Career Craft for Claude Code

A coaching-first Claude Code project for helping job seekers understand, practise, and improve CV/resume writing, as well as tailoring applications, auditing content, redesigning sections, and producing print-ready A4 layouts.

The project does not normally jump straight to a finished rewrite. It diagnoses the writing, explains the principle, asks focused evidence questions, gives the user a structure to try, and then reviews and refines the user's draft.

Quick start

1. Unzip the project and open a terminal in the resume-refiner folder.
2. Put source files and job descriptions anywhere inside the project (the suggested locations are workspace/input/ and workspace/job-descriptions/).
3. Start Claude Code with claude.
4. Type / to see the project skills, then invoke one of the commands below.

Claude Code now recommends project skills in .claude/skills/. They still appear and run as slash commands. Existing .claude/commands/ remains compatible, but skills can bundle supporting instructions and references.

Commands

Starting from a PDF

Most people arrive with a PDF rather than Markdown. /import-resume converts it into an editable source file that every other command can read:

/import-resume @workspace/input/jane-cv.pdf

or type

/import-resume 

then drag the file in.

The result is written to workspace/resumes/jane-cv.md. The import preserves the candidate's exact wording — it converts format, not prose — and flags anything it could not read cleanly with [VERIFY: ...] markers for the user to check against the original.

Claude reads the PDF itself. There is nothing to install — no poppler, no Python, no conversion tool — and scanned or image-only PDFs import the same way as ones with a text layer.

Because the text is transcribed rather than copied, proofread the [VERIFY: ...] markers against the original before using the file. Dates, metrics, and contact details are the places worth a second look.

Example:

/refine-personal-statement @workspace/input/jane-resume.md
Target role: Senior Frontend Developer
Tone: confident, concise

The first response will normally explain what is working, what is unclear, ask a few questions, and give a drafting scaffold. After the user replies with evidence or a new attempt, Claude provides feedback and a grounded refined version.

Experience practice:

/coach-experience-writing
Responsible for onboarding new team members.

The coach will ask what the user actually created, the scale, who was affected, and what changed. It will not silently turn the sentence into an invented achievement.

/tailor-to-jd @workspace/input/jane-resume.md @workspace/job-descriptions/frontend.md

/design-resume-layout @workspace/output/jane-refined.md
Style: ATS clean
Maximum length: 2 pages

/render-resume @workspace/output/jane-refined.md
Layout: professional accent
Accent colour: #D6A84B

Arguments may also be pasted directly after the command. When no file or text is provided, the command asks for the minimum information required.

Recommended learning workflow

1. /import-resume if the starting point is a PDF
2. /audit-resume
3. /coach-experience-writing on one or two representative bullets
4. /refine-personal-statement
5. /tailor-to-jd
6. /refine-full-resume
7. /restructure-resume
8. /design-resume-layout
9. /render-resume
10. Proofread the exported PDF before submission

For an urgent application, the user can explicitly request direct rewrite or text only. The no-fabrication rules still apply.

Layout modes

• ATS Clean — single column, minimal styling, safest for automated parsing.
• Professional Accent — single column with restrained colour and stronger hierarchy; suitable for most technology, finance, product, and corporate roles.
• Modern Two Column — compact sidebar plus main experience column; best when a human-readable version is appropriate. Create a separate ATS version when unsure.

Output principles

• No fabricated facts or numbers.
• No generic one-size-fits-all profile.
• No keyword stuffing.
• No automatic conversion of assisted or participated into unsupported leadership claims.
• No assumption that every bullet needs a numerical metric.
• No skill bars, rating dots, or charts that ATS tools cannot interpret.
• Content comes before visual decoration.
• Markdown remains the editable source of truth; HTML/CSS is the presentation layer.

Project map

.claude/skills/       Slash-command workflows
rules/                Shared editorial and design rules
references/           Detailed checklists and layout recipes
templates/            Starter resume and HTML/CSS templates
examples/             Before-and-after examples
scripts/              Project validation
tests/                Structural and behavioural tests
workspace/resumes/    Imported Markdown resumes (git-ignored)

Testing

The commands are prompts, so testing splits into two tiers — see tests/README.md for the full explanation.

bash tests/run-tests.sh              # structural: fast, deterministic, free
bash tests/run-behaviour.sh --judge  # behavioural: runs the commands, grades them

tests/run-tests.sh checks skill frontmatter, that every referenced file exists, that the README commands table matches the skills on disk, and that candidate material stays git-ignored. Run it after editing the project; it wraps scripts/validate.sh, so there is no need to run that separately.

tests/run-behaviour.sh runs the coaching commands headlessly against tests/fixtures/poor-resume.md and grades the responses against the pass criteria in tests/fixtures/poor-resume-evaluation-guide.md — mainly to catch the tool regressing into a rewrite-first assistant. It costs tokens and takes about a minute per scenario, so it is not a per-commit check.