Leverage vs noise

High-leverage AI = high-volume, repetitive, well-defined tasks where quality variance is costly. Low-leverage AI = creative one-offs you already love doing.

The 4-layer personal AI stack

• Brain - Claude / GPT for reasoning
• Memory - Notion stores prompts, templates, learnings
• Automation - n8n for scheduled flows
• Integration - MCPs for persistent capability

Bring the team along

Run prompt clinics. Share templates in a single Notion. Apply growth-experiment mindset to AI adoption: small bets, fast feedback, double down on winners.

Evals: how you know an AI workflow is actually good

Most AI workflows are shipped on vibes. An eval is just a repeatable test set for a prompt or agent - the AI equivalent of a regression suite, and the single highest-leverage habit for anyone putting AI into a real process.

1. Collect 20 real inputs. Not synthetic ones - real tickets, real interview transcripts, real briefs, including the messy edge cases.
2. Write down what "good" looks like for each: the key facts that must appear, the format, what must never appear.
3. Run the prompt on all 20 and grade pass/fail. A 70% pass rate is fine for a draft assistant and unacceptable for anything customer-facing.
4. Change one thing at a time - prompt, context, or tier - and re-run. Keep the score in a sheet with the date.
5. Re-run when the model changes. Provider updates silently shift behaviour; your eval set is the only way you'll notice.

For subjective outputs, use an LLM-as-judge: a second prompt that scores the output against your rubric. Spot-check the judge against your own grading on 10 cases before trusting it.

Shipping AI features, not just using AI

Using AI in your workflow and putting AI in your product are different jobs. If you own an AI feature, these are the decisions nobody else will make for you:

Decision	What to get right
Cost & latency	Know your cost per request and p95 latency before launch. Route easy cases to the cheap tier; stream output so waiting feels shorter.
Failure UX	Design the unhappy path: low confidence, refusal, timeout. Never let a broken generation look like a confident answer.
Hallucination handling	Cite sources, show the retrieved snippet, and make it trivial to verify. "Trust me" is not a UI.
Human in the loop	Decide what AI drafts vs decides vs executes. Irreversible actions need a confirmation step.
Feedback loop	Thumbs plus a reason code, logged with the input. This becomes your eval set.
Guardrails	Input validation, output filtering, rate limits, and a kill switch you can flip without a deploy.

Rule of thumb for scoping: AI should reduce the cost of a step a user already takes. If the feature only works when the model is perfect, it isn't ready.

The compounding PM

Tools change every quarter. Taste and judgment for where AI creates leverage compounds for a career. Invest there.