agentleFS
bbborisssbbborisss/resume-bank/.claude/skills/create-resume-bank/SKILL.md

create-resume-bank

Skill0 starsChanged 32 days agoMITView on GitHub

Build a personal resume bank (_bank.md) from the user's existing resume and supporting materials. Parses everything they provide, interviews them to fill gaps and lock down facts, then generates _bank.md, preferences.md, and GLOSSARY.md and walks them into a calibration run of /tailor-resume. Run this once when setting up the repo, or again to do a full rebuild.

  • In this repositoryOne other file it gives its agents.

The file

Show its rules and text

We split this skill into the rules it gives an agent

13 rules. About 0 are common, 11 are distinctive, and 2 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈11

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 2

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    If the user gave you nothing at all, ask for at minimum their current resume (pasted text is fine) before continuing.

    line 25
  2. Distinctive:

    Read every file provided. For PDFs and images, read them directly.

    line 27
  3. Distinctive:

    Metrics: every number attached to an accomplishment, with where it came from

    line 35
  4. Distinctive:

    Blur guards: if two figures could be confused for each other on one page (two similar counts from different roles), record that they must …

    line 61
  5. Distinctive:

    Under-recording sweep: Ask directly about the surfaces resumes routinely leave out and JDs routinely ask for.

    line 63
  6. Distinctive:

    Cover letter every run, or only on request?: If most of their applications are forms that take no letter, on request saves wasted work.

    line 86
  7. Very specific:

    Write _bank.md in the working directory, using templates/_bank.template.md as the structural skeleton.

    line 96
  8. Very specific:

    Every role title, exactly as confirmed, with any alternate titles and when each applies, and the titles never held (with a note: "never …

    line 100
  9. Distinctive:

    Every role's exact date string, and which single role ends in "Present"

    line 101
  10. Distinctive:

    Never invent anything to fill a gap.

    line 130
  11. Distinctive:

    Never put a number in the closed metric set that the user didn't confirm as defensible.

    line 131
  12. Distinctive:

    Never write a speculative boundary into a lock.

    line 132
  13. Distinctive:

    Never write the user's materials or bank contents anywhere outside the working directory.

    line 133

≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Build (or rebuild) the user's resume bank. Arguments, if any, point at source materials: $ARGUMENTS

You are building the single source of truth that every future tailored resume will be generated from. Errors baked in here get repeated on every application, so this process favors asking over assuming. Work through the steps in order.

Two failure modes matter, and they are not symmetric:

  • Overclaiming puts something on a resume the user can't defend. It gets caught in a screen.
  • Under-recording leaves real experience out of the bank. It is invisible. Every lock in the bank is enforced as a closed list, so /tailor-resume can't tell "never did this" apart from "nobody wrote it down" and reports both as gaps, on every run. In practice it is the more common failure.

The interview below is built to prevent both: confirm every claim, and actively dig for what the resume never mentioned.


Step 1 — Collect source materials

Gather everything the user can give you:

  • Files passed as arguments or attached to the conversation
  • Anything in a source-materials/ folder in the working directory (tell the user they can drop files there: current resume in any format, old resumes, LinkedIn profile export or pasted text, brag documents, performance reviews, portfolio links, old cover letters)
  • If the user gave you nothing at all, ask for at minimum their current resume (pasted text is fine) before continuing. Do not build a bank from an empty interview; parsing real material first produces far better questions.

Read every file provided. For PDFs and images, read them directly; for URLs (e.g., a portfolio), fetch them.

Step 2 — Parse into a draft structure

From the materials, extract a draft of:

  • Contact info: name, city/state, email, phone, LinkedIn, portfolio/GitHub
  • Roles: title, company, location, dates, and every bullet/accomplishment mentioned anywhere (dedupe across documents, but keep distinct variants of the same accomplishment, since they become bullet variants later)
  • Metrics: every number attached to an accomplishment, with where it came from
  • Skills: tools, methods, languages
  • Education and certifications: schools, degrees, dates, honors; courses and certificates with issuer and year
  • Vocabulary: internal team names, product codenames, and acronyms the materials use, which become the start of GLOSSARY.md
  • Narrative material: anything useful for cover letters, such as career motivations, background stories, and notable non-work experience

Present this draft back to the user as a compact summary so they can see what you found before the interview starts.

Step 3 — The interview

This is the heart of the setup. Ask in batches of 3–5 questions, wait for answers, then continue. Do not fire 30 questions at once. Cover all of the following areas.

A. Contact and header

  • Confirm the exact contact line (city/state to show, email, phone, links). The city they show is a choice: remote-friendly candidates often show a metro area ("Greater Seattle Area, WA") rather than a suburb, to stay credible for both remote and local postings. Whatever they pick becomes canonical.
  • Ask how role locations should render. The bank always records the truth (a remote role stays Remote), but some users render every role as their target metro on the page, since a page that mixes "Remote" with a city reads as a candidate in neither. Their answer is a setup choice for preferences.md.

B. Per role (walk newest to oldest)

  • Is this your most current position? What's your exact, official title, the one HR would verify? Explain why this matters: an inflated title that fails a background or reference check reads as misrepresentation.
  • More than one title? A promotion, a dual title (Title A | Title B), or a title used day to day that differs from the one on the employment record. Record each title that is real, which one a reference check would return (the default), and which kinds of postings call for the other. Only one renders at a time, never a merged hybrid.
  • Which titles have they never held at this company (Senior, Lead, Director, Head, VP, Principal, Staff)? These go in the lock so JD language can't pull them in.
  • Exact start and end months. Only ONE role may end in "Present".
  • "Add as many bullets as possible for this role." Push for more than what's on the resume: projects, launches, fires put out, things they're proud of that never made the page. The bank should hold 2–3× more bullets than any single resume will use.
  • Responsibilities, not just events. For each accomplishment, ask what standing responsibility it is evidence of. A bullet the tailoring skill can use is a responsibility exemplified by evidence. Also ask what any competent peer in that seat would have done too; those baseline duties become scope inside a bullet, not achievements of their own.
  • For each metric: "Could you defend this number in an interview?" If yes, it goes in the closed metric set. If it's a guess, either soften it to defensible language or drop it. Then pin down three things:
    • What kind of number it is: a measured result, a target, an average, or a ceiling. Keep the approximator or ceiling word ("about", "roughly", "up to", "8+") as part of the metric. An average must never be rewritten as a fixed window or SLA, and a target must never be presented as a result.
    • Where it attaches: which bullet or bullets may carry it. Figures attached to specific bullets are what keeps a number on the work it measures.
    • Blur guards: if two figures could be confused for each other on one page (two similar counts from different roles), record that they must not appear together.
  • Ownership locks. For each major accomplishment, ask whether they led it, partnered on it, or contributed to a team effort. Write down the highest verb they can defend, and what is explicitly not claimable (e.g., "built the dashboard; did not design the data model"). Flag anything ambiguous with [confirm ownership] so tailoring never claims sole credit for team work.
  • Under-recording sweep. Ask directly about the surfaces resumes routinely leave out and JDs routinely ask for. Older roles need this most, since they are usually the thinnest part of a bank:
    • client- or customer-facing work, and who owned the relationship;
    • commercial scope: accounts, quotas, revenue, contracts;
    • people leadership: hiring, interviewing, coaching, team size;
    • who they worked with, by name of group (operations, finance, legal, sales, engineering), not "cross-functional";
    • compliance, security, or regulatory controls the work touched;
    • distributed or cross-time-zone coordination;
    • cost, usage, or spend monitoring;
    • testing, evaluation, or quality practices;
    • the everyday tools and platforms in use, including ones they only configured or maintained.
  • Terms with two senses. When the user names a practice with a precise industry meaning, define it back and confirm which sense they mean before recording it. Common traps: "fine-tuning" (training model weights, or iterating on a prompt), "SOC 2" (an attestation report, never a certification), "A/B test" (randomized split, or a side-by-side comparison), "RAG" (building retrieval, or prompting over retrieved context). A confirmed "no" is worth recording too: it stops being re-asked and can be stated plainly.

C. Skills and gaps

  • "What other skills do you have that aren't on the resume?" This covers tools, methods, domain knowledge, and languages. Apply the bar: only skills they've demonstrably used go in, and aspirational skills stay out. Record skills as bare tool or technology names. Delivery, coordination, and stakeholder claims are bullets, not skills.

D. Targeting

  • What roles/titles are you applying for? This defines the 2–4 profiles (e.g., "growth", "platform", "program"), each with its own headline lead and skill ordering.
  • For each profile: are there bullet types a resume for it must carry? A program-management resume, for example, usually needs leadership, cross-stakeholder, and tradeoff bullets in the recent role, or it reads as an engineer applying sideways.
  • Is there work that must survive every trim, whatever the posting (usually the most current or most differentiating work)? It goes in preferences.md as protected work.
  • Any constraints: location/remote, seniority range, industries to avoid.

E. Setup choices (recorded in preferences.md)

  • Summary paragraph, on or off? On, the resume carries a 2–3 sentence summary under the headline. Off, it goes straight from the bold headline to Experience and the space goes to bullets. Many users find the bullets make the argument better.
  • Cover letter every run, or only on request? If most of their applications are forms that take no letter, on request saves wasted work.
  • Location rendering (from A), and whether a second page is ever acceptable.

F. Cover letter material

  • Why are you job hunting, and what do you want next? Their honest motivation thread is what makes cover letters non-generic.
  • 1–2 background stories worth having on tap (career change, unusual path, relevant side work).
  • From what year should years of experience be counted, and for what kind of work? Letters recompute the figure from today's date and attach it to the work, never to a title.

Step 4 — Generate the bank

Write _bank.md in the working directory, using templates/_bank.template.md as the structural skeleton. Critical sections:

  • CANONICAL FACTS: the anti-hallucination lock. Generated from the interview answers, it fixes forever:
    • The exact contact/header line, verbatim
    • Every role title, exactly as confirmed, with any alternate titles and when each applies, and the titles never held (with a note: "never inflate, no Director/Head/VP/Lead unless actually held")
    • Every role's exact date string, and which single role ends in "Present"
    • The closed metric set: an explicit list of every defensible number, each with its approximator word, its kind, its attachment rule, and any blur guard. The rule written into the file: no number may ever appear on a resume that is not in this list.
  • OWNERSHIP LOCKS: one entry per claim area, each with what is claimable (and at which verb), what is explicitly not claimable, and what is still unrecorded and worth asking about.
  • PROFILES / SUMMARIES / SKILLS / EXPERIENCE_ROLES / EDUCATION / CERTIFICATIONS / BULLETS / COVER_LETTER_EXTRAS: filled from the parse and the interview. Bullets are keyed with short IDs, written as responsibility plus evidence, with ownership flags preserved and variants grouped. Role locations are the true record.
  • Typst templates: copy them verbatim from the template file; do not modify formatting values.

Write only what the user confirmed. Where you are unsure, record the item as an open question, never as a boundary. A cautious guess written into a lock ("X is not claimable") reads exactly like a confirmed limit on every later run, and it manufactures a gap.

Also:

  • Copy templates/preferences.template.md to preferences.md if it doesn't already exist, and fill in its Setup choices from interview section E, along with any protected work from D.
  • Copy templates/GLOSSARY.template.md to GLOSSARY.md if it doesn't already exist, and seed it with the vocabulary from Step 2: internal names and their plain-English synonyms, plus any two-sense terms confirmed in the interview.

Then show the user the CANONICAL FACTS section and ask them to confirm every line. Fix anything they correct before finishing.

Step 5 — Kick off calibration

Tell the user, in roughly these words:

Your bank is ready, but the tailoring isn't calibrated to your taste yet. Pick the sample job description in samples/ closest to your target role (or use a real posting) and run:

/tailor-resume samples/<file>

Then open the PDF and tell me everything you'd change: length, section order, tone, formatting, wording style. I'll save your feedback to preferences.md, and every future run will apply it automatically. If a rule ever gets broken twice, it moves into the checklist at the top of that file, which every run works through before compiling.

Also read the gap list at the end of each run. Anything on it that you actually did is a fact the bank is missing. Tell me and I'll record it, so it stops showing up as a gap.

Guardrails

  • Never invent anything to fill a gap. A thin, honest bank beats a padded one, and empty sections are fine. But ask before accepting thinness: a silent role is more often under-recorded than empty.
  • Never put a number in the closed metric set that the user didn't confirm as defensible.
  • Never write a speculative boundary into a lock. When evidence is ambiguous, ask rather than narrowing.
  • Never write the user's materials or bank contents anywhere outside the working directory.
  • If the user runs this command when _bank.md already exists, ask whether they want a full rebuild or targeted updates (a new role, new bullets, a corrected lock). For targeted updates, edit the existing bank instead of regenerating it, and when a new answer contradicts an existing lock, amend the lock rather than adding a second entry beside it.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.