a pack in ../../knowledge/regions/. Never silently default to US. When region was inferred from weak signals (clues, not an explicit location), write it as [VERIFY] so downstream skills treat it as soft. Then confirm where they're hunting (target.regions). The chosen pack dictates which region-specific fields to collect (IN → current/expected CTC + notice period; US → work authorization per ../../knowledge/work-authorization.md) and what to omit (e.g., don't collect photo/DOB).

3. Ask only for gaps

Compare what you extracted against reference/profile-schema.md. Then ask a short, batched set of questions for only the missing or [VERIFY] fields, prioritizing the track's must-haves, the region's required fields (per the region pack), and anything find-jobs/score-fit need (target roles, locations, remote preference, salary floor; work authorization for US; current/expected CTC + notice period for IN). Don't interrogate — 5–8 crisp questions max, grouped, with sensible defaults offered.

3a. The single optional links-gap prompt

After populating links: from extraction, run one track-aware optional prompt to surface what the resume didn't include. Use the link gap matrix in reference/track-presets.md to decide what to suggest.

Shape:

"I found these links: <list of canonical keys filled>. For <track> roles, candidates also commonly include: <strongly + optional gaps from the matrix>. Want to add any? (Paste or skip.)"

One prompt. Dismissible. No follow-ups. The user can paste any handle/URL or skip entirely.

3b. Two closing context prompts

The context: block is read by find-jobs, score-fit, and tailor-resume to bias direction — never rendered into a PDF or external message. Ask once:

1. "In 2–3 sentences, what are you actually optimizing for over the next 12–24 months?" → context.career_goal (≤ 600 chars).
2. "Anything else I should know that won't go on the resume — geo constraints, life situation, things to avoid?" → context.additional_info (≤ 400 chars). Skip if the conversation already covered it; record what surfaced.

Both fields are optional but high-value.

3c. The single optional proof-point gap prompt

Gaps aren't just blank fields — the highest-value gap is evidence a strong candidate for this target would show but the resume doesn't. Compute the expected proof points for the user's target.tracks (+ AI archetype if applicable) from the "Proof-point coverage" section of reference/track-presets.md (which points at the archetype proof points in ../../knowledge/ai-roles.md for AI tracks). Check coverage against the extracted Experience/Projects/Achievements, then ask one target-framed prompt:

"Strong <target track/archetype> candidates usually show <X / Y / Z>. From your resume I can see <X>; I don't see <Y, Z> — anything to add there? (Paste or skip.)"

This is the fix for the case where a missing portfolio/case-study (marketing) or eval/LLMOps evidence (AI Platform) is the candidate's #1 gap for their target, not a generic blank. Frame it by the target, never invent the evidence. One prompt. Dismissible. No follow-ups. Shares the relevance vocabulary in ../../knowledge/relevance.md.

4. Write the profile

Write profile/master-profile.md following reference/profile-schema.md exactly (YAML frontmatter + markdown sections). Rules:

• Never fabricate. If a fact is unknown, leave the field empty / omit it; mark anything uncertain inline as [VERIFY].
• Quantify achievements where the source gives numbers; don't invent metrics.
• Capture the user's voice/constraints in the Notes section (tone, hard limits like no-relocation, sponsorship needs) — downstream skills rely on it.
• Populate every link the extractor found (canonical keys to their named fields; everything else under links.other). Keep unfamiliar URLs even if you can't classify them.
• If the file already exists, update in place (preserve user edits; only fill/correct).

5. Confirm

Show a tight summary (identity, target, # roles captured, # links captured, gaps still [VERIFY]) and tell the user they can now run /find-jobs. Keep tone encouraging and concise — no emoji.

Heads-up on setup (mention once, now). Run python3 scripts/doctor.py to check dependencies. Everything works with no install except producing resume PDFs, which needs rendercv (a one-time install: python3 -m venv skills/render-resume/.venv && skills/render-resume/.venv/bin/pip install "rendercv[full]"). Flag this here so it's installed before the user reaches /tailor-resume, not mid-render. (API keys are all optional; doctor lists which strengthen find-jobs/research.)

Next steps

Profile saved. Next: /find-jobs to pull matching roles — or /apply-to-job <url> to evaluate one specific job end-to-end.