Garry Tan Shares Agent Skill for Reproducible Data Reports
The practical insight is that documentation standards now need two audiences at once: humans skim conclusions while their agents need executable recipes. Teams running multi-agent workflows can adopt this skill directly, since the appendix format makes any shared report verifiable and rebuildable without pinging the author. It pairs well with other Claude-style skills as a default for quantitative memos.
Garry Tan, CEO of Y Combinator, published a new agent skill called "methodology-appendix" designed for teams whose members send PDFs and markdown reports to each other and use AI agents to interpret them. The skill instructs an agent to append a Methodology / Replication Appendix to every data-backed report or memo by default, without asking.
The appendix is structured as a working recipe rather than a citation list, with six sections covering access requirements and a snapshot date, signal definitions and the column or threshold choice that mattered, verbatim queries, derivation pseudocode over named constants, the rebuild command, and drift/limits. The test is that a reader with an agent and equivalent data access can regenerate the document end to end from the appendix alone, without contacting the author. The body of the report states conclusions plainly with a finding in the first sentence, while all process detail moves to the appendix.
The skill fires automatically for analysis memos backed by database or API queries, PDFs sent to partners or executives, pages with quantitative findings, and anything with tables of counts, rates or percentiles, and is skipped for narrative writing, meeting notes and personal memos. It addresses a growing workflow problem: colleagues increasingly delegate document comprehension to their own agents, so reports need to be machine-replicable, not just human-readable.
I send around PDFs and markdown files to my coworkers so much and they use their agents to make sense of the world too, so I started using this skill methodology-appendix to make sure their agents could replicate what my agents found.
--- name: methodology-appendix description: "Every data-backed report or memo ends with a Methodology / Replication Appendix: the signal choice and its sensitivity, verbatim queries, derivation pseudocode over named constants, the rebuild command, and drift/limits. Appended by default, without asking." ---
# Methodology / Replication Appendix
## Rule
Any report, PDF or page whose conclusions rest on data queries ends with a **Replication Appendix**. It is a working recipe, not a citation list. The test:
> A reader with an agent and equivalent data access can regenerate the document > end to end from the appendix alone, without contacting the author.
If a reader would have to ask "which column did you use?", "what date was this pulled?" or "how do I rebuild the PDF?", the appendix failed.
## Body voice
The body states conclusions plainly. No process narration ("I was asked...", "on the first pass..."), no self-grading. The first sentence after the title contains a finding. Everything about how you got there goes in the appendix at the end, written for forensic recreation.
## When this fires
Always, automatically, for: - analysis memos backed by database or API queries - any PDF sent to partners, executives or external stakeholders - pages presenting quantitative findings - anything with tables of counts, rates or percentiles
Skip for narrative or qualitative writing, meeting notes, personal memos, short ephemera, and documents with no data spine. Do not ask permission.
## The six sections
### 1. Header
```markdown ## Replication Appendix
A reader with an agent and equivalent data access can regenerate every number and rebuild this document from this appendix alone.
**Access required:** <tool/CLI + auth, internal APIs> **Snapshot date:** <ISO date>. Live-data documents won't reproduce exactly after this date; see Drift. ```
### 2. Signal definitions and the choice that mattered
The most important section. Analyses usually turn on one column or threshold choice. Name it, name the alternatives, state the impact.
```markdown ### Signal definitions
| Field | What it is | Why chosen / rejected | |---|---|---| | `score_a` | Ensembled score | **Primary.** More predictive per <source>. | | `score_b` | Single-model score | Rejected as primary; kept for contrast. | | final_decision | Human decision object | Not queried. Known gap. |
**Outcome sensitivity:** switching from score_b to score_a moved the headline count from N1 to N2 and changed claim X. A replication that changes the signal column will not match these numbers. ```
Always include an **Outcome sensitivity** line. If a different defensible choice changes the conclusion, the reader must know before quoting it.
### 3. Verbatim queries
Every query that produced a number, copy-pasteable, in document order, labelled by the finding it supports. Never paraphrase SQL. Never write "similar to the above." Add schema notes a replicator would otherwise get wrong: enum meanings, non-obvious joins, columns that seem to exist but don't.
### 4. Derivation pseudocode
Queries give raw counts; the document has derived figures. Show the arithmetic, with assumptions as named constants a reader can swap.
``` ITEMS_PER_DAY = 8 # observed median, <period> YIELD_HAIRCUT = 0.10..0.15 # marginal cases convert below observed rate
gap = count(status = REJECTED and score >= threshold) extra_days = gap / ITEMS_PER_DAY added_wins = gap * YIELD_HAIRCUT ```
Different constants change the recommendation, not just the number, so they must be visible.
### 5. Rebuild command
Exact commands from source markdown to verified PDF, plus a visual check (e.g. `pdfinfo` page count, `pdftoppm` a couple of pages and look at them). Name the source of record: the PDF is a render, never the original.
### 6. Drift, limits and known gaps
- **Live data moves.** What was mid-flight at snapshot; what is closed and stable. - **Coverage gaps.** Fields only populated for part of the range, and how tables were restricted rather than shown with misleading denominators. - **What this cannot see.** Inputs to the real decision that the data lacks. - **Retractions.** Any claim withdrawn between versions, and why.
## Procedure
1. Log queries verbatim as you run them. Reconstructed SQL is how appendices go wrong. 2. Write the analysis. 3. Append the six sections before rendering. 4. Name the signal choice and its outcome sensitivity. 5. Render, check visually, then send. 6. On revision, update the appendix in the same commit as the finding. An appendix describing a superseded method is worse than none.
## Quality bar
- [ ] Could a reader re-run this without asking me anything? - [ ] Is every headline number traceable to a pasted query? - [ ] Is the signal/threshold choice named, with alternatives and impact? - [ ] Are derived figures arithmetic over named constants? - [ ] Is the rebuild command present and correct? - [ ] Are drift, coverage gaps and retractions stated?
## Anti-patterns
- **Citation theater:** table names instead of runnable queries. - **Reconstructed SQL:** retyped from memory, drifted from what ran. - **Silent signal choice:** one column presented as "the score" when a defensible alternative changes the answer. - **Buried assumptions:** constants hard-coded in prose. - **Stale appendix:** findings revised, appendix untouched. - **Asking first:** this is a default, not an offer.
This report is categorized under AITOP's "How-to" desk, focusing on US frontier AI labs and Silicon Valley technology leaders. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows. Key entities and technical topics under continuous tracking include Claude, prompt engineering, Claude Skills, data analytics.
Source attribution: curated from Garry Tan (@garrytan).