Dev Log #006 — The Brainstorm Tool That Ran On Itself
Nine runs, 604 ideas, zero decisions. The fix wasn't more ideas — it was a hard gate the tool couldn't skip.
A structured brainstorming tool had produced 604 ideas across 9 runs and converted exactly zero of them into a real decision. Pointed at its own failure, it generated 48 candidate fixes and ranked them. The fix that shipped wasn't a better idea generator — it was a mandatory step the tool could no longer silently skip, plus a forced choice at the end of every run.
What happened
The brainstorming tool has a built-in step called a bias audit — check the idea list for blind spots, weight sensitivity, coverage gaps — and it was declared mandatory in the tool's own instructions. Across 9 runs it ran 5 times. Skipped 4. Nobody enforced it because "mandatory" lived in a sentence, not in code.
The bigger number was worse: 604 ideas generated, 0 converted into a chosen next step. Every run ended in a ranked list. No run ended in a decision.
I pointed the tool at its own transcript history and ran it on itself — 48 candidate fixes, one category (tooling), full pipeline including the audit it kept skipping. Ranked, the top fixes weren't about generating better ideas. They were about generating fewer excuses to stop before choosing one.
Why it matters
"Later" is not a smaller number than "never." It's the same number, deferred until nobody's counting.
The shipped fix has two parts. First, the bias-audit step is now enforced by the tool itself: any ranked output produced before that step runs carries a visible ⚠ PRELIMINARY watermark that only a completed audit removes. Declaring a step mandatory in a prompt is a request; a watermark the tool can't turn off itself is a gate. Second, every run now ends by asking for a same-session pick — three to five items, numbers are enough — instead of stopping at a table and calling that done.
What I learned
A tool that measures its own output is a different category from one that doesn't. The 604/0 number existed the whole time; nobody had assembled it into one place before this run, because doing so meant reading 9 separate transcripts and counting. Nine ideas about better ideas wouldn't have moved that number. One idea about what happens after the ideas did.
Next
The tool now writes a one-line record to a tracking file after every run — date, idea count, whether the audit actually happened, what became of it. In 90 days there will be enough of those lines to know whether the fix held or just moved the failure somewhere less visible.
The skill itself is downloadable — definition, the scoring script with the gate in it, and a filled-in example that runs with no setup: Brainstorm Engine — Skill Skeleton.
$ head -12 devlog-006.md
$ cite devlog-006
Citation id SV-0412 is stable. It resolves at
https://stillvalid.dev/c/SV-0412 even if this artifact moves to another section,
which a bare URL does not survive. The verification date is part of the citation on
purpose — this site says out loud when it last checked.
[Dev Log #006 — The Brainstorm Tool That Ran On Itself](https://stillvalid.dev/log/devlog-006) — stillvalid, SV-0412 (devlog, verified 2026-08-17)