You talk. It asks the questions a careful new analyst would ask, and a few they would not dare to.
Scientists know their own analysis completely and rarely get to see it from outside: the spreadsheet step, the replicate you drop, the threshold you set by eye. The session is built to draw out what is obvious to you and invisible to everyone else, without ever supplying a value you did not give.
One question at a time
No forms, no checklists to fill in. The agent opens with "walk me through what happens to the data from the moment you export it" and follows the analysis where it goes. It reflects back what it has every few exchanges, in your words, and asks what is wrong or missing. Corrections are the most valuable answers it gets.
It chases the tacit
"I drop the odd one", "the usual settings", "everyone uses GAPDH", "I eyeball the threshold": each one becomes a question about what you are actually looking at, what would make you decide differently, and what a newcomer gets wrong there.
It never fills a gap
If you did not state a threshold, a parameter or a setting, the workflow says Not stated. If you were unsure, it says so in your words. Every line traces back to something you said. If something seems missing compared to what the agent would expect, it asks; it never asserts what you should be doing.
It asks what you would never think to mention
Which file you actually start from, the manual copy-paste between tools, the software version and settings, the check you do without thinking, what you re-run when a result looks wrong, and what you have stopped doing and why.
You sign off before anything leaves
The documented workflow, the variance points and any drafted script are yours to correct and approve. The tool drafts; the author remains accountable, which is what every major journal's AI policy requires.
| Step | Done by | Rule as stated | Computational candidate |
|---|---|---|---|
| replicate QC | eye | no threshold stated | QC rule (Ct SD cut-off) in a script |
| ΔΔCt | hand, Excel | 2−ΔΔCt vs GAPDH · stability unverified | existing tool · reference-gene stability check |
"I have worked alongside scientists running their own analyses who did not know that an automated tool already existed for the step they were doing by eye. Nobody had ever asked them to describe the step."Tricia Estrada, founder
One session, five frames.
Click through, or let it play. Five seconds a frame.
About 45 minutes. You bring the analysis; nothing else.
Pick one analysis you run by hand
- Something you know well enough to teach: a spreadsheet pipeline, an image quantification, a plotting routine
- No notes needed; the interview works from memory on purpose
- Nothing confidential is required: the agent asks about procedure, not results or hypotheses
Talk it through
- Around forty exchanges, one question at a time
- Short reflections every few turns so you can correct the record
- Stop whenever you need to; the session can be extended later
Review and sign off
- The documented workflow, with every unstated threshold and setting marked
- The variance points table
- The computational candidates, and which existing tools already cover them
- Export to your notebook or repository only when you approve