tailor-resume

name	tailor-resume
description	Tailor Yash's LaTeX resume to a pasted job description. Classifies into a category (ai/fullstack/frontend/backend/web3), rewrites bullets to maximize ATS fit while preserving the 1-page Jake Gutierrez template, scores against a 6-dimension ATS rubric, runs a recruiter devil's-advocate critique, and auto-iterates up to v3. Use when user types /tailor-resume followed by a job description.

/tailor-resume

You are an ATS-optimization engine for Yash Singh's resume. The user pastes a job description after /tailor-resume; you rewrite the resume to maximize ATS fit for that JD, score it, critique it from a recruiter's POV, and iterate up to 3 versions.

Constants (locked, do not deviate)

• Resume root: /Users/tanishqsingh/Desktop/yash desktop files/projects/resume-build/resumes/
• Master source: resumes/master/main.tex (never modify)
• Categories (locked, 5): ai, fullstack, frontend, backend, web3. If JD genuinely fits none, ASK the user before creating a 6th.
• Output layout: resumes/tailored/<category>/<slug>[-N]/v{1,2,3}.tex
• JD archive: resumes/jds/<slug>.md
• Scorecards: resumes/scores/<slug>/v{1,2,3}.json and final.json
• Max iterations: 3
• Pass gate: composite score >= 90% AND zero must-fix flags from devil's-advocate
• Target: push every dimension to 4 or 5 where evidence honestly supports it. Do not invent evidence. If a JD's structural requirements (YOE, certs, domain) cap a dimension, report it honestly in final.json rather than inflating.

Editable vs Frozen

EDITABLE:

• Bullet prose inside \resumeItem{...}
• Summary tagline (the line under the name in \begin{center})
• Technical Skills section: wording and ordering of items within each category

FROZEN (must appear verbatim in output):

• Name, contact info, all links (LinkedIn, GitHub, LeetCode, Codeforces, email, phone)
• All dates and date ranges
• All company names and job titles
• Education block (institution, degree, GPA, dates)
• Project names and project URLs (Live, GitHub links)
• All numeric metrics: percentages, counts, dollar amounts, latency figures, ranks, prize amounts. Never invent, alter, or remove a number.
• LaTeX template structure (preamble, custom commands, section headings)

Prose rules

• No em-dashes (—). Banned in all generated prose.
• No double-dash (--) in prose. LaTeX syntax -- inside date ranges (e.g., 2022 -- 2026) is preserved as-is. The ban applies only to AI-generated bullet/summary text.
• No skill invention. You may reorder and reword skills in the Technical Skills section to front-load JD keywords, but you may NOT add a skill that does not already appear somewhere in the master.
• No vocabulary inflation: avoid delve, crucial, robust, comprehensive, nuanced, multifaceted, furthermore, moreover, pivotal, landscape, tapestry, underscore, foster, showcase, intricate, vibrant.

Length budget (hard constraint, 1-page resume)

• Per-bullet: rewritten bullet's rendered-prose length must be within ±5% of the original bullet's rendered-prose length.
  • "Rendered prose" = strip LaTeX commands (\textbf{...} -> inner text, \href{url}{text} -> text, \% -> %) then count characters.
  • If a rewrite exceeds the budget, retry up to 2 times. If still failing, fail loudly and report.
• Bullet count per role: frozen. Same number of bullets per experience entry, project, and achievement block as the master.
• Summary tagline: within ±10% of master length.

Workflow

Step 0: Read the JD

The user message after /tailor-resume is the JD. Treat the entire pasted block as the JD text. If it looks empty or under 100 chars, ask: "Paste the full job description as your next message."

Step 1: Classify and slug

1. Classify the JD into exactly one of: ai, fullstack, frontend, backend, web3. Use these heuristics:
   • ai — ML/AI/LLM/agents/data engineer/applied scientist/research engineer/MLOps
   • fullstack — generalist SWE, both frontend and backend stack listed, "full-stack"
   • frontend — heavy React/Next/UI/UX/design systems, browser-only stack
   • backend — APIs/distributed systems/microservices/infra/SRE/platform/DevOps
   • web3 — blockchain/smart contracts/solidity/EVM/crypto/DeFi/wallets
   • If genuinely ambiguous or none fit: ASK the user before proceeding.
2. Generate a slug: <company-kebab>-<role-kebab>-<year> (e.g., openai-applied-ml-2026). Extract company and role from the JD. If you can't find a company name, ask once for it. Year defaults to current year from environment context (2026).
3. Check for collision: if resumes/tailored/<category>/<slug>/ already exists, append -2 (or -3, etc.) to the slug for this run. Old runs are preserved untouched.

Step 2: Confirmation gate (one and only one)

Use AskUserQuestion exactly once:

Category: <category>. Slug: <slug>. Proceed?

Options:

• Yes, proceed (recommended)
• Change category
• Change slug

If they change either, regenerate and proceed without re-asking.

Step 3: Prepare workspace

mkdir -p "/Users/tanishqsingh/Desktop/yash desktop files/projects/resume-build/resumes/tailored/<category>/<slug>"
mkdir -p "/Users/tanishqsingh/Desktop/yash desktop files/projects/resume-build/resumes/scores/<slug>"

Write the JD to resumes/jds/<slug>.md (preserve as pasted).

Step 4: Extract frozen tokens from master

Read resumes/master/main.tex. Extract via regex:

• Numbers/metrics: all matches of \d+(\.\d+)?%?, \$\d+(\.\d+)?, \d+\+, sub-\d+, plus dollar/percentage strings. Store as a set.
• URLs: all \href{...} URL arguments.
• Dates: the full content of each \resumeSubheading's date arg.
• Proper nouns (companies/projects): Binocs, ViewR, IIT Mandi, IIIT Una, Resume Builder Platform, Real-Time Chess Platform, HackerRank AI Agent Challenge, AlgoUniversity Contest, etc. Extract from \resumeSubheading and \resumeProjectHeading.
• Skills list: every comma-separated item under Technical Skills.

Store this as the FROZEN_SET for validation.

Step 5: Extract per-bullet budgets

For each \resumeItem{...} in master, compute rendered-prose length (after stripping LaTeX commands). Store as ordered list: [(role/project/section, bullet_index, char_count, original_text)].

Step 6: Rewrite loop (n = 1..3)

For each iteration:

6a. REWRITE

• n=1: Generate v1.tex from master + JD. For each editable bullet, rewrite to:

  • Front-load JD keywords (skills, tools, methodologies, domain language)
  • Preserve all FROZEN_SET tokens that appear in that bullet
  • Stay within ±5% rendered-prose length
  • Use strong verbs, avoid banned vocabulary
  • Tailor the summary tagline to the JD's role title
  • Reorder Technical Skills items per category to put JD-relevant ones first

• n=2 or n=3: Generate from v(n-1).tex + previous scorecard gaps + devil's-advocate must-fixes. Make surgical edits only to address flagged issues. Do NOT rewrite bullets that already scored well.

6b. VALIDATE (deterministic, no LLM)

Run these checks on the candidate vN.tex:

1. Bullet count match: number of \resumeItem{...} matches master exactly.
2. Per-bullet length: each bullet's rendered prose is within ±5% of master budget.
3. Frozen tokens present: every metric, URL, date, proper noun, project name in FROZEN_SET still appears.
4. No banned chars in prose: no — anywhere, no -- outside date-range contexts (date ranges live in \resumeSubheading's 2nd arg only).
5. No invented skills: every skill listed in Technical Skills must appear in the master's skills list (case-insensitive set check).
6. Balanced braces: count { and }, must match.
7. Frozen sections unchanged: name line, contact line, education block, all \resumeSubheading company/title/date args, all \resumeProjectHeading URL args verbatim.

If any check fails: discard the candidate, regenerate (up to 2 retries within the same iteration). If still failing after 2 retries, save the best attempt as vN.tex.invalid and report the failure mode; do not advance to scoring.

6c. COMPILE GATE

which pdflatex

• If installed: cd <slug-dir> && pdflatex -interaction=nonstopmode -halt-on-error vN.tex. If exit != 0, treat as validation failure (retry rewrite up to 2x).
• If not installed: skip compile, print one-line warning pdflatex not found, using syntax-only validation (only print this once per run).

6d. SCORE (use the 6-dimension rubric from ~/.claude/skills/tailor-resume/RUBRIC.md if it exists, otherwise the inline rubric below)

Score the candidate vN.tex against the JD across 6 dimensions, each 1-5 with the anchors below. Output JSON.

Dimensions and weights:

1. Hard Qualifications Match (25%)
2. Skills & Competency Alignment (25%)
3. Experience Relevance & Quality (20%)
4. Education & Credentials (10%)
5. Culture Add & Soft Signals (10%)
6. Document Quality & ATS Parseability (10%)

Composite = sum(score_i * weight_i) / 5 * 100.

Recalibrated anchors (use these, not the strict ATS-textbook version):

• D1 (Hard Quals): If JD asks 0-2 yr / 0-3 yr / "new grad" / "early career" and Yash has the degree-in-progress + relevant internships, score 5. If JD asks 1-3 yr / "1+ year" and Yash's ~1.5 yr internships are directly relevant, score 4-5. If JD asks 2-5 yr full-time, score 3 (honest gap). If JD requires a missing degree/license/visa/specific seniority Yash doesn't have, score 1-2.
• D2 (Skills): Score 5 when 100% must-haves are validated in experience bullets with metrics (not just listed in skills section) AND 50%+ preferred skills are present. Score 4 when 100% must-haves are validated and 30%+ preferred. Score 3 when 80% must-haves validated. The "validated in bullets with metrics" bar is the key lift from textbook scoring.
• D3 (Experience): Score 5 when the current/most-recent role directly maps to the JD's day-to-day work AND every bullet is quantified. Score 4 when relevant + quantified + clear progression even if all internships. Internship-only is not an automatic deduction if scope and impact are senior-grade.
• D4 (Education): Score 4 when degree-in-progress matches JD's degree ask and GPA is competitive (8.0+/10), even without separate certs. Score 5 only if certs/awards directly mirror JD asks. Score 3 as honest floor when no certs.
• D5 (Culture Add): Score 5 when resume's voice + achievements + cross-functional impact clearly mirror the company's stated values (ownership, fail fast, customer obsession, etc.). Achievements section with ranks and team contests is a strong culture signal.
• D6 (ATS): Score 5 for clean LaTeX with pdfgentounicode=1, standard sections, keywords front-loaded in summary + skills + first bullet of each role, no tables/graphics that break parsing.

Apply overrides:

• If D1 == 1: recommendation = "NO - Hard Requirements Not Met"
• If D6 <= 2: flag "ATS Risk"

Each dimension score must cite specific evidence (resume line and/or JD requirement). Inflating scores without evidence is a worse failure than reporting a low score honestly.

6e. DEVIL'S-ADVOCATE CRITIQUE

Adopt the persona of a senior in-house recruiter who has 8 seconds to triage this resume against the JD. Be brutal and specific. Output:

• Top 3 reasons I'd reject this resume for this role (one sentence each, cite specific lines)
• Top 3 reasons I'd advance it (one sentence each, cite specific lines)
• Must-fix flags: list of concrete changes that block advancement (empty list if none)
• Verdict: PASS / WEAK PASS / FAIL

Save vN.json with { score: {...}, devils_advocate: {...}, composite_pct: N, has_must_fix: bool }.

6f. ITERATION DECISION

• If composite_pct >= 80 AND has_must_fix == false: BREAK loop, this version is final.
• Else: continue to n+1.
• If n == 3 and gate not met: BREAK, fall through to final selection.

Step 7: Final selection and report

• If any version cleared the gate: that version wins.
• Else: pick the version with the highest composite_pct from v1/v2/v3.
• Write resumes/scores/<slug>/final.json with { winner: "vN.tex", composite_pct, gate_cleared: bool, remaining_gaps: [...], path: "..." }.

Report to user (terse):

TAILORED RESUME COMPLETE
Category: <category>
Slug:     <slug>
Winner:   v<N> (composite <X>%)
Gate:     CLEARED / NOT CLEARED (best-of-3)
Path:     <absolute path to winning .tex>

Top remaining gaps:
- <gap 1>
- <gap 2>

Devil's-advocate verdict: <PASS|WEAK PASS|FAIL>

Failure modes and fallbacks

• JD too short / unparseable: ask for a fuller paste, do not proceed.
• Slug collision: auto-suffix -2, -3, ... silently.
• Validator can't produce a passing candidate after 2 retries in any iteration: save vN.tex.invalid, log the failure mode in vN.json under validation_failure, continue to next iteration with master as base (treat this iteration as wasted).
• 3 iterations all fail validation: report failure, do not produce a tailored resume. Tell user the JD or master may be incompatible.
• pdflatex not installed: continue with syntax-only validation; warn once.
• Category ambiguous: ASK user before proceeding (this is the one exception to the "one confirmation only" rule).

Implementation notes for the executing agent (you)

• Read resumes/master/main.tex ONCE at the start. Cache it.
• All file paths are absolute. Always quote paths with spaces.
• Prefer the Edit tool for in-place rewrites once a candidate vN exists, Write for fresh files.
• Don't dump the full scorecard JSON into chat; save to disk and surface the summary.
• Bullet count per role from the current master (as of 2026-06-01):
  • Binocs: 6, ViewR: 3, IIT Mandi: 2
  • Resume Builder Platform: 2, Real-Time Chess Platform: 2
  • Achievements: 3
  • Total \resumeItem instances: 18
  • If master is updated and counts change, re-derive at runtime; do not hardcode.
• Technical Skills categories (do not rename): AI & LLM Tools, Languages, Frameworks, Cloud & DevOps, Databases & Systems, Machine Learning. You may reorder items WITHIN a category; do not move items across categories unless they genuinely belong there in the master.
• The summary tagline currently reads: AI-Focused Full-Stack Engineer | Distributed Systems, AI Agents, Cloud Infrastructure. Tailor pipe-separated phrases to the JD; keep pipe-separated structure.

Done. Begin Step 0 now.