The problem
Every great AI tool gets stale the day after you build it.
Garry Tan and the YC partners use AI to do everything: review applications, draft feedback, summarise office hours, surface pattern matches across cohorts. They built tools for each one.
And every week, a partner would hit a wall. "I wish I could ask the AI to do X." But X was a brand new ask — no tool for it yet. They'd write a one-off prompt, get a half-good answer, move on. The gap between what the AI could do and what they wanted it to do never closed. It just kept widening.
What they built
A registry that watches itself. And a "dream cycle" that closes the gap.
YC's system has two pieces. The first is a tool registry — every AI tool the partners use, stored in one place, versioned, observable. When a partner uses a tool, it logs what worked, what didn't, what the partner wished it could do.
The tool registry (Friday afternoon)
application-reviewused 47x this week ✓
draft-followup-emailused 23x this week ✓
summarise-office-hoursused 14x this week ✓
cohort-pattern-matchused 8x this week ✓
competitor-pulseused 3x this week ✓
⚠ gap:"diff this batch's pitch decks against last batch" (asked 3x, no tool)
The second piece is the dream cycle. Every night, the system reviews the gaps and tries to build the missing tool. It writes a draft skill, tests it against the requested examples, iterates, and proposes it for review.
Overnight, the dream cycle runs
22:00
Scan the registryFind all the "wished I could do X" gaps logged this week.
22:14
Draft a candidate toolFor each gap, write a prototype skill that would close it.
22:38
Test against the original asksRun the prototype on the examples partners asked about. Does the output look right?
03:12
Refine, iterate, retryIf it failed, rewrite. Sleep is cheap. Compute is cheap. Iteration is free.
07:00
Propose for reviewThe tool shows up in the partner Slack channel: "I built deck-diff overnight. Want to try it?"
The result: the registry grows itself. Monday morning, a partner shows up to find a brand new tool waiting. They try it. If it works, they use it. If it doesn't, the failure logs back into the registry and the next dream cycle takes another pass.
The result
350+ tools. The AI is smarter every day.
YC's tool registry passed 350 named, versioned, observed tools — almost all of them written by the dream cycle, not by hand. Partners reach for an AI tool first because one usually exists for whatever they need to do. And if it doesn't, they trust that one will tomorrow.
The gap between what the model can do and what we need it to do should close itself.
— YC's internal AI principles