AI Cv Scanner

AI CV Scanner

Scans your full Claude Code conversation history to extract concrete evidence of AI usage, then drafts answers to user-provided questions.

Input

The user provides questions to answer — either:

• Pasted directly in the prompt
• Referenced from a file (e.g. "answer questions in questionnaire.md")
• A general ask like "summarize my AI experience for a CV"

If no specific questions are given, produce a general AI experience summary covering: tools used, proficiency level, concrete project examples, API/integration experience, and biggest wins.

Phase 1: Gather data

Run these scripts from this skill's scripts/ directory:

node <skill-dir>/scripts/build-index.mjs > /tmp/cc-session-index.json
node <skill-dir>/scripts/scan-setup.mjs > /tmp/cc-setup-scan.json
node <skill-dir>/scripts/extract-evidence.mjs < /tmp/cc-session-index.json > /tmp/cc-evidence.json

• build-index.mjs — extracts summaries + first prompts from all session indexes
• scan-setup.mjs — scans ~/.claude/ for skills, commands, MCP servers, hooks, CLAUDE.md files
• extract-evidence.mjs — pre-parses Claude Code and Codex sessions into clean user/assistant text, plus per-session toolCounts, messageCount, and model (hard signal for scope and sophistication). Each session also carries a subagents array (up to 8, name/description/agentType/model/toolCounts/assistantTexts) — on delegated sessions the implementation happened there, so the parent's own counts understate the work

Read all three output files. This is your map of the user's AI history and extracted evidence.

Phase 2: Spawn subagents (parallel, use model: sonnet)

Launch 3 sonnet subagents in parallel. Each gets:

1. The full session index from /tmp/cc-session-index.json
2. The setup scan from /tmp/cc-setup-scan.json
3. The parsed evidence from /tmp/cc-evidence.json
4. Their specific mission below

IMPORTANT: Tell each subagent to:

• Search the evidence JSON for relevant sessions by keyword matching on user and assistant text fields
• Use the session index for summary, first prompt, project, and date context
• Use per-session toolCounts as hard evidence of how AI was used — Skill/Agent/mcp__* counts signal advanced orchestration (custom skills, subagent fan-out, MCP); high Edit/Write/Bash counts signal real multi-file building, not just chat. Read the subagents entries alongside the parent: their toolCounts are where a delegated session's building shows up
• Use messageCount to rank the highest-effort sessions (the "very high message count" sessions worth deep-diving)
• Cross-reference setup scan findings when the mission involves tools or configuration
• Do NOT open raw JSONL files; the evidence file already contains the extracted signal
• Return structured findings, not raw dumps

Agent 1: "Project Impact Scanner"

Search for sessions where the user built substantial things with AI — entire features, multi-file refactors, complex integrations, ambitious tasks.

For each find, extract: project name, what was built, how AI helped, estimated scope/impact. Look for moments the user expressed surprise at speed/quality, sessions with very high message counts, parallel subagent workflows.

Agent 2: "Advanced Usage Scanner"

Search for sessions involving: custom skills, MCP setup, hooks, CLAUDE.md authoring, prompt engineering, workflow automation, subagent orchestration, design systems.

Cross-reference with the setup scan — for each tool/technique found in config, find a conversation that shows it being USED or CREATED.

Categorize findings by capability level: basic chat, code generation, custom prompts, skills, context rules, MCP, workflow automation. Note gaps honestly.

Agent 3: "API & Integration Scanner"

Search for sessions mentioning: API, LLM, chatbot, anthropic, openai, sdk, system prompt, RAG, embedding, function calling, tool use, webhook, automation.

Also check repos for LLM API code: grep -r "anthropic\|openai\|llm\|chatbot" ~/git/*/src/ --include="*.ts" --include="*.py" -l

For each find: what was integrated, how (SDK, API, RAG), what the use case was.

Phase 3: Synthesize

After all 3 agents return, combine their findings to draft answers to the user's questions. If the user provided a structured questionnaire, match findings to each question.

Output format

Present each answer as a draft with cited evidence (which project, which session). Flag sections where evidence is thin. After presenting drafts, ask the user which sections need adjustment, more evidence, or a different tone.

Language

Match the user's input language. If the questions are in Czech, answer in Czech. If English, answer in English.