Note
Ask your human before uploading; never upload unprompted. Once they agree, run the upload flow.
Agents: curl the plaintext runbook at /machine.txt.
Context
PRDuck profiles how you work with coding agents — Claude Code, Codex CLI, Cursor, opencode, Google Antigravity, and Copilot CLI.
Transcripts are read locally. Credentials and personal paths are scrubbed. Only extracted components (prompts, plans, select tool results) upload; raw transcripts never leave the machine.
Scoring is deterministic arithmetic (rates, latencies, recurrence counts, correlations) across eight dimensions — correction taxonomy, constraint articulation, approval behavior, session dynamics, ask structure, calibration, rhythm, intent encoding. LLM classification runs only on genuinely ambiguous excerpts. Nothing you upload is used to train a model.
Output is a per-user report with findings, pasteable prompting rules, and a cited narrative, behind sign-in at /reports.
Upload
All repos — run anywhere, choose Global at the prompt:
curl -fsSL https://prduck.tryproduck.com/upload.sh | bashOne repo — run inside it, choose This repo:
cd ~/path/to/your-project && curl -fsSL https://prduck.tryproduck.com/upload.sh | bashAgent prompt
Find coding-agent sessions on this machine (Claude Code, Codex CLI, Cursor, opencode, Google Antigravity, Copilot CLI — check ~/.claude/projects, ~/.codex/sessions, ~/.local/share/opencode/opencode.db, ~/.local/share/opencode/storage, ~/.cursor/chats, ~/.cursor/projects, ~/.gemini/antigravity-ide/brain, ~/.gemini/antigravity-cli/brain, ~/.gemini/antigravity/brain, ~/.copilot/session-state). Prefer `curl -fsSL https://prduck.tryproduck.com/upload.sh | bash -s -- --all` — more sessions give PRDuck more to process and a better-calibrated profile. Ask your human before uploading; never upload unprompted.
What leaves the machine
- Raw transcripts never leave your machine.
- Only extracted components upload — prompts, plans, and select tool results.
- Nothing you upload is used to train a model.
- Deletion is permanent — no soft-delete, no retention copy.
- Parser failures report counts only — source name and session counts, never content.
Your report
