apply — evidence-first, human-governed application operations

apply is a working system for turning one verified career corpus into job-specific documents, searching public job listings, submitting through heterogeneous ATS forms, and preserving exactly what was sent.

The goal is not to automate judgment or mass-submit generic résumés. The goal is to make every claim traceable, every target deliberate, every submission recoverable, and every irreversible browser action human-owned.

Repository status: public. This began as a private operations repository and ran that way through every application it records. In August 2026 its whole history — every commit, not a squashed snapshot — was rewritten to remove what could not be published, and the result is what you are reading.

So the record is real but not complete, and the gap is deliberate. Personal data, third-party personal data, received assignments, internal diagrams, and the submitted binaries are gone; the companies, the answers, the rejections, and the judgment errors are not. docs/HISTORY-REDACTION.md states what was removed and by what method, and PUBLICATION.md is the contract that governs it.

The loop

public job cards/posts
        │
        ▼
search + location evidence ──► ranked candidates ──► human chooses one role
                                                        │
verified fact corpus ──► role-specific résumé cut ◄─────┘
        │                         │
        └─────────────────────────┴──► ATS form + human final submit
                                           │
                                           ▼
                              immutable submitted files + ledger

The boundaries are as important as the arrows:

• Search does not log in or submit.
• A résumé cut may reorder and select facts, but does not invent them.
• Public adjacent work is not presented as proof of a private deployment.
• Agents may prepare a form; a human owns personal fields, consent, and the final Submit action.
• The launcher keeps changing, so each application stores a copy of the files actually uploaded.

Repository map

Lane	Path	Responsibility	SSOT
Discovery	.claude/skills/linkedin-jobs/	public guest search, detail enrichment, location evidence, explainable ranking	SKILL.md
Document launcher	resume/	one fact corpus → role-specific résumé cuts	resume/body.org
Submission operations	applications/	JD snapshot, form answers, uploaded-file snapshot, outcome ledger	applications/LEDGER.md
Dossier launcher	dossier/	competency statement, portfolio, and evidence package from one org corpus	dossier/AGENTS.md
Handoff	NEXT.md	the next concrete move, not a backlog or history archive	NEXT.md

The narrative documents — this file, PUBLICATION.md, docs/, the license files — are in English. The operating contracts the agents actually follow, the root AGENTS.md and one per lane, are in Korean, because that is the language the work was done in. Translating them would mean maintaining two versions of a contract, and a contract that drifts is worse than one you have to translate.

Multi-agent collaboration

This project separates roles instead of asking one agent to search, write, browse, and judge at once.

Role	Owns	Must hand off
Explorer	candidate collection, source text, location confidence, suggested résumé cut	URL, quoted constraints, unknowns, ranking reasons
Document steward	factual selection, role vocabulary, public evidence routes, reproducible builds	exact artifact path and verification result
Browser operator	ATS observation, field preparation, upload mechanics	unresolved personal fields and the final-submit gate
Meta reviewer	cross-lane contracts, contradictions, public/private boundary, next-session handoff	gaps and durable fixes, not another submission lane
Human	target choice, truthful personal values, consent, final submission	the final decision and any private value that must not be stored

This is role coordination, not an agent hierarchy. Parallel work succeeds only when each lane owns a different artifact and reports checkpoints rather than silently editing another lane's SSOT.

One morning as a systems test

On 2026-07-28 the workspace went from no general submission lane to an end-to-end operating loop:

• 418 deduplicated job cards collected across Korean/English queries;
• location evidence kept three-valued instead of treating “unknown” as “wrong”;
• six résumé cuts generated from one fact source;
• four actual applications submitted through two ATS families;
• uploaded artifacts copied into per-application snapshots;
• one potentially serious release error caught: a job-specific résumé was still configured as the LinkedIn default and could have leaked into unrelated applications.

The useful outcome is not the application count. It is that failures became contracts: default-file preflight, visible evidence links, browser-only personal values, per-ATS observations, and a ledger that can be checked against each application record.

Verification

# Search/ranking logic — network-free regression suite
.claude/skills/linkedin-jobs/test_ljobs.py
python3 -m py_compile \
  .claude/skills/linkedin-jobs/ljobs.py \
  .claude/skills/linkedin-jobs/test_ljobs.py

# Application ledger ↔ target record ↔ uploaded snapshot
applications/check.py

# Rebuild and inspect all résumé cuts
(cd resume && ./run.sh all)

# Publication gate — this repository's own history, not an export tree
./scripts/check-public.sh --repo .

# Application ledger plus the public-contract invariant
applications/check.py --public

linkedin-jobs uses an undocumented public LinkedIn guest endpoint. It is best-effort, rate-limited, uses no login cookies, and may break when LinkedIn changes markup. See its SKILL.md for the exact contract and limitations.

Privacy and publication

The private workspace intentionally preserves company names, application answers, and submitted artifacts. That is incompatible with publishing this object database as it stands. A .gitignore change or a file deletion does not erase old commits.

The published record is therefore the whole decision history, sanitized — not a single-commit allowlist export. This repository is that sanitized result, not a plan for one. Four rules produced it:

1. Rewrite every reachable commit, blob, path, and identity field with git-filter-repo, keeping commit count, topology, timestamps, and messages intact. Preservation is verified, not assumed.
2. Remove what cannot be published — personal data, third-party personal data, compensation figures, received assignments, raw third-party copies, submitted binaries — and record the categories and method of every removal, never the values.
3. Push the sanitized history into a new, private object database and re-verify it from a clean clone; never flip this private repository's own object database to public.
4. Run the publication gate and read every file manually before the visibility change.

See PUBLICATION.md. The final visibility change remains a human decision, and it is effectively irreversible: once public, third-party archives and caches hold independent copies.

Licensing

• Software, scripts, tests, build definitions, and configuration: MIT — LICENSE.
• Reusable method documentation and public-safe operating contracts: CC-BY-4.0 — LICENSES/CC-BY-4.0.txt.
• Personal résumé wording, application answers, and actual career records: no reuse license (LicenseRef-Personal-Record).
• Exact path mapping: LICENSING.md; reserved and third-party material: NOTICE.md.

The licenses do not grant privacy, publicity, trademark, or employment-confidentiality rights, and do not relicense third-party job postings, received assignments, or quoted company material.

Author

This project is maintained by Junghan Kim (김정한, GLG).

• Digital garden
• Homepage
• GitHub profile
• LinkedIn