AI decision loop
The problem
The company's signals lived scattered: leads in the pipeline, clients in production, finances in Stripe, affiliates somewhere else. Decisions came from gut feel and nobody checked afterward whether they paid off. With no decision memory, the same mistakes kept coming back.
What we built
A repo-native decision loop, inside the app itself:
- Typed collectors pull signals from the four sources in Neon (leads, clients, finances, affiliates).
- Claude synthesizes a daily brief: what changed in 24h and 1 to 3 prioritized recommendations.
- The founder accepts a recommendation and it becomes a decision, with a hypothesis, a target metric and a review date.
Every Monday, the weekly review pulls each due decision's metric, compares it against the baseline and records whether the bet worked.
How a decision becomes measurable
Each recommendation carries what it needs to be measured later, not just text:
- Decision type: build, focus, growth or health, classified by intent.
- A machine-pullable target metric: source, query, measurement window, success direction and the baseline at decision time.
- An Outcome Score (-1, 0 or 1) set by the weekly review, shown next to the measured-vs-baseline math.
Criteria and sources are versioned in the repository: they change with the code, with history.
Architecture
The daily brief
Architecture decisions
Repo-native, not a separate service
The loop lives inside the Next.js app itself, on the same Neon. The signals are already there; an external service would only add syncing and latency with no gain.
The loop closes on measurement, not on a suggestion
The recommendation isn't the end. A decision only counts when it has a target metric and a review date, and the weekly review records the result. That's what turns a hunch into learning.
Vocabulary discipline
A proposed recommendation is never called a decision. The word decision only applies once the founder accepts. That boundary keeps a candidate from being confused with a commitment.
Outcome
- Four signal sources consolidated into a single daily brief.
- The loop closes: every decision has a target metric and is measured in the weekly review.
- Zero manual spreadsheet: the brief and the review arrive by email, with no collection work.
Stack
How we executed
We built the loop in thin increments: first the collectors and the decision schema, then the daily synthesis, finally the weekly measurement. Each part shipped to production running before the next one started.
“I stopped deciding on gut feel. Every bet has a hypothesis and a date to look back, and on Monday I see whether it worked. The company became a loop that learns.”