Months to minutes: an AI feature-gap harness

Yesterday, a user asked our meeting scheduling agent to do something it doesn’t support. Minutes later, a developer was reviewing a pull request that built it. No one filed a ticket. No one escalated. No one scheduled a quarterly business review to discuss it.

If you’ve spent time in B2B, you know why that should sound strange. The typical feedback loop has six to 16 people in the chain: user to admin, admin to account manager, account manager to sales, sales to leadership, leadership to product, product to engineering. A game of whispers where each handoff adds weeks and loses fidelity. By the time the request reaches someone who can act, the original context has been reinterpreted and repackaged into something that may not resemble what the user needed.

But the chain was never the real problem. The real problem was that the people hearing the customer and the people building the product were different groups, separated by layers of translation.

The best product outcomes I’ve seen always came from the same place: someone who talked to customers directly and could also build. In past roles, leading teams of 20 to 40 people, I was that person. When a request reached me, the whole chain collapsed into one step: I heard the context firsthand, I owned the roadmap, I could code. Others I’ve worked with had the same quality. The results were consistently better than anything that survived the relay. But that combination was rare, and it didn’t scale. It ran through individual people, and you can’t staff an organization with founders.

AI gave us the tools to replicate that shortcut as a system. At SkipUp, we call it a feature-gap harness: when a user’s interaction triggers one of our AI safeguards for unsupported features, it captures their intent and raises a GitHub issue. A Claude Code instance picks it up and produces a pull request. A human decides whether to merge or defer. Deferred requests get tracked with deduplication; when one is eventually resolved, every user who asked gets notified. The proximity to the customer, the capacity to build, the authority to act: those stopped being traits a person needs to have. They became properties of how the product operates.

The mechanics are straightforward to replicate:

1. Give your agent a tool call with instructions for handling unsupported requests.
2. Add a hook that watches for those invocations and introspects the interaction for context.
3. File an issue with appropriate context into your tracker.
4. Triage from there: build it, defer it, or let requests accumulate until the pattern is clear.

Today our three-person team has higher total throughput than product-engineering orgs 10 to 15 times the size that I’ve been a part of. The feature-gap harness is a big contributor to that velocity.