# The Debt Collector: Your AI Said "Done." Did It Actually Run?
You approve your AI's fix. You watch it write "done." You move on. And that fix — the retry
logic, the config edit, the one-line patch you already said yes to — just flies into an approved
folder and sits there, uncounted, forever. Nothing ever comes back to actually run it. That's
suggestion debt: every approved-but-never-executed fix quietly rotting in a backlog file, feeling
finished when it's actually just filed. This page hands you The Debt Collector — the exact
scanner script, the skill procedure it runs on, and a worked example of a rotted suggestion
getting found and executed with a receipt.
## Why suggestion debt happens
Approving a suggestion FEELS like finishing it. You say yes, your AI (or you) write it down
somewhere as "approved," and the brain moves on to the next thing — the same relief you'd get
from actually doing the work, for none of the actual work. Nothing forces a second pass. Weeks
later you rediscover the same broken thing you already agreed to fix, still broken, because
"approved" was never the same thing as "executed."
The full step-by-step (5 minutes)
1. Drop the scanner in. Copy debt-scan.js from the kit anywhere convenient — it's zero
dependencies, pure Node standard library. Point it at your own suggestion stores with a few
environment variables (a backlog file, a flags file, any project's improve/SUGGESTIONS-*.md
— defaults are sane if you don't have these yet).
2. Run it. node debt-scan.js --md prints a markdown table of every pending item it finds,
deduped, classified, and ranked by age.
3. It re-checks before it runs anything. For each item, the skill's run procedure re-opens
the actual file or config the suggestion targets — it never trusts that "approved" still means
"needed." Already fixed by someone else in the meantime? It closes the item as ALREADY-DONE
instead of duplicating work.
4. Safe ones execute; the rest get a manual path. Anything self-only and reversible (a file
edit, a config change) runs directly, archive-first. Anything that touches money, messages
anyone, publishes, or deletes without a backup gets REFUSED with the exact manual command
printed instead — the scanner never guesses on anything with a blast radius.
5. Every flip carries a receipt. The source store's status line changes from "open" to
"done" with a one-line evidence trail, and a separate receipt file records exactly what ran
and how it was verified. Nothing closes without evidence a cold reader can re-check.
On the first real run of this exact scanner, it found a hundred and sixty items that had been
approved and never executed — five got executed clean, re-checked and receipted, in the first
pass.
What's in the kit
- debt-scan.js — the drop-in scanner script, sanitized and dependency-free
- SKILL.md — the list / show / run / sweep procedure it follows
- A worked example — a fake rotted suggestion, the scan output that finds it, and the exact
receipt shape a real run produces
Comment DEBT on today's video and the kit lands in your DMs — or grab it right here with the
button above. Fair warning: the comment word rotates out when tomorrow's drop lands at 4pm.