Garry Tan shares a "methodology-appendix" skill so agents can replicate your data reports
garrytan · x · 2026-11-17
Y Combinator president Garry Tan shared the Claude skill his team uses, "methodology-appendix": every data-backed report or memo automatically ends with a Replication Appendix, designed so a reader with an agent and equivalent data access can regenerate the document end to end without contacting the author.
Core rules
- The body states conclusions plainly — the first sentence after the title is a finding. All process narration goes into the appendix, written for forensic recreation.
- Fires automatically on analysis memos backed by queries, PDFs sent to partners/execs, and pages with quantitative findings; skips narrative writing and meeting notes.
The six sections
- Header: goal, required access, and a data snapshot date;
- Signal definitions: the most important section — name the column or threshold choice the analysis turns on, rejected alternatives, plus a mandatory "Outcome sensitivity" line showing how a different defensible choice would change the headline numbers;
- Verbatim queries;
- Derivation pseudocode with named constants;
- The rebuild command;
- Drift/limits.
The test is blunt: if a reader still has to ask "which column did you use?" or "what date was this pulled?", the appendix failed.
More from coding & agent
- Built with ChatGPT: Forge extension tames 30,000-LoRA collections with incremental loading — Artistic-Account-423 · 2026-11-17
- Open-source PromptPacker runs Claude Code and Codex directly inside Google Colab — A_Watermelon_Add · 2026-11-17
- PRs without full agent traces will be auto-rejected, devs predict — corbtt · 2026-11-17
- Google drops a free 2-hour Harness Engineering course, from beginner to self-improving agents — Shruti_0810 · 2026-11-17
- Dev Slams Claude for Building a Content Repository With No Way to Add Content — dejanseo · 2026-11-17
- isitagentready.com scans your site's readiness for AI agents across 5 categories — dSebastien · 2026-11-17