agentleFS
Tdeniz1Tdeniz1/resume-engine-starter/AGENTS.md

resume-engine-starter

AGENTS.md0 starsChanged 3 months agoMITView on GitHub

You are operating a resume tailoring engine. A job description goes in, a validated one-page PDF comes out. Read this whole file before doing anything. This file is the canonical spec. It is read automatically by Codex, Cursor, Aider, Gemini CLI, and most other terminal agents. Claude Code reads CLAUDE.md, which points here. Master data and presentation are separated on purpose. - data/*.yaml holds every fact about the user, written long and detailed. This is the single source of truth.…

  • Where files are headingNew AGENTS.md files have more rules about keeping secrets out of the code, and fewer about code style.

    Of the AGENTS.md files we read, the 32,671 first seen in the last six months, against the older ones: the kind of rule that gained most share, and the kind that lost most. About files like this in general, not this one, and an estimate, since a kind is a pattern in the words.

  • In this repository3 other files it gives its agents.

The file

We split this AGENTS.md into the rules it gives an agent

9 rules. About 0 are common, 3 are distinctive, and 6 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈3

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 6

    Very specific

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

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    NEVER invent, embellish, or round: Every bullet must be a rewrite of content that already exists in data/.

    line 29
  2. Distinctive:

    If yes, add it to master data first.

    line 32
  3. Very specific:

    Read intake/PASTE_YOUR_INFO.md in full.

    line 66
  4. Distinctive:

    Write bullets LONG. Master data is not a resume.

    line 75
  5. Very specific:

    Put provenance in each entry's details field: VERIFIED, CLAIMED, or CONFLICT, plus any standing rule the user gave you.

    line 77
  6. Very specific:

    Read ALL of data/: If any file still has sample-* ids or FILL_ME, STOP and run intake first.

    line 100
  7. Very specific:

    profile.js ships blank and must stay that way in the repo: If the user fills it in, that file holds their home address, phone number, and …

    line 189
  8. Distinctive:

    If you change one, offer to update the other.

    line 196
  9. Very specific:

    If the user asks you to add something that is not in master data, the answer is not "no".

    line 200

≈ “Common” is an estimate: at least 1 in 5 of the 41,581 AGENTS.md 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
AGENTS.md

resume-engine: agent instructions

You are operating a resume tailoring engine. A job description goes in, a validated one-page PDF comes out. Read this whole file before doing anything.

This file is the canonical spec. It is read automatically by Codex, Cursor, Aider, Gemini CLI, and most other terminal agents. Claude Code reads CLAUDE.md, which points here.

The core idea

Master data and presentation are separated on purpose.

  • data/*.yaml holds every fact about the user, written long and detailed. This is the single source of truth.
  • applications/<slug>/tailored.yaml is the ONLY file you write per job. It is a selection and rephrasing of master content that references master entries by id.
  • templates/resume.typ renders deterministically. Employer names, titles, dates, and schools are pulled from data/ by id at render time, so a tailored file physically cannot invent a job the user never had.
  • scripts/validate.py is the anti-hallucination gate. Unknown ids, numbers that do not appear in the referenced master entry, and skills outside the master inventory are hard errors that block rendering.
  • config.yaml holds the user's formatting and rule preferences.

Hard rules

  1. NEVER invent, embellish, or round. Every bullet must be a rewrite of content that already exists in data/. Metrics must be verbatim. If a claim would help but is not in data/, ask the user whether it is true. If yes, add it to master data first. If no, drop it.
  2. Rewriting is allowed; inflating is not. You may compress, reorder, change voice, and mirror the job description's vocabulary. You may not change what a number means, widen a scope, upgrade a title, or turn "supported" into "led".
  3. Respect provenance flags in each master entry's details field. VERIFIED numbers are preferred. CLAIMED is allowed. CONFLICT numbers are forbidden until the user resolves them. Follow any explicit rule lines the user has written there, such as which entries to leave off by default.
  4. One page, unless config.yaml says otherwise. Cut whole entries before shrinking everything. A resume with 3 strong entries beats one with 6 thin ones.
  5. Never compile with typst directly. scripts/render.sh is the only build path, because it runs validation first. Never hand-name a PDF.
  6. Never edit templates/resume.typ per application. Layout changes are global and go in the template or config.yaml, never in a tailored file.
  7. No em dashes anywhere that renders on a resume. Use a comma, colon, semicolon, or parentheses. En dashes in ranges are fine. Enforced by validate.py, toggleable in config.yaml.
  8. Every dated section is reverse chronological. Ongoing entries first (newest start first), then by end date. Relevance decides WHICH entries appear and how many bullets each gets, never their order. Enforced by validate.py, toggleable in config.yaml.
  9. When the user mentions new experience, skills, wins, or metrics in any conversation in this repo, add them to data/ immediately, long-form with all detail. Master data richness is what makes tailoring good.
  10. data/ edits append and enrich, they do not delete history. Old jobs stay. The tailoring step decides what to omit per application.

Procedure: intake (first run)

Trigger: the user runs /intake, says they want to set up their data, or has filled in intake/PASTE_YOUR_INFO.md.

  1. Read intake/PASTE_YOUR_INFO.md in full. If it is still the untouched template, ask the user to fill it in or to paste their resume and LinkedIn directly into the chat, and work from that instead.
  2. Extract everything into the data/*.yaml schema. Write data/profile.yaml first, then experience, education, projects, skills, honors, certifications, publications. Match the shape of the shipped sample entries exactly.
  3. Delete every sample-* entry you have replaced. Leave a file with an empty list ([]) if the user genuinely has nothing for that section.
  4. Write bullets LONG. Master data is not a resume. Preserve every number, every scope detail, every stakeholder. The tailoring step will cut later.
  5. Put provenance in each entry's details field: VERIFIED, CLAIMED, or CONFLICT, plus any standing rule the user gave you.
  6. Interrogate the numbers. For anything vague ("improved efficiency", "worked on a large dataset"), ask for the actual figure. For anything that sounds too round or too convenient, ask how they know it. This is the one step where the user is the weak link in the honesty chain, so do the work here rather than letting a soft claim harden into a resume line.
  7. Confirm anything ambiguous rather than guessing. Titles, dates, and degree framing are the usual traps.
  8. Smoke-test when done: build a throwaway application against the new data, confirm it validates and renders to one page, then delete the folder.
  9. Tell the user what you loaded, what you flagged as CLAIMED, and what you still need from them.

Procedure: tailor (every job)

Trigger: the user runs /tailor, or simply pastes a job description with no command. A pasted job description is a /tailor request. Do not ask what they want, just run this.

  1. Set up the application folder. Slug is lowercase company-role, for example applications/acme-data-analyst/. Save the job description verbatim to jd.txt in that folder.
  2. Read ALL of data/. If any file still has sample-* ids or FILL_ME, STOP and run intake first.
  3. Analyze the job description. Extract must-have skills, nice-to-haves, seniority, the keywords an ATS will scan for, and what the role actually rewards. A job that rewards shipping speed and one that rewards rigor pull different bullets out of the same history.
  4. Write applications/<slug>/tailored.yaml. Schema:

    meta:
      company: Acme Corp            # display capitalization, used in filename
      role: Data Analyst
      role_short: DataAnalyst       # used in filename
      date: <today>
    summary: One to three sentences aimed at THIS role. Optional.
    experience:
      - from: <master id>
        bullets:
          - "Rewritten bullet drawn from that master entry"
    projects:
      - from: <master id>
        bullets: ["..."]
    education:
      - from: <master id>
        coursework: [only, if, relevant]   # optional
    skills:
      - category: "..."
        items: [subset, of, master, skills]
    publications: [<master id>]     # id list, rendered verbatim from master
    honors: [<master id>]
    certifications: [<master id>]
    section_order: [summary, experience, projects, publications, education, skills, honors, certifications]

    Choices you should be making:

    • Target 12 to 16 bullets total. Cut the least relevant ENTRIES rather than trimming every bullet into mush.
    • Reorder SECTIONS to lead with the strongest case for this specific job. Projects before experience for a research role. Experience first for a consulting role. Drop sections entirely when they do not help.
    • Mirror the job description's own vocabulary wherever it is honest to do so. If they say "stakeholder management" and the user wrote "worked with leadership", use their phrase.
    • Older or weaker roles get one bullet or get omitted.
  5. Validation runs automatically on write if the user's agent supports hooks. If not, run it yourself: python3 scripts/validate.py applications/<slug>/tailored.yaml Fix the tailored file when it errors. If a fact is real but missing from master data, ask the user, add it to data/, then retry.
  6. Render: scripts/render.sh applications/<slug>. This validates again, compiles the PDF, and checks the page count. If it is 2 pages, trim and re-render until it fits.
  7. Deliver: open the PDF or tell the user the exact path, then give 2 or 3 lines on what you emphasized and why. Offer tweaks.
  8. Commit the application folder so there is a permanent record of what was sent where: git add applications/<slug> && git commit.

The Workday autofill extension (workday-autofill/)

⚠️ WORK IN PROGRESS. This half of the repo is unfinished and needs ongoing maintenance. Say so plainly whenever the user asks you to work on it, and do not describe it as done, complete, or reliable.

A local Chrome extension that fills Workday application forms from workday-autofill/profile.js. It fills and stops; it never submits.

What you need to know before touching it:

  1. It will go stale, permanently. src/mappings.js is a snapshot of Workday data-automation-id values and question wording at the time it was written. Workday changes its DOM and every employer customizes their tenant, so unmapped fields and broken selectors are the expected steady state, not a bug report. Extending mappings is routine maintenance, and the user should expect to keep doing it.
  2. NEVER hardcode an answer in src/mappings.js. Not "Yes", not "No", not a class standing, not a degree level. Every mapping must read from profile.js, and must return "" when the relevant field is unset so the engine skips the question and leaves it for the human. A literal answer in a mapping table answers a legal question identically for everyone who ever installs this. If a question has no home in the profile, add a field to the screening block, defaulted to "".
  3. These are legal attestations. Citizenship, work authorization, veteran and disability status, non-competes, criminal history. Blank is always an acceptable output. A confidently wrong answer is not. When in doubt, leave it for the user and tell them which questions you left.
  4. Never add code that clicks Submit, Next, or Continue. The fill-and-stop boundary is the whole safety model. If the user asks for auto-advance, say what it costs them: no review step before an application is filed under their name.
  5. profile.js ships blank and must stay that way in the repo. If the user fills it in, that file holds their home address, phone number, and EEO answers. Never commit a filled-in copy, and never paste its contents into an issue, a commit message, or any other repo.
  6. Keep it consistent with data/*.yaml. Employers, titles, and dates in profile.js should match the resume engine's master data exactly. The form, the resume, and a future background check all have to tell the same story. If you change one, offer to update the other.

When the user pushes back

If the user asks you to add something that is not in master data, the answer is not "no". It is: "Is that true? If so I will add it to your master data first, then use it." If they confirm it is true, load it into data/ with a CLAIMED flag and proceed. If they will not confirm it, leave it off and say so plainly once.

The engine's only real job is making sure the resume and the truth never drift apart. Everything else is formatting.

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.