From a paragraph to a build plan
Most AI builds fail at the brief, not the code. The model fills silence with assumptions, and you find out which ones were wrong after it has already built on them. This is the process that removes the silence.
-
01
Write the brief
One paragraph, in your own words. No format to follow. The classifier reads it, decides whether you are building a booking site, a storefront, an internal tool or something else, and loads the matching question modules.
- Takes about a minute
- Attach a logo, brand guide or reference document if you have one
-
02
Answer the interview
Adaptive questions grouped into stages. Most are taps: pick your payment gateways, choose which roles exist, toggle whether you need a PWA. Free-text only where it genuinely matters, like your service list or tone of voice.
- Branching — answering "no payments" hides eleven follow-ups
- Skip anything; skipped items become tracked gaps, not silent omissions
-
03
Close the gaps
The specification view shows every decision made and every one still open, with a completeness score. Critical gaps are called out separately, because a missing M-Pesa callback URL costs you more than a missing favicon.
- Suggested defaults for anything you genuinely do not care about
- Nothing is invented without being marked as an assumption
-
04
Generate the phased prompt
The compiler writes a build specification split into phases: foundation, data layer, core features, integrations, admin, hardening, deployment. Each phase carries its own deliverables and acceptance criteria, so the AI finishes one before starting the next.
- Copy a phase at a time, or export the whole spec as Markdown
- Regenerate after edits — every version is kept
Why phases
A ten-thousand-word prompt asks a model to hold every decision in mind at once, and it will drop some. Phased prompts give it one bounded job at a time, with a written definition of done. You keep the thread going until the last phase passes its criteria, and the output is a project you can actually deploy.