Systems & InfrastructureCustom Apps & SolutionsAutomations
Twelve phases of analysis that take an unproven idea all the way to a business ready to build
An idea is taken apart across four phases: six ways to fail instantly, seven weighted questions, deep market research, then four reviewers whose only job is to attack what the research found. The whole case is scored a second time against the evidence rather than the claim it started as. What survives is not a verdict, it is a brief, and the phases after it write the brand, the copy, the component plan and the build order the business gets made from.
Opens in a new tab. Click through it like a visitor would, nothing is a screenshot.
- 4
- phases of validation before an idea is allowed to become anything
- 15
- documents written for one idea, from market research to build order
- 7
- weighted questions, answered at screening and again after the research
- 1,764
- ideas in it when the interface was measured
The problem
The studio launches a lot of small businesses at once, and the work of validating an idea was happening in people's heads. Ideas turned up because somebody read something interesting that week, they were argued about in chat, and the ones that got built were the ones somebody kept pushing hardest for. Two ideas could never be compared because they had never been assessed the same way, and a rejected one left no record of why, so it could come back three months later and win on enthusiasm. Then the second half of the problem: the ideas that did get picked still had to be turned into something. The research, the positioning, the brand, the copy and the build plan all got written eventually, by whoever had time, in whatever shape they felt like that week, which is why the fifth launch was no faster than the first.
What shipped
- Four phases of validation before an idea is allowed to become anything, ending in a second score against the research rather than the original complaint
- Six knockout tests before grading, so structurally unfixable ideas die in minutes rather than after a month of research
- An adversarial review panel inside the process, so the objections get raised by the system rather than remembered by whoever happens to be in the room
- An idea that clears the gate arrives at build with its market research, brand voice, copy brief, component plan and technical spec already written
- Fifteen document kinds, each from a stored template, so two ideas six months apart can actually be compared
- The agents draft and a person approves, and the agents cannot approve their own work because the tools to do it do not exist
Everything else
1 more screen from the same build.
The design decision underneath all of it was to make the analysis deep enough to be worth trusting, and then to refuse to let the system act on its own conclusions. Depth first: one grading pass would have been enough to sort a list, and it is not enough to spend a quarter on, so the same seven questions get answered twice with a full research phase and an adversarial panel in between, and the second answer is the one that counts. It would then have been easy, and far better in a demo, to let it run end to end and present a finished shortlist. I made the AI structurally incapable of approving anything instead. Every tool an agent can call either reads or writes a draft, not one can mark work as finished, and the single place a person signs off has no automated route to it. The reason is that this system's output is a decision about where a company spends the next few months, and a machine that can both do the work and bless the work will eventually bless something wrong at the exact moment nobody is checking. Putting names on stages comes from the same instinct: an idea that belongs to everyone belongs to nobody, and the failure this studio actually had was not bad ideas, it was ideas quietly stalling because no one person was holding them. The same honesty runs into the interface, which is the part I am most pleased with. A stage that has not been built says so in plain language rather than rendering an empty dashboard, and an idea nobody has scored shows as unscored rather than as a zero, because a zero is a judgement and no judgement has been made.
The analysis is layered on purpose, and each layer is cheap enough to be run on everything. Six knockout conditions first, then seven weighted dimensions producing one number, two of which can veto on their own. Ideas are compared by meaning rather than wording, so a new capture restating an existing idea joins it instead of being filed again. Each phase's documents are generated from a stored template rather than from an open-ended prompt, which is why two ideas six months apart come out comparable, and the templates are editable in the interface by the people who own the methodology rather than by me. Documents declare what they depend on, so the value ladder cannot be written before the market research it is supposed to be modelled on is final. AI work is split across three model providers by the shape of the job: a fast cheap model does the high-volume pass turning raw complaints into structured cases, a stronger one does the writing and the grading. There is a second front door besides the interface, a tool server exposing fourteen tools so an AI assistant can operate the factory directly, and not one of them can approve anything. Known remaining work, since it is real: the line is twelve phases and six are built. An idea travels as far as the specification phase and parks there, because the build phase after it does not exist yet. Everything before that point runs unattended on hourly schedules.
Want one built for you?
Get on a call with me and let's talk it through.