A Claude Project that red-teams product decisions against your own feedback log

Solo founders have nobody to challenge their product calls. In this case a founder shipped a feature to 51 beta testers that was dropped within a week, although a tester had already said in the feedback log that they did not want it.

1) Create a Claude Project for decisions. 2) Write standing instructions: do not cheerlead, surface failure modes and positions that were dismissed, check past decisions before answering, and never call an idea good because the user wants it to be. 3) Add knowledge files: a jobs-to-be-done file (one plain sentence per core user job), a user feedback log (dated, tagged quotes added after each support chat or beta round), a decisions log (one dated entry per meaningful product call) and an anti-patterns list (short named mistakes that recur). The founder suggests starting with only the first two. 4) For each decision, fill in five fields: the decision, why now, evidence for, evidence against, and what happens if you are wrong. An empty 'evidence against' field is treated as a warning sign. 5) Run a three-part prompt: a premortem with ranked failure modes, each with an early warning sign and a cheap test; a red-team memo making the strongest case against, citing dated feedback; and a decision memo with a verdict (go, no-go or test first) and kill criteria that can be checked within 30 days. 6) Save the memo with a date into the decisions log and schedule a 30-day review. Use a new chat per decision inside the same project. The author later found that running a full premortem on every idea caused paralysis, so it is now a gate only for expensive or hard-to-reverse decisions.

In a live demo, the setup flagged that the founder was about to re-ship a feature she had already dropped months earlier with the same reasoning, and pointed out that the study she relied on was about supervised children rather than adult app users. The article says the method had only been used inside that founder's young company as of July 2026, so long-term results are not known. No business numbers are reported.

Written by a newsletter author about another founder's system after a live session, not by the founder herself. Only one demo result is described, with no measured effect on the product. The author sells Claude Code courses, toolkits and a paid newsletter tier, so the source is labelled vendor. The full prompts are described rather than all printed in full.

A Claude plan with Projects, and the habit of logging dated user feedback and product decisions.

Begin with only the user-jobs file and the feedback log. Add the decisions log and anti-patterns list once you have a few decisions recorded.

Decision: [what you plan to do]. Why now: [...]. Evidence for: [...]. Evidence against: [...]. If I am wrong: [...]. Using the feedback log and decisions log in this project: 1) list eight ranked ways this could fail, each with an early warning sign and a cheap test; 2) write the strongest case against it, quoting dated feedback; 3) give a verdict of go, no-go or test first, with kill criteria I can check within 30 days. Tell me if I have made this decision before.

• GenAI Unplugged (Dheeraj Sharma) ↗ Vendor case study, published July 26, 2026

Source published: July 26, 2026. Last reviewed here: October 27, 2026. Spot a mistake? Tell us.