Tools
Small things I’ve built for researchers — playful, self-contained, no accounts.
Research Values — Would You Rather
A forced-choice game that surfaces your research values. Pick the fate you’d rather — and watch a picture of what you actually trade for take shape. Two modes: a scored questionnaire that maps your choices onto ten research trade-offs (curiosity vs real-world impact, rigor vs novelty, depth vs breadth, autonomy vs security, integrity vs expedience, and more) into a research-values fingerprint; and a lightweight icebreaker for lab retreats and seminars. There are no wrong answers — the reasoning is the point. Everything runs in your browser; nothing is sent anywhere.
Research Pitfalls — The Study Game
A text-only game for the first year of a PhD. Run a study: thirteen decisions across five phases (question, design, execution, analysis, reporting), two meters (weeks spent, rigor), and every wrong turn names the pitfall you walked into — leakage, untuned baselines, single-seed conclusions, HARKing, vacuous bounds, “it is easy to see that”, sunk cost — with what it smells like, what to do instead, and what to read. It ends in a postmortem with a reading list built from your mistakes. Train your taste gamifies the discriminative skill from Developing taste on two tracks, taste for developing ideas and taste for writing and evaluating: pick the stronger of two idea pitches or abstracts, grow a weak idea, triage a batch, spot the flaw in a plan, say how sure you are, and get scored on accuracy and calibration, plus a taste profile that places you on six spectra such as generalist–specialist and method-led–question-led. Machine learning (experimental and theory) is the main focus; a domain selector re-voices the traps for computational, lab, and social science. Scope an idea sizes up a project on novelty (ten named kinds, from a new method to connecting fields), impact, feasibility, killability and fit, and shows how three scorers rated the same idea. A field guide lists every pitfall with its sources, and every technical term has a hover tooltip. Everything runs in your browser; nothing is sent anywhere.
Research Projects — The Management Game
The sister game, for the year you first lead a project. Run a project: nine months, a supervisor with no time, a senior collaborator with many projects, a master’s student you are responsible for, and a deadline. The four people get a personality each run (a supervisor who wants to be asked or one who trusts you to decide, a student who asks early or hides being stuck), and when the sound move depends on who they are, the reveal says so. Every decision moves six hidden dials (alignment, trust, buffer, credit, energy, loop) and seeds the classic problems: misaligned expectations, silence, scope creep, late feedback, hero mode, the authorship fight, the dropped ball, burnout. Most can be prevented or repaired, and the debrief traces every problem that fired, and every one that did not, back to the decision behind it, then places you on five management spectra. Say it better drills the hard messages: the collaborator who went quiet, the authorship opener, the reminder, feedback on a late section, disagreeing with your PI. A field manual lists every problem with what it looks like, what causes it, how to prevent and repair it, and what to read, drawn from lab-culture discussions and the Path to PhD newsletter. Everything runs in your browser; nothing is sent anywhere.
Principles for (agentic) research
The talk, given at the 50th Machine Learning Summer School on 8 September 2026, as an interactive page, with the talk’s own drawings and the argument behind each principle, not only the headline. Five parts: why do research at all, how taste is trained, what an agent is (the same model, reached through six rungs of access and permission), how to run one as a supervisor, and an audit for the next task you think of delegating. Every agent practice is attached to the research principle it applies, and the tools appear as complete, copyable examples: a CLAUDE.md for a research repository, a hook and its script, a permissions rules block, a test file, a skill file, a sub-agent prompt that emulates a research group, a Makefile, a postmortem template. Interactive where it earns it: a ladder explorer, a keep/drop test for skills, a delegation calculator, a jagged-frontier sorter, and the audit. Your answers stay in your browser.