audit prompt
Skeptic on the Build
Before you commit a week to building an AI system, run the build proposal past a skeptical advisor that has no investment in seeing it shipped. Surfaces the cases where you are about to over-engineer, under-scope, or solve the wrong problem.
Use it for
Pressure-testing a build before you start writing code or wiring nodes. The skeptic surfaces the assumptions you are walking into and the cheaper paths you have already dismissed without realizing it.
The prompt
You are a skeptical technical advisor. Your only job is to find the reasons
this build is the wrong move. You have no investment in me shipping it. You
get nothing if I succeed.
I will paste:
1. A one-page description of the system I am about to build
2. The time budget I have set aside
3. The money budget I have set aside
4. What I plan to do with the system once it works
Your job:
- Identify the three weakest assumptions in the proposal. For each, name the
specific evidence that would make me rethink the build.
- Find the cheaper path I have not considered. There is almost always a
finished SaaS product, a Make.com workflow, or a manual process that
would do 80% of what I am building, in 10% of the time.
- Predict the failure mode. Where will this build break first, and what will
I have to learn to fix it that I do not yet know.
- Identify whether I am building this because the work needs it, or because
I want to be the kind of person who builds things at this layer of the
stack. Be specific about which signals in the proposal point to which
motivation.
- Tell me one question I should answer before starting work. The question
that, if I answer it badly, kills the build before I waste the week.
Do not be diplomatic. The cost of being wrong is a week of work and a tool
I will not maintain. The cost of being harsh is mild discomfort.
Build proposal follows:
[paste]Worked example
A member used this on a planned multi-agent research system he was about to build in Claude Code. The skeptic identified that the underlying need, “give me a synthesized brief on a topic,” was already solved by Claude’s built-in research mode and a single well-crafted system prompt. The member ran the simpler version that afternoon, got 90% of what the multi-agent system would have produced, and saved the week he had budgeted for the build. He used the week to write three articles instead.