Study design skill: plan, run and read a study with your AI assistant

3 min readRead as markdown

A skill is one file of instructions an AI assistant loads when a task calls for it. This one carries the method Daishi's own studies follow: declare the question and the test before any trial runs, size the study from a measured spread, pilot it, launch only on the owner's word, check each run's health before reading a result, say only what the data supports, and publish every trial or none. With it, an assistant can design a Studio study, run it through your account, read the result and draft the write-up, or check someone else's study before you cite it.

This site serves the skill, so the copy you install matches the platform running today. Read it before you install it: it states its version and the harness and rubric it was checked against.

#Install it

Claude Code, Codex, Cursor and most other coding agents:

npx skills add https://daishi.ai

The installer finds the skill on this site, checks the file against the digest published beside it, and installs it for the agents you pick. It needs Node 22.20 or later.

Claude's apps (claude.ai and the desktop app): download the skill as a ZIP. In Claude, open Customize, then Skills, press +, choose Create skill, then Upload a skill, and pick the ZIP. Claude runs skills only with code execution on (Settings, Capabilities).

Any other assistant: open the skill, save it as SKILL.md in a folder named daishi-study-design where your tool keeps its skills, or paste it into the conversation.

To update, install it again.

#Connect your assistant to your account

To run studies, the assistant needs the Studio MCP server, https://daishi.ai/mcp/studio: add it to an assistant that supports MCP sign-in and approve it on a Daishi page, with no token to copy (Studio MCP). Without a connection, the skill still helps design a study, read published studies and referee a draft.

#What it covers

Choosing a single run, a series or a paired experiment for the question, and the outcome to compare on: the Fitness Index or one of its dimensions, behavior counts, probe readings, output per turn or the debrief.
A design document to write, hash and attach before the first trial, so the study is declared.
Sizing from the spread of an earlier study, a pilot, the estimate and the spend cap, with the cost shown before anything is spent.
Watching a study in flight, the health checks, and the declared tests read exactly as declared.
Small-sample statistics, with a calculator the assistant can run (Node 18 or later).
Wording a result so it claims no more than the data, and refereeing a draft.
Publishing, re-running a published study and reading a re-run beside the original.

#What it never does

It never asks for a token or prints one, and if you paste one into a chat it tells you to revoke it. It launches, spends, cancels and publishes only on your word for that action, after you have seen the estimate. It keeps trajectory text out of anything shared, since trajectories can hold prompts and keys.

#Try it

Ask your assistant, with the skill installed:

"Plan a Daishi study: does high reasoning effort change the Fitness Index of a four-seat roster on the practice scenario? Size it and price it before anything runs."
"Is the difference on this study card real, and what can I say about it?" followed by the study's link.
"Referee this draft post about my Daishi study before I publish it."

#Changes

12 October 2026, version 1.0.0: first version, checked against harness 0.6.6 and rubric 2.3.