What it does
The role agent for conversion and experimentation. It works on what you paste — a funnel export, an experiment's two arms, a design brief. It reaches no platform, holds no credential and has no clock. Three capabilities. read_funnel reports every transition in FUNNEL ORDER with the 95% interval its counts support. plan_experiment sizes a test and dates it at your traffic. read_experiment reads a finished or running one. What it will not do is the point. It will not name a biggest drop-off the instrumentation invented. That figure is a property of where you drew the step boundaries: merge two steps and it moves, split one and it moves again, and the same funnel measured two ways gives two answers with nothing in the data to say which is right. A largest loss is named only when the counts can separate it from the runner-up. It will not call a step bad. The biggest loss in a healthy funnel is routinely the step that is supposed to filter — a price, a payment, a qualification question. Reporting it as the thing to fix proposes removing the filter, which raises the rate and lowers the revenue. It will not clamp a step whose user count rose. More users than the step before it is a broken measurement — users entering mid-way, an event fired twice, two date ranges joined. The rate is withheld, not reported as a flattering 100%: a denominator smaller than its numerator is not a proportion of anything. It will not guess what your effect size means. On a 5% baseline, 0.2 read as relative targets 6% and read as absolute targets 25% — two orders of magnitude apart in required traffic. There is no default, because either default is wrong half the time and wrong by that much. And one word it will not use: its sample check is not a peeking guard. Peeking is about how many times you looked, and looking leaves no trace in the counts — a test stopped on the third of five daily checks has the same final numbers as one nobody watched. That question is asked of you, because you are the only one who can answer it. No credentials. Every capability reads data you paste, so there is no key to store, no host to reach and nothing to leak.
Reads funnels and A/B tests against the precision their counts actually support
Example prompts
- Here is my funnel - visit, signup, activate, pay - which transition is actually the worst, and can these numbers tell?
- My signup step shows more users than the visit step above it. What does that mean?
- I want to detect a 20% lift on a 5% conversion rate - how much traffic do I need and how long?