mcp-overleaf-server: Skill for Claude Code

Multi-job entry

Replaces hand-writing a jobs JSON. Ask how many jobs, render that many entry cards as a form in the chat, then run the existing batch pipeline on what comes back.

$ARGUMENTS may already contain the job count (e.g. /jobs 4).

Step 0 — infer before asking

Skip any question you can already answer:

• A number in $ARGUMENTS or in the user's message ("I have 3 jobs") → that's the count, skip step 1.
• Job postings already pasted in this conversation, attached as files, or referenced by a readable local path/URL → skip straight to step 3. Read the source first; do not make the user paste it again.
• job.json in the repo root filled in with real values → offer to use it instead of the form.

Never ask for something the conversation already told you.

Step 1 — how many jobs

Call mcp__visualize__read_me with modules: ["elicitation"] first — it carries the canonical form chrome and the current design rules. Then render form 1 from forms.md with mcp__visualize__show_widget.

It asks three things at once: job count (1–30), résumé vs CV, and cover letters. Stop and wait for the reply — it arrives as the user's next message.

Step 2 — the job cards

Render form 2 from forms.md, repeating the job block exactly N times with the index incremented. Only include the per-job cover-letter pills when step 1 answered "let me pick per job".

Cap at 30 (MAX_PLAN_JOBS in src/core/batch.ts). If more are wanted, run the first 30 and say so. Above ~8 jobs the form gets long — render it anyway rather than splitting it; the batch itself is chunked later.

Step 3 — parse the reply

The submitted answers arrive as one line of Label: value pairs, with any value over 200 characters replaced by (N chars — see below) and repeated verbatim under a --- Full content --- fold. Read the fold — that's where the job descriptions are.

Build a list of { company, position, jobDescription, jobUrl, questions }:

• Job N posting → jobDescription. Required. Drop any job whose posting is blank and say which.
• Job N company / Job N role → if blank, read them out of the posting text. They are almost always in the first lines. Say what you inferred so it can be corrected.
• Job N link → jobUrl, omit if blank.
• Job N questions → split on newlines into questions, omit if blank.