Stress-test a go/no-go decision on your company idea with two AI memos and a written kill criterion

A solo builder with a long sunk cost needs an honest answer on whether a project should become a company, without letting the investment bias the answer.

Before asking any model, the author set a kill criterion (about two years of full-time work to find a business) and had an agent write an index of everything built, abandoned and live, with verdicts and links, so the evaluation started from the record and not the pitch. Five gates were written down with failure conditions in advance: differentiation from a cheap index fund, evidence, regulatory fit, economics for one person without outside funding, and the personal path. Two models (Claude in Claude Code, and Composer in Cursor) each wrote a memo independently under rules such as not treating years invested as evidence and not softening a stop verdict to protect sunk cost. A third model (Claude Opus) reviewed both and found the same two factual errors in each, which the author took as a sign the two memos were not independent opinions. Over the next weeks the author logged 25 counterarguments under 'arguments tried and retired': 16 were retired, 6 kept as partial corrections, 3 left open (two needing a securities lawyer). A re-check on 24 years of data instead of 8 reversed the most favourable correction. The verdict stayed the same and the author stopped building toward a company.

Both memos concluded 'personal only': keep running the system on personal money, do not found a company. None of the 25 counterarguments changed the verdict, and the author stopped building toward a company and returned to engineering leadership. The post says live results were not statistically distinguishable from cheaper alternatives.

Single firsthand account in a regulated finance niche, so the specifics do not transfer to most online businesses; the method does. The author acknowledges both models only knew the record they were given and may share blind spots, and that the counterarguments were generated while trying to overturn the verdict. Memos and prompts are described, not published in full here. Nothing is sold.

A written record of what was built and its results; access to at least two AI models or coding agents; willingness to set the stop criterion before asking.

• DEV Community (redgeoff) ↗ Firsthand write-up, published October 20, 2026

Source published: October 20, 2026. Last reviewed here: October 20, 2026. Spot a mistake? Tell us.