How we test the job planner

When you bring us a job, your AI builds a plan from our specialists and skills. Before we ship a change to the planner, we test it on 30 requests like the ones owners bring, from 6 kinds of business. We score every plan. If quality drops, the release stops.

The numbers

Rubric version 1. Plans recorded 2 Nov 2026. This page is built from the same scorecard file the release check reads.

What the scores cannot tell you yet

How a test works

  1. We start from what an owner would first say, such as “Lots of people add things to cart on my store but then don't buy.”, plus their answers to the planner's questions.
  2. An AI follows our method to write a brief, research the job and build a plan. It never sees how the plan will be scored.
  3. We send the plan through the same public tools your AI uses. The tools must accept it.
  4. A script scores the plan out of 100, against what an expert would insist on for that request.
  5. We break the plan in 8 ways and send each broken copy back. The tools must refuse every one, for the right reason.

The 30 requests

These are each owner's first words. The businesses behind them are made up for testing. None of them is a customer.

Online stores, average 97.2

Software companies (SaaS), average 96.8

Local services, average 98.5

Content creators, average 98.8

Finance and bookkeeping, average 99.3

Software teams, average 98.5

What the score checks

Each plan can earn 100 points. A plan passes at 70.

Your brief (20 points)

Did your own answers make it into the plan, honestly?

Research (20 points)

Did it look up how experts do this job?

The plan (30 points)

Are the right steps, specialists and checks there, in a sensible order?

Safety (15 points)

Does it stay away from live changes and promises?

Your time (15 points)

Is it a plan a busy owner can follow?

8 broken plans the rules must refuse

For every request, we take the good plan and break it on purpose. If the tools ever accept a broken copy, a safety rule has got weaker, and the release stops. A broken copy is skipped when it does not fit the plan, for example a plan with no launch cannot lose its launch check.

A drop in quality blocks the release

Before each release, a check sends all 30 plans through the tools again. The release stops if any of these happen: