MLGalusha/resume-tailor

resume-tailor

This skill produces applications/<app_id>/resume.md — a tailored resume for a specific job. It is the most truthfulness-critical skill in the system. Every claim it makes must be defensible in an interview against the actual shipped code in Mason's projects.

---

When to invoke this skill

• Mason says "write the resume for the <company> one"
• Mason says "apply to <company>" and the previous step was role-intake
• The role-intake skill recommends applying and Mason confirms

Do not run this skill if:

• role.md for the app does not exist → run role-intake first
• role.md has recommended_action: skip — hard filter → ask Mason; don't auto-run on a skipped role
• role.md has recommended_action: skip — comp below floor → ask Mason
• Mason is just thinking about a role and hasn't decided → wait

---

Prerequisites

Before running:

1. applications/<app_id>/ directory exists
2. applications/<app_id>/role.md exists and has been run through role-intake
3. profile/skills.md, profile/wins.md, profile/repos.md, profile/voice.md, templates/resume-base.md all exist

---

Research-backed contract (2026 updates)

Rebuilt on notes/research/resumes-2026.md and notes/research/SYNTHESIS.md. Five non-negotiables:

1. Resume is an index, not the story. Its job is to earn a click on the GitHub / demo / blog. Optimize for one-page density and project-first structure, not for comprehensive coverage.
2. Single column, one page for anything under 5 YoE. Mason is 0 YoE professionally — always one page. The 1.5-page exception previously allowed is retired.
3. Projects-first ordering. Selected Projects comes immediately after Summary. Skills comes after projects, not before. Education last.
4. JD-phrase coverage sweet spot is 50-70%, not higher. Above 70% triggers the "too mirrored" suspicion filter on LLM screeners and human reviewers. If you catch yourself pulling more than 70% of the JD's phrases into the resume, back off — leave some of them on the table and let the projects do the talking.
5. Every project bullet must have a real engineering metric. For Mason's projects (none have end-user counts), use latency, p95, eval scores, token cost, commit counts, endpoint counts, dataset size, LOC, build time, deploy frequency. Never fake user counts. Never invent metrics. If a project has no defensible number, that's a signal the bullet shouldn't be on the resume.

---

The truthfulness contract (READ THIS FIRST)

A tailored resume may only contain claims that trace to one of:

1. Tier A of profile/skills.md (production-validated, shipped in a real project)
2. Tier B of profile/skills.md — but only with hedged language ("familiar with", "experience with"), never as a strength
3. A profile/projects/*.md file where the claim is in the "Claim guardrails" section
4. profile/wins.md

Never:

• Invent years of experience Mason doesn't have
• Soften a 🔴 Gap from role.md into a claim
• Copy JD language into the resume if Mason can't defend it
• Claim frameworks or tools Mason hasn't used
• Put Tier C (learning/gaps) items on the resume at all
• Claim something "we" built when Mason built it solo
• Overstate a side project's status ("shipped in production" when repos.md says "functional internal tool, not publicly deployed")

When in doubt, omit. A shorter resume that's 100% defensible beats a longer resume with one line you can't talk to.

If you find yourself wanting to make a claim that isn't supported: either (a) ask Mason for a citation and add it to the profile first, then use it, or (b) note it in the gap report and leave it off the resume.

---

The pipeline

1. Read the inputs (targeted, not exhaustive)

Read in this order:

• applications/<app_id>/role.md — the whole file
• profile/skills.md — whole file, you'll be citing from it constantly
• profile/voice.md — voice rules, banned phrases, formality dial
• profile/wins.md — whole file, source of quotable bullets
• profile/repos.md — to decide which repo links go on this resume
• profile/comp.md — for comp signaling (don't put comp on resume, but you need it for the "summary" tone)
• templates/resume-base.md — the max-volume source that the tailored resume is a subset and reordering of
• Then selectively read profile/projects/*.md for projects you intend to lead with (usually 2–4, not all 5)

Don't read every project file. Read only the ones you'll cite.

2. Decide the lane-specific shape

Read role.md's lane field and adapt the resume shape:

| Lane | Summary tone | Projects to lead with | Skills emphasis | Key differentiator | |---|---|---|---|---| | founding | Shipping speed, solo execution, range | VTR, Staffclaw, OpenClaw | Agent-directed dev, full-stack TS, GCP | "Shipped VTR solo in 3 weeks, agent-directed" | | ai-engineer | Production AI pipelines, cost-tiered models, HITL | VTR, PianoTranscriber, SonicGen | OpenAI API (GPT-5.4 + GPT-5.4-mini tiering), HITL, structured outputs | "Production vision pipeline on OpenAI Responses API with strict schemas and human gate" | | fde | Ability to teach customers to build with agents, translation skill | VTR, OpenClaw | Agent-directed dev, reverse-engineering, customer-facing framing | "I direct agents on real production systems and could teach your customers to do the same" | | fullstack-ai | Full-stack TypeScript, AI integration, shipping | VTR, Staffclaw | TypeScript end-to-end, React 19, Express, OpenAI API | "TypeScript end to end, AI in production" | | local-raleigh | Shipping work, bootcamp + local context | VTR, Staffclaw, PianoTranscriber | Full-stack, bootcamp, local | Softer on agent-directed angle unless the company signals it's agent-friendly | | off-lane | Do not tailor. Ask Mason if he really wants to apply. | — | — | — |

3. Select projects (2–4, not all 5)

From profile/projects/, pick the 2–4 projects that best match the JD's priorities (read from role.md required list). Ordering rule:

1. VTR is almost always #1 — it's your only shipped production project with real users. The only exception is a role that's purely about ML training depth (in that case lead with PianoTranscriber and put VTR second).
2. #2 slot goes to the project whose projects/*.md "Claim guardrails" section most directly hits the JD's required list.
3. #3 slot is for range — pick something that shows a different capability than #1 and #2.
4. #4 slot only if you still have space. Usually skip.

For each selected project, write 3–5 bullets that combine what projects/*.md says with verbs and numbers from wins.md. Each bullet must be traceable to a specific line in one of those files.

4. Tailor the Summary

Rewrite the Summary section from resume-base.md to lead with the lane's framing. Three sentences max. Must include:

• What Mason is (AI-native engineer / full-stack TS engineer / founding-engineer-shaped)
• One concrete shipped thing (VTR, almost always)
• What he's looking for (one clause, matched to this role's lane)

Do not include comp numbers on the resume. Do not include a list of "years of experience."

5. Tailor the Skills section

From profile/skills.md Tier A and Tier B, pick the subset that maps to this JD. Ordering rule:

1. Agent-directed development stays near the top for agent-friendly companies; drop it below Languages for traditional companies
2. Languages — list only the ones in Tier A (shipped code), order by relevance to this JD
3. Frontend / Backend / AI-ML / Cloud / Other — order the headers by this JD's priorities (read the JD's required list)
4. Do not list Tier C items. Ever.
5. Do not pad with items the JD doesn't care about. If the JD is for a backend Python role, don't include React just because it's in Tier A.

6. Repo links — check repos.md

For any project you link, the link URL comes from profile/repos.md:

• VTR — link voicesthatremain.com, never the repo

• Staffclaw — link the public GitHub repo. The Staffclaw narrative arc is fixed and important. Lead with it in this exact order:

  1. OpenClaw mapped all 92 Teamworxs API endpoints in one agent session — including the manager and owner endpoints Mason had no access to as a cook, with full request/response shapes. This is the enabling discovery.
  2. Reverse-engineered the entire app, not just the employee side — Mason set out to rebuild Teamworxs's manager schedule editor too, using only the discovered API surface.
  3. Drop-in replacement architecture (this is the killer signal): Staffclaw is a transparent client over Teamworxs's actual backend. Users log in with their real Teamworxs credentials, Staffclaw stores the Teamworxs session cookie, every read/write proxies through Teamworxs's API on the user's own cookie. Migration cost from Teamworxs to a finished Staffclaw would be literally zero — log in with the credentials you already have, no data export, no company onboarding. Verified in server/src/routes/auth.ts and server/src/services/teamworxs.ts.
  4. Multi-source external APIs as input for an auto-scheduler — Alamo Drafthouse public showtime/ticket data, weather APIs, local-events data, history analyzer. The auto-scheduler is the hard unsolved problem; the multi-source data layer exists to feed it, not as a vanity dashboard. Verified in services/alamo.ts, services/weather.ts, services/showtime-poller.ts, services/history-analyzer.ts, routes/demand.ts, cron/weather-fetcher.ts.
  5. Paused when VTR took priority.

  Do NOT lead Staffclaw with "I built a schedule dashboard" or "I built a pre-shift briefing." The pre-shift briefing is a byproduct. The architecture (drop-in replacement via credential proxying) and the discovery (full API surface including manager-side via agent) are the parts that prove unique technical thinking.

• PianoTranscriber — link the public GitHub repo

• SonicGen — link the public GitHub repo (github.com/MLGalusha/SonicGen). Standalone audio dedup engine. Use the "written by hand, no coding agents" framing as a counter to "only ships via agents" interview objections — this is one of SonicGen's most valuable signals.

• job-tracker — link only if the JD mentions agent-directed work explicitly

• OpenClaw, the private truth-engine dev repo, line-finder — no link. Reference in prose only if relevant. (For SonicGen, always link the public MLGalusha/SonicGen repo, never the private truth-engine dev repo it was developed in.)

If you catch yourself writing a URL that isn't in repos.md, stop and re-check.

7. Length discipline

• Hard target: one page. Single column. No exceptions for Mason's current experience level.
• Projects-first structure. Summary → Selected Projects → Skills → Education. Do not reorder.
• Cut first: the Background note about the cooking job. Include only if the lane is local-raleigh or Mason explicitly wants it in.
• Education stays short. Bootcamp + CS50. No fluff. No "6-month intensive" credential-padding per voice.md rule 3.

7.5. JD coverage cap check

After drafting bullets and the skills section, count the JD phrases you've echoed back in the resume. If coverage is above 70%, pull some back. The research calls this the "too mirrored" suspicion filter — reviewers and LLM screeners both flag resumes that read like a JD restatement. Target 50-70% coverage. Leave some JD phrases on the table deliberately and let the lead project do the talking.

If coverage is below 40%, the lane is wrong and you should flag that to Mason before continuing.

7.6. Engineering metrics check

Scan every project bullet. Each one should have a concrete engineering number. Acceptable number shapes for Mason's projects:

• Commit counts, LOC, endpoint counts, dataset size (bytes, rows, files)
• Latency (ms, p95), eval scores, model accuracy, token cost
• Build time, cold-start time, deploy frequency
• Time budget ("shipped in 3 weeks"), solo-execution claim

Unacceptable numbers for Mason's projects:

• User counts for projects without real users
• "Increased X by Y%" when there's no baseline
• "Processed N records" when N is fabricated
• Any round number that isn't traceable to profile files

If a bullet has no defensible number, either find one from wins.md / projects/*.md or cut the bullet.

8. Voice audit (blocking)

Before writing the file, scan your draft against profile/voice.md. Blocking violations:

• Any banned phrase from rule 3 (passionate, rockstar, synergy, extensive experience, helped build, strong fundamentals, results-oriented, self-starter, capstone, intensive in credential-padding context)
• Any word from rule 15's banned LLM vocabulary (leveraged, spearheaded, delve, seamless, robust, cutting-edge, state-of-the-art, tech-savvy, adept, pivotal, tapestry, ecosystem, empower, foster, streamline, game-changing, best-in-class, world-class, transformative, innovative without concrete justification, disrupt, etc.)
• Any em-dash in a sentence (rule 14) — acceptable only in headers and date ranges
• Any stacked hyphen compound adjective (rule 14)
• Any approximation tilde (rule 14)
• Any pipeline arrow in prose (rule 14)
• Any brand-name drop in bullet prose (rule 12) — brand names belong in the Skills section only
• Any framing that makes Mason sound like he builds LLM infrastructure (rule 11)
• Any AI-washing of Mason's projects (rule 18)
• Any self-deprecation (rule 9)
• Agent-directed framed as a caveat instead of a capability (rule 4)
• Any summary sentence that spends more than ~10 words describing a single project (rule 13)
• Any bullet that uses "Responsible for" instead of a first-person declarative verb

When a violation is found, rewrite the whole bullet. Don't patch.

9. Write the file

Write to applications/<app_id>/resume.md. Use markdown. Structure:

# Mason Galusha

Raleigh, NC · [email protected] · 304-546-5850
[github.com/MLGalusha](https://github.com/MLGalusha) · [linkedin.com/in/masonlgalusha](https://www.linkedin.com/in/masonlgalusha) · [voicesthatremain.com](https://voicesthatremain.com)

## Summary

<tailored 2-3 sentences>

## Selected projects

### <Project Name> — <year>
**<link>** · <one-line framing>
- <bullet with concrete verb + number>
- <bullet>
- <bullet>

### <Project 2> — <year>

**Title format rule:** project headers are **just the project name and the year** — no "Solo builder," no "Solo builder + architect," no "Primary builder," no "(UNC Chapel Hill AI bootcamp capstone)." Keep it minimal: `### <Name> — <year>`. Solo / role information belongs in the body of the bullets if it matters at all, not in the header.

## Skills

**<Category>:** <tier-A items>
**<Category>:** <tier-A items>

## Education

**UNC Chapel Hill AI Bootcamp** — 6-month intensive, May–December 2024
- <1 line>

10. Log the artifact via application-logger

After writing the file, invoke application-logger to record that the resume was produced:

cd ~/Workspace/job && pnpm log resume_updated --app_id=<app_id> --path=applications/<app_id>/resume.md --summary="Initial tailored resume for <role>"

11. Write a gap report alongside

In the same directory, write applications/<app_id>/resume-gaps.md:

# Resume gap report — <company> <role>

Claims the JD wanted that were NOT put on the resume because they aren't supported by the profile. These go to the `interview-prep` skill as study topics, NOT on the resume.

- **Gap 1** — JD wants X. Profile has Y (closest). Reason it's still a gap: ...
- **Gap 2** — ...

## Claims on the resume that might get probed in an interview

- Claim 1 — defensible because: <citation to profile or project file>
- Claim 2 — ...

This file is how interview-prep knows what to prepare for later.

12. Report to Mason

Report:

• Path to the resume
• 1-line summary of the shape (which projects you led with, which skills you emphasized)
• The gap list in 2–5 bullets
• Any truthfulness concerns you had while writing (if any)
• Anything you wanted to put on the resume but couldn't — so Mason can decide to add evidence to the profile if it exists

---

Rules and guardrails

• Every bullet must be defensible. If Mason can't talk about the specific line in an interview, cut it.
• No "X years of experience" claims. Mason doesn't have years of professional experience. He has shipped projects. Let the projects do the talking.
• No claims beyond resume-base.md. The base is the ceiling. Tailored resumes are a subset and reordering — never stronger than the base.
• Agent-directed framing stays confident. Never "I also use AI tools" — always "I build primarily by directing coding agents." See voice.md and wins.md.
• Do not flatter the JD by mirroring its language if Mason can't back it up. If the JD says "5+ years of Kubernetes in production" and Mason has no Kubernetes experience, do not write "Familiar with modern container orchestration." Just leave it off.
• Comp lane awareness: do not put numbers on the resume, but know which lane you're writing for. A resume for a $100k local Raleigh role reads differently than one for a $140k remote AI-native role.
• Write once, revise once. If the first draft feels wrong, read it against voice.md and fix. Don't ship a first draft.

---

What this skill does NOT do

• Does not write the cover letter (that's cover-letter-writer).
• Does not do company research (that's company-research).
• Does not apply to the job or submit the resume anywhere.
• Does not mutate application status — that's application-logger when Mason actually submits.
• Does not write to data/ files directly. Uses pnpm log via the application-logger skill to record the artifact.
• Does not overwrite an existing resume.md without asking first. If a tailored resume already exists for this app_id, ask Mason whether to revise or replace.