Tailored & Humanised Resume Skill

Overview

Builds job-winning resumes that are both strategically optimised for a specific role and genuinely human in tone. Most AI-generated resumes fail not because the content is wrong, but because they sound like a machine wrote them. This skill fixes both problems together.

Two tracks run simultaneously:

• Relevance track: Surface the right experiences, match them to what the role needs, structure the story intelligently
• Humanisation track: Strip out AI tells at every stage so the final document reads like the applicant wrote it themselves

Guiding principle: A person's shot at a job should come down to their actual experience, not how polished their resume writing is. This skill closes that gap without fabricating anything.

Author: Siv Souvam

Reference Files

Read these files as needed during each phase:

File	Read during	Purpose
research-prompts.md	Phase 1	Company/role research templates, success profile synthesis
matching-strategies.md	Phase 3	Confidence scoring formula, reframing strategies
branching-questions.md	Phase 2.5	Experience discovery conversation patterns
truthfulness-guardrails.md	All phases	Anti-hallucination rules, title reframing constraints, metric verification
ats-compliance.md	Phase 4	ATS formatting rules, DOCX structure requirements
multi-job-workflow.md	When multi-job detected	Batch processing workflow

CRITICAL: Always read truthfulness-guardrails.md before starting any resume generation. Its rules apply to every phase.

When to Use

Use this skill when:

• User has a job description and wants a resume tailored to it
• User has existing resumes to draw from (PDF, DOCX, markdown, or plain text)
• User wants to optimise their application for a specific role or company
• User needs help finding and articulating experiences they haven't documented yet
• User wants their resume to sound natural, not AI-generated

Not designed for:

• Writing a resume from scratch with no existing material (needs at least one source resume in any format)
• Cover letters or LinkedIn optimisation (separate workflows)

Quick Start

Required from user:

1. Job description (text or URL)
2. One or more existing resumes (PDF, DOCX, markdown, or plain text — any combination)
3. Resume location: uploaded files in conversation, or a directory path

Workflow:

1. Build library from existing resumes → Phase 0
2. Research company/role → Phase 1
3. Create template (with user checkpoint) → Phase 2
4. Optional: Branching experience discovery → Phase 2.5
5. Match content with confidence scoring → Phase 3
6. Generate MD + DOCX + Report → Phase 4
7. User review → Optional library update → Phase 5

Multi-Job Detection

Before starting single-job workflow, check for multi-job signals:

• Multiple JD URLs
• Phrases like "multiple jobs", "several positions", "batch"
• Multiple company/role mentions

If detected, ask user if they want multi-job mode. If yes, read multi-job-workflow.md and follow that workflow instead. Single-job workflow below is unchanged.

---

Phase 0: Library Initialization

Goal: Build fresh experience database from user's existing resumes, regardless of format.

Supported input formats: PDF, DOCX, Markdown (.md), Plain text (.txt)

Process:

1. Locate resumes:

   • If user uploaded files directly: use uploaded files from /mnt/user-data/uploads/
   • If user provides a directory path: scan that directory
   • Accept any mix of PDF, DOCX, MD, and TXT files

2. Parse each resume by format:

   • PDF: Use the pdf skill to extract text. If the PDF is image-based (scanned), attempt OCR. Warn user if extraction quality is low.
   • DOCX: Use the docx skill or python-docx to extract text content, preserving section structure where possible.
   • Markdown: Read directly — section headers map to roles, bullets map to achievements.
   • Plain text: Read directly — use heuristics (date patterns, company names, indentation) to identify sections.

3. Normalise into common structure: Regardless of source format, extract into the same in-memory database:

{
  "roles": [{
    "role_id": "company_title_year",
    "company": "Company Name",
    "title": "Job Title",
    "dates": "YYYY-YYYY",
    "bullets": [{
      "text": "Full bullet text",
      "themes": ["leadership", "technical"],
      "metrics": ["17x improvement", "$3M revenue"],
      "keywords": ["cross-functional", "program"],
      "source_file": "resume.pdf",
      "source_format": "pdf"
    }]
  }],
  "skills": { "technical": [], "product": [], "leadership": [] },
  "education": []
}

4. Auto-tag content: themes, metrics (extract numbers/percentages/dollars), keywords
5. Announce: "Built library from {N} resumes ({X} PDF, {Y} DOCX, {Z} other) — {A} roles, {B} unique bullets"

Format-specific gotchas:

• PDF: Multi-column PDFs may extract with jumbled text order. If the parsed text looks garbled, warn the user and ask if they have a DOCX or text version.
• DOCX: Tables used for layout (common in designed resumes) may parse as fragmented text. Flag if structure looks wrong.
• Scanned PDFs: OCR quality varies. If text extraction yields very little content, tell the user: "This looks like a scanned PDF. I got limited text from it. Do you have a text-based version?"

Output: In-memory database ready for matching, with source file and format tracked per bullet.

---

Phase 1: Research

Goal: Build "success profile" — what actually makes someone good at this job, beyond the JD's literal text.

Read research-prompts.md for full templates. High-level:

1. JD Parsing: Extract explicit requirements (must-have vs nice-to-have), technical keywords, implicit preferences, red flags, role archetype
2. Company Research: WebSearch for mission/values/culture, engineering blog, recent news
3. Role Benchmarking: Search LinkedIn for similar role holders, analyze common backgrounds and terminology
4. Success Profile Synthesis: Combine into structured profile with core requirements, valued capabilities, cultural fit signals, terminology map, risk factors

Checkpoint: Present success profile summary to user. Wait for confirmation before proceeding.

Fallback: If research is limited (obscure company, WebSearch fails), fall back to JD-only analysis and ask user for context.

---

Phase 2: Template Generation

Goal: Create resume structure optimised for this specific role.

Process:

2.1 Analyse library: Extract role archetypes, experience clusters, career progression from user's existing resumes.

2.2 Role consolidation decision: When user held multiple roles at same company, decide whether to consolidate or keep separate. Present both options with rationale. Different companies are ALWAYS separate.

2.3 Title reframing: Adjust titles to emphasize aspects most relevant to target role.

⚠️ CRITICAL — Read truthfulness-guardrails.md for title reframing rules. Key constraints:

• NEVER claim work you didn't do
• NEVER inflate seniority beyond defensible
• Company name and dates MUST be exact
• Always show original title alongside proposed reframing
• Flag any reframing where new title implies a different job level

2.4 Generate template skeleton:

## Professional Summary
[GUIDANCE: {X} sentences emphasizing {themes from success profile}]

## Key Skills
[STRUCTURE: {2-4 categories based on JD structure}]

## Professional Experience
### [ROLE 1 - Most Recent/Relevant]
[TITLE OPTIONS: A: {option}, B: {option}, Recommended: {choice}]
[BULLET ALLOCATION: {N} bullets]
Bullet 1: [SEEKING: {requirement type}]
Bullet 2: [SEEKING: {requirement type}]

### [ROLE 2]
...

## Education

Checkpoint: Present template to user showing structure, consolidation decisions, title options, and bullet allocation. Wait for approval.

---

Phase 2.5: Experience Discovery (OPTIONAL)

Goal: Surface undocumented experiences through conversational branching interview.

Trigger: After template approval, if gaps identified (confidence < 60%):

"I've identified {N} gaps or weak matches:
- {Gap 1}: {Current confidence}%
- {Gap 2}: {Current confidence}%

Would you like a structured brainstorming session? (10-15 minutes typical)"

Read branching-questions.md for full conversation patterns.

Core approach: For each gap, start with open probe → branch based on answer (direct experience / indirect / adjacent / personal / no) → follow up systematically → capture immediately.

Capture structure per experience:

• Context (where/when), Scope (scale, duration, impact)
• Which gaps it addresses
• Draft bullet
• Confidence improvement estimate

After discovery: For each captured experience, user decides: Add to current resume / Add to library only / Refine further / Discard.

⚠️ CRITICAL: During discovery, help the user articulate what they did — but NEVER fabricate details, metrics, or specifics they didn't provide. See truthfulness-guardrails.md.

---

Phase 3: Assembly

Goal: Fill approved template with best-matching content, with transparent scoring.

Read matching-strategies.md for the full scoring system.

For each template slot:

1. Score candidates from library using weighted formula:

   • Direct match (40%): Keywords, domain, technology, outcome overlap
   • Transferable (30%): Same capability in different context
   • Adjacent (20%): Related tools, methods, problem space
   • Impact (10%): Achievement type alignment
   • Overall = (Direct × 0.4) + (Transfer × 0.3) + (Adjacent × 0.2) + (Impact × 0.1)

2. Present top 3 matches with confidence bands:

   • 90-100%: DIRECT — use with confidence
   • 75-89%: TRANSFERABLE — strong candidate
   • 60-74%: ADJACENT — acceptable with reframing
   • <60%: WEAK/GAP — flag to user

3. Handle gaps (confidence < 60%):

   • Option 1: Reframe adjacent experience (show before/after with truthfulness justification)
   • Option 2: Flag for cover letter
   • Option 3: Omit bullet slot
   • Option 4: Use best available with disclosure
   • User decides.

4. Apply content reframing where needed (see matching-strategies.md):

   • Keyword alignment, emphasis shift, abstraction level, scale emphasis
   • Always show original → reframed with explanation

Checkpoint: Present complete mapping with coverage summary, reframings applied, gaps identified, overall JD coverage %. Wait for user approval.

⚠️ CRITICAL: Every reframed bullet must pass the truthfulness test in truthfulness-guardrails.md. Never add metrics, facts, or specifics not provided by the user.

---

Phase 4: Generation

Goal: Create professional multi-format outputs.

4.0 Humanisation Pass:

Before compiling final outputs, run humanisation on all generated text. Apply the humanizer skill patterns. Focus on resume-specific AI tells:

• Promotional language: "passionate about", "results-driven", "proven track record" → Replace with specific facts
• Significance inflation: "pivotal role", "key contributor" → Concrete verbs and numbers
• Vague attribution: "led efforts to improve X" → "Reduced X by Y% by doing Z"
• Rule-of-three padding: "Built, deployed, and optimised X while fostering collaboration" → Break apart or cut
• Sycophantic summary openers: "Highly motivated professional" → Direct statement of what they do

Concreteness test: If you can remove a phrase and the bullet still means the same thing, remove it. If a phrase could appear on anyone's resume, replace it with something specific.

4.1 Markdown generation: Compile mapped content into clean markdown using user's formatting preferences.

4.2 DOCX generation: Use the docx skill. Read ats-compliance.md for ATS-safe formatting requirements.

4.3 PDF generation (optional): Use the pdf skill if user requests PDF.

4.4 Generation summary report: Create metadata report with target role summary, content mapping summary, reframings applied, source resumes used, gaps addressed, key differentiators, and interview prep recommendations.

Present to user: List files created, quality metrics (JD coverage %, direct matches %, newly discovered experiences), and ask for review decision.

---

Phase 5: Library Update (CONDITIONAL)

After user reviews generated resume:

Option 1 — Save to library: Move files to library directory, rebuild database, preserve generation metadata (source resumes, reframings, match scores, discovered experiences).

Option 2 — Need revisions: Collect feedback, make changes, re-present.

Option 3 — Save but don't add to library: Keep files in current location, don't enrich database.

---

Error Handling

Insufficient library (1-2 resumes): Warn about limited matching options, emphasize discovery phase value, proceed with available content.

No good matches (<60% for critical requirement): Present options transparently — run discovery, reframe best available, omit slot, note for cover letter. Never force matches.

Research failures: Fall back to JD-only analysis, ask user for context about company culture/tech/structure.

Vague JD: Flag missing detail areas, ask user for additional context, work with what's available.

Resume too long: Present pruning suggestions ranked by relevance score, let user decide priority.

Generation failures: DOCX/PDF fail → fall back to markdown-only with error details.

General principles:

• All checkpoints allow going back to previous phase
• Progress saved between phases where possible
• Transparent about limitations at every step
• User always has final decision authority

---

Usage Examples

Example 1: Internal role (same company)

• Library build → 29 resumes found
• Research → Internal culture + role benchmarking
• Discovery → Surfaces side projects and undocumented work
• Assembly → 92% JD coverage, 75% direct matches
• Result: Highly competitive application leveraging internal experience

Example 2: Career transition

• Research → Cross-domain transfer patterns identified
• Template → Reframes titles to emphasize transferable aspects
• Discovery → Surfaces volunteer work and graduate research in target domain
• Assembly → 65% coverage, gaps flagged for cover letter
• Result: Bridges technical skills with new domain

Example 3: Career gap

• Template → Includes startup period as legitimate role
• Discovery → Surfaces entrepreneurial skills (fundraising, product dev, team building)
• Assembly → Gap becomes strength showing initiative
• Result: Gap framed as valuable experience

Example 4: Multi-job batch (3 similar roles)

• See multi-job-workflow.md for complete batch processing workflow
• Shared discovery → per-job tailoring
• 3 resumes in ~40 minutes vs ~45 minutes sequential