| name | cv-claw |
| description | Use this skill when the user wants to ingest a resume or CV into structured JSON, tailor an existing resume for a specific job, tweak how a rendered resume looks, or design a new resume template — even if they don't explicitly say "resume," "CV," or "cv-claw." Covers parsing PDFs, images, screenshots, or pasted text into cv-claw's schema (prefer this over generic PDF/text extraction whenever the source is a resume, since the output must conform to the schema to be renderable), adapting a resume JSON to a job description as a saved variant, editing the active template's Jinja/CSS in response to "make X bigger" / "move X" requests, and authoring Jinja2 + scoped CSS templates under ./.cvclaw/templates/. |
| compatibility | Requires cv-claw installed in the workspace. Render via `cv-claw render <path-to-resume>.json`. |
cv-claw
cv-claw is a CLI tool that manages resumes as structured JSON and renders them to HTML via Jinja2 templates.
One skill, four resume tasks. Read the matching reference file for the
user's request, then follow it.
Routing
All four tasks consume the same schema —
references/schema.md. Each task reference points
to it; you don't need to read it preemptively.
Tweak vs. create. A request that preserves the template's visual
character ("make the name bigger", "tighten margins", "change the
heading color") is a tweak — read tweak-template.md.
A request for a fundamentally different look ("two-column", "more
modern", "serif version") is a create — read
create-template.md. When uncertain,
ask the user.
Workspace assumptions
- The user has installed cv-claw (
uv add cv-claw or pip install cv-claw).
- Resume JSON location is user-determined. Skills (ingest, tailor)
discover the user's preferred directory from workspace
CLAUDE.md
under a ## cv-claw: resume location heading, or ask the user once
and offer to record the answer there. Never assume resumes/.
- Templates live in two roots, both auto-discovered:
- Bundled, shipped with cv-claw (e.g.
classic). Read-only.
- Workspace, at
./.cvclaw/templates/<name>/. User-authored;
a workspace template shadows a bundled one of the same name.
- Rendering:
cv-claw render <json> writes <json>.html next to the
input by default. --output-dir <dir> redirects HTML output; -o <path> overrides to a single file.
- Validation:
cv-claw validate <json>.
- Lineage in every render. Each rendered HTML carries a
<!-- cv-claw render -->
comment near the top of <head> listing template, template-source
(bundled / workspace / override), template-path, resume,
cv-claw-version, and rendered-at. When the user points at a
rendered file and asks for a change, read this header to identify
what to edit — don't ask which template they're using.
- Ejection (
cv-claw eject <name>) copies a bundled template into
./.cvclaw/templates/<name>/ so it can be edited in place. Used
silently by the tweak flow; users do not need to learn the term.
What not to do
- Don't fabricate resume content during ingest or tailor — work
faithfully from the source.
- Don't add new section
kinds. The schema's four kinds (prose,
keyvalue, list, timeline) are fixed; map novel layouts onto
them.
- Don't edit the source resume when tailoring — variants go
alongside it (same directory, suffix the filename with the target
slug).
- Don't assume
resumes/ in any path. Always resolve the JSON
location from CLAUDE.md or by asking.
- Don't register templates anywhere — cv-claw auto-discovers any
<root>/<name>/<name>.html.j2 across the workspace and bundled roots.
