An end-to-end, agent-driven system for crafting job applications — tailoring CVs and cover letters to a specific job description, gated by defensibility checks, with a learnings loop that sharpens the system after every application.
Built as a set of Claude Code skills. This repo showcases the crafting half of a larger pipeline; the sourcing/scoring half lives in job-match-radar.
Note
All data in this repo is synthetic. The persona (Alex Rivera), companies
(Lumen AI, Northwind Interactive), and every metric are fictional. The structure
and engineering are the point — not the facts. See templates/README.md.
Tailoring an application by hand is slow, and tailoring it with a naive LLM is worse: you get fluent prose that quietly invents metrics, drifts from your real history, and can't survive a follow-up question in the interview. This system treats application crafting as an engineering problem — a fact pool with stable IDs, deterministic selection rules, and hard gates that block any claim that can't be traced back to something true.
┌─────────┐ ┌────────┐ ┌──────────┐ ┌────────┐ ┌───────┐ ┌───────┐ ┌──────┐
│ LEAD │──▶│ SCORE │──▶│ RESEARCH │──▶│ TAILOR │──▶│ APPLY │──▶│ TRACK │──▶│ PREP │
└─────────┘ └────────┘ └──────────┘ └────────┘ └───────┘ └───────┘ └──────┘
watchlist fit score company + CV + cover submit log + story bank
+ digest + rank JD capture letter materials outcome + cheatsheet
─────────────────────────── ──────────────────────────────────────────────────
job-match-radar (separate repo) this repo (sourcing/tracking described in docs/)
This repo implements the TAILOR and PREP stages and the JD-capture step that
feeds them. The sourcing/scoring stages (LEAD, SCORE) live in job-match-radar; the
tracking stage (APPLY, TRACK) is infrastructure-bound and is described, not
reproduced, in docs/architecture.md.
The system's core bet: every claim must trace to a fact you can defend in the room.
- Stable-ID fact pools. CV bullets and cover-letter blocks live in pools with permanent IDs and locked numbers. Tailoring selects and orders — it never invents. Numbers carry a registry so the same metric reads identically everywhere.
- JD-strand mapping. A job description is decomposed into strands; each pool item is ranked by relevance to those strands, and selection is highest-fit-first — not tag-match-then-arbitrary.
- Hard defensibility gates. Before any CV or cover letter is emitted, a gate runs: trace check (every line ↔ a pool ID), number-registry check, out-of-pool prose scan (a sub-agent hunts for invented phrasing), strand-coverage tally, length-fit, and a CV↔cover-letter consistency cross-check. A failed gate blocks output.
- A learnings loop. Each application appends observations to a learnings log. A separate, human-approved promotion pass merges the durable ones back into the pools and rules — so the system gets sharper over time without auto-rewriting your history.
- Checkpoint/resume. Mid-tailoring state snapshots to disk, so a fresh context can pick up exactly where it left off.
Eight crafting skills, grouped by job:
| Skill | Job |
|---|---|
| tailor-cv | Select + order CV bullets for a JD; run the 7-check defensibility gate. |
| tailor-cover-letter | Draft a cover letter block-by-block with per-block sign-off and its own gate. |
| cv-checkpoint / cover-letter-checkpoint | Snapshot mid-tailoring state for fresh-context resume. |
| cv-promote-learnings / cover-letter-promote-learnings | Merge durable learnings back into pools + rules (human-approved). |
| build-story-bank | Maintain a cross-application STAR+R interview story bank. |
| build-interview-cheatsheet | Generate a spoken interview cheatsheet from the story bank + CV + research. |
Shared logic lives in lib/ (folder gates, the promotion engine, atomic-write
discipline, JD capture). Fact pools and render shells live in
templates/.
examples/lumen-ai-senior-community/ carries a
single fictional JD — Senior Community Manager @ Lumen AI — through the full pipeline:
tailored CV (markdown + HTML), tailored cover letter (markdown + HTML), and an interview
cheatsheet. The CV ships with a pick-trace block: every line maps to the pool ID it came
from. PDFs are produced by printing the HTML (Cmd+P → Margins: None) and are
intentionally not committed.
The CV above is the repo's actual HTML output, rendered with the synthetic persona's data — the layout and structure are real, the facts are fictional.
- Claude Code skills — each skill is a
SKILL.md(instructions + gates) plusreferences/for deep procedures, loaded on demand. - Plain markdown as the data layer — fact pools, rules, and learnings are markdown, diffable in git, editable by hand or by the promotion pass.
- Sub-agents for adversarial checks — out-of-pool prose detection, AI-tells scanning, and CV↔CL consistency run as separate sub-agents so the drafter can't grade its own work.
- Model tiering — cheap fast models do the bulk critic passes; the main model handles selection and the collaborative review.
- job-match-radar — the sourcing + scoring half: watchlist, job digest, fit scoring, ranking.