The agent stack I actually run my consulting on

Most "AI for operators" posts are a list of tools nobody uses past week two. This is the opposite — the actual stack behind a consulting practice that runs on it, day in, day out.

The three layers

I think of the stack in three layers, not by tool category:

1. Capture — anything that pulls reality into the system. Transcripts, screenshots, emails, CRM exports.
2. Synthesis — anything that turns capture into a decision-ready artifact. Claude, embeddings, structured outputs.
3. Distribution — anything that gets that artifact to where it's useful. Notion, Slack, email, the doc.

The mistake is to start by picking tools. Start by picking which layer your work is bottlenecked on.

What sits in each layer

• Capture: call transcripts (Granola → Notion), CRM activity exports (weekly), inbox triage (Gmail filters + Claude triage agent).
• Synthesis: Claude as the workhorse. Project files for client context. A small set of repeated prompts I've refined for ~18 months.
• Distribution: Notion for the client-facing artifact, Slack for the team-facing one, email for the founder-facing one. Three audiences, three formats.

The review loop

The thing nobody talks about: the stack only works if there's a review loop. Every Friday I read every artifact the system produced that week, score them, and adjust one prompt or one filter. Compounding happens in the loop, not the tools.

If you don't have the loop, you don't have a stack — you have a subscription bill.