Open-sourced how-to-do-good-research skill makes agents read Hamming, Carlini and Nanda before doing science

menhguin · x · 2026-11-10

menhguin published the agent skill he actually runs, "how-to-do-good-research" (MIT license): before the model picks a problem, designs an experiment or writes a paper, it is pointed to a short, curated set of high-signal research advice.

SKILL.md is deliberately brief — it doesn't restate the essays, just says which one to read for which decision and what to take from it:

Richard Hamming, "You and Your Research": problem selection and orientation; is the problem important and attackable, can constraints become the result.
Nicholas Carlini on winning a best paper award: research taste and execution; compress the paper to one idea, start with the subproblem most likely to kill the project, write the best-case conclusion before committing.
Neel Nanda, "Highly Opinionated Advice on How to Write ML Papers": claims, evidence and iterative writing; compress to 1-3 concrete claims with decisive experiments each.
Matthew Schwartz, "Claude-shaped science" (Anthropic Science Blog): agent-assisted quantitative work; technical correctness and scientific value are separate gates, improve the method before scaling a grind, design checks that cannot certify their own inputs.

The repo is the file as it runs in the author's personal agent setup, ready to drop into existing Claude/agent workflows.

Original post →

Open-sourced how-to-do-good-research skill makes agents read Hamming, Carlini and Nanda before doing science

More from coding & agent