I Built the Resume Mechanism First, Then Went to Bed
Delegating overnight research that gets back up on its own after hitting the token limit, and a post-hoc audit of April's idea scores
The idea repository is a repo I built in April where six personas debate and score ideas out of 75. Since what I’m doing now looks unlikely to make money, I had it pull new monetization ideas, and the first thing I did was doubt the score of the idea that got 72 in April and became a real service. That night I delegated the research wholesale and went to bed.
What I set out to do
I reopened the idea repository I’d built in April. It’s the repo where six personas (realist developer, startup PM, investor, AI native, end user, idea bank) debate and score ideas out of 75. I’d concluded that what I’m doing now is unlikely to make money, so I had it pull new monetization ideas again.
The first thing I did was doubt the old scores. The idea that scored 72 in April and became the disclosure-summary service — I said its unit price and break-even looked wrong, and the personas ran a post-hoc audit. The conclusion stung.
v7 (72 points) was built on an assumed paid-conversion rate of 35%. Across 34 rounds of debate, nobody verified this number.
Realistic conversion for a B2C premium service is 2–5%. Rescored: 53 points, a 19-point demotion. A footnote was added that much of the 20-point climb from v1 to v7 was self-reinforcement on top of an unverified assumption. A rule was born that day — every idea calculates “break-even at 3% paid conversion” before entering the debate.
How I asked — filters as questions, resume mechanism first
For the new idea hunt I did two things differently.
First, I defined the filter as a question, not an adjective. “Bring me promising ideas” filters nothing. Instead I had every idea face the question “Couldn’t our developer just build this in-house instead of using an external service?” and told it to bring only the areas where that’s impossible. Because the criterion is decidable, candidates actually started dying in droves.
Second, around 11 p.m. on Aug 11, as I went to bed, I delegated overnight research. The key wasn’t the task instruction but the condition I put in front of it.
I’m going to sleep. Keep researching and writing things up until you hit the token usage limit. The 5-hour limit resets in 1 hour 40 minutes; if you hit the limit before then, build something that restarts you first, then proceed.
The resume mechanism Claude built had two parts. A cron job injects a “research auto-resume” prompt into the session every hour (it’s session-scoped, so it dies with the session), and that prompt tells it to read the handover document RESEARCH_LOG.md and continue from the “next batch” section. Since the document is updated after every batch, wherever it gets cut off, the next resume picks up. Research rules were baked in too — verify market size, incumbents, regulation, and pricing on the web, and always leave sources.
Where it broke — what was left in the morning was a kill list
The hourly resume prompt ran more than ten times through the night. When I opened it in the morning, there were zero winners. Human-approval infrastructure: “build it yourself in a weekend.” Agent memory: “a pgvector wrapper.” Webhook infrastructure: open-source alternatives — every candidate had a one-line cause of death attached, nothing but a growing kill list, and the area where new regulation had supposedly spiked demand was dead with “the government covers it for free.” My first reaction was: so what am I supposed to do with this, there’s nothing to build?
Result — the method changed, and candidate #1 died in two days
Reading it again, the night’s real output wasn’t an idea but a change of method. The approach of deducing ideas from constraints discarded itself after eight consecutive kills, with the diagnosis that “the intersection of the constraints is effectively an empty set,” and the log recorded a new starting point. Start from manual-service markets where money is already changing hands — demand is already validated, and because the customers aren’t developers, the “just build it yourself” filter doesn’t even apply.
On that heading, a Shopify barcode label app came out the next day. A replacement app targeting the spot where the platform’s own free app had been broken by an update, sitting at a 2.2 rating with piling one-star reviews. I reran the persona panel in a compressed 8 rounds, it passed, and a repository was even created. And two mornings later, I killed it. I judged there was no commercial case. I told it to delete, waited, and in the end deleted the repo with my own hands.
Lessons
- The heart of overnight delegation isn’t the task instruction but the resume mechanism. The limit will be hit, so have it build the thing that gets back up after a cutoff (cron + handover document) first, then assign the work.
- Give filters as questions, not adjectives. One decidable question like “couldn’t we just build it ourselves?” filters far more than “bring me good ones.”
- The assumptions behind AI-assigned scores need to be re-audited against real measurements. Even 34 rounds of debate self-reinforced on a single unverified number, and one re-audit dropped 19 points.
- The kill list is an output too. A list with a one-line cause of death per entry, plus screening rules, remained and made judging the next candidates noticeably faster.
- Even candidate #1, picked by a night of research, died in two days. Research only reduces the cost of discarding; it doesn’t make the decision to go ahead for you.