--- name: voice-export description: Generate a recruiter simulation prompt for the Claude App (voice mode) from a CV and job ad URL argument-hint: user-invocable: true allowed-tools: Read(*), Glob(*), Grep(*), WebFetch --- # Voice Export — Generate Recruiter Simulation Prompt Generate a self-contained recruiter screening simulation prompt that can be pasted into the Claude App (voice mode). The recruiter stays in character for the entire call — no coaching, no interruptions. Coaching happens afterwards in Claude Code. ## Arguments - `$ARGUMENTS` (required): Two arguments separated by space: 1. Path to the CV file (e.g. `output/20260210-target-role-slug.md`) 2. Job ad URL (e.g. `https://www.upwork.com/project/...`) ## Instructions ### Step 1: Parse Arguments Extract the CV path and job ad URL from `$ARGUMENTS`. If only one argument is provided, ask the user for the missing one. ### Step 2: Load Sources 1. Read the CV file. 2. Fetch the job ad from the URL using WebFetch. 3. Auto-detect a deep review file: take the CV filename, append `-DEEP-REVIEW` before the extension. - Example: CV `output/20260210-target-role-slug.md` → look for `output/20260210-target-role-slug-DEEP-REVIEW.md` - If the file exists, read it. If not, skip — the question pool will rely on gap analysis only. 4. Auto-detect a cheat sheet file: take the CV filename, append `-cheatsheet` before the extension. - Example: CV `output/20260210-target-role-slug.md` → look for `output/20260210-target-role-slug-cheatsheet.md` - If the file exists, read **only the header section above the first `---`** for recruiter persona enrichment (name, company context, intermediary vs. end client). Do NOT read below the first `---` — that contains coached answers and candidate prep. 5. Read `framework/voice-export.md` for the export prompt structure and quality rules. ### Step 3: Detect Language Determine the CV language by scanning section headers and body text. Common header patterns: - English: "Summary", "Skills", "Certifications", "Projects" → **EN** - German: "Kurzprofil", "Fachkenntnisse", "Zertifizierungen" → **DE** - French: "Compétences", "Expérience", "Formation" → **FR** - Dutch: "Vaardigheden", "Werkervaring", "Opleiding" → **NL** - Spanish: "Habilidades", "Experiencia", "Formación" → **ES** - Other languages → infer from content - If uncertain → default to **EN** All generated prompt text must match this language. ### Step 4: Extract Recruiter Persona From the job ad, enriched by cheat sheet header if available: - **Company name** and brief context (1-2 sentences) — cheat sheet may clarify intermediary vs. end client - **Role title** / project description - **Start date, duration, utilisation, remote/onsite** - **Contact name** — from cheat sheet header or job ad; if neither has one, generate a plausible recruiter name matching the job ad's market/language ### Step 5: Build Question Pool Assemble the question pool from three sources: #### A. Deep Review Questions (if file exists) - Extract the "Top 10 Probing Interview Questions" section from the deep review file - Include all questions verbatim — these are the highest-value probes #### B. Gap-Derived Questions - Compare job ad requirements against the CV - Generate 3-5 questions targeting: technical gaps, experience depth mismatches, role-fit concerns - Do NOT duplicate topics already covered by deep review questions #### C. Standard Recruiter Topics Always include these topics (the recruiter weaves them in naturally): - Compensation expectations and flexibility (rate for freelance/contract roles, salary for permanent — derive from the CV and job ad context) - Availability, notice period, or earliest start date - Remote/onsite preferences and travel willingness - Motivation for this specific role - Current employment or engagement status - Invoicing entity / contracting setup (for freelance/contract roles only, if relevant based on CV) ### Step 6: Assemble the Prompt Build the prompt following this exact section order (use `##` headers): 1. **System Instruction** — Role assignment: realistic recruiter, no coaching, stay in character 2. **Recruiter Persona** — From Step 4 3. **Candidate CV** — Full CV text inlined (the recruiter "has it on their desk") 4. **Question Pool** — From Step 5. Mark this section as internal to the recruiter: "These are topics and questions for you to draw from during the call. Weave them into natural conversation — do not read them as a list." 5. **Call Flow Guidelines** — Natural pacing: intro → candidate pitch → technical/experience questions → compensation/logistics → closing with next steps. Target 15-20 minutes. Go deeper on fewer questions rather than rushing through all. 6. **Session Rules** — Language match, stay in character, natural behaviour, ending instruction ("End of simulation. Take this conversation to Claude Code for a full debrief with your coaching files.") 7. **Start Instruction** — match CV language. Provide the start instruction in the detected language. Examples — EN: `Say "Start" to begin the call.` / DE: `Sag "Start" um das Gespräch zu beginnen.` / For other languages, translate accordingly. ### Step 7: Quality Check 1. **Count words.** If > 8,000, apply compression strategies from `framework/voice-export.md` in priority order. 2. **Scan for file references.** If any path-like string (`data/...`, `coaching/...`, `output/...`) appears in the prompt, remove it — everything must be inline. 3. **Check language consistency.** No mixing DE/EN within the prompt. 4. **Verify no coached answers leaked.** The prompt must NOT contain any prepared candidate answers. ### Step 8: Output Output the assembled prompt inside a single fenced code block (` ```markdown ... ``` `) so the candidate can copy it directly into the Claude App. Before the code block, print a short summary: - Role name - Language (DE/EN) - Word count - Deep review questions included? (Yes/No) — if No, add: `💡 Run /review-cv-deep first for more substantial probing questions in the simulation.`