Thinking / field notes2025—26

Ideas sharpened by the work.

A deliberately small collection about the decisions between model capability and product reliability. Written to clarify a point of view, not fill a feed.

Product field notes, decision tokens, and an evaluation map arranged on a work table.

01Essay2026 · 5 min

AI-native transformation, not another chatbot

A contrarian case for rebuilding operations, decisions, customer experience, and internal workflows around AI—not attaching another bot to the company that already exists.

Chatbots, voicebots, and copilots are features. The larger opportunity is to ask how the company itself would be built if AI had been available from day one.

Working notes

Map how work and decisions move before choosing an AI interface.
Redesign roles, handoffs, controls, and learning loops as one operating system.
Measure changed outcomes and cycle time, not the number of prompts sent.

Read the full essay

02Essay2026 · 5 min

The pipeline is the product

What building an autonomous content operating system taught me about where automation should stop, and why the model is only one component of a reliable workflow.

Deterministic code should own state, models should own judgment, and people should retain accountability through an approval mechanism the system cannot route around.

Working notes

Bind approval to the exact content version; material edits reset review.
Treat idempotency as a product requirement for unsupervised workflows.
Record unavailable performance data honestly, or the learning loop will optimize on fiction.

03Essay2026 · 4 min

Memory makes the agent

Why I self-host an agent’s memory on a small local model and a private vector store—and why memory is a product sovereignty question, not an API checkbox.

Tools and loops let an agent act, but memory lets it accumulate. The user’s ability to understand and control that memory will determine whether agent products earn trust.

Working notes

Use small local models for narrow, high-frequency extraction work.
Keep the most personal dataset in an agent system inspectable and privately hosted.
Make infrastructure reproducible; a hand-configured memory stack is eventually lost infrastructure.

04Essay2026 · 5 min

AI products fail on power, not models

The biggest risk in AI product development is often not model quality. It is who gains agency, who absorbs uncertainty, and who is accountable when the system is wrong.

A model changes the distribution of decisions inside an organisation. If the product does not make that redistribution explicit, resistance and shadow workflows will beat technical quality.

Working notes

Map whose judgment the system augments, replaces, or makes visible.
Design escalation around accountability rather than hierarchy alone.
Measure adoption through changed decisions, not prompt volume.

05Field note2026 · 4 min

Evals before code

Why a small, opinionated evaluation set is the most useful first artifact for an AI product team.

Writing an eval set forces the team to turn taste into an operational boundary. It exposes disagreement early, before a polished prototype makes weak assumptions feel inevitable.

Working notes

Start with real failure cases, not an abstract benchmark.
Separate model quality from workflow quality and interface quality.
Keep the evaluation set alive as users and data change.

06Field note2026 · 4 min

Confidence is a product primitive

Thresholds, fallbacks, and human review are not backend implementation details. They are the experience of reliability.

Every probabilistic system needs a legible boundary between action, clarification, and escalation. Product teams should own that boundary explicitly.

Working notes

Use the cost of a wrong action to set the confidence policy.
Preserve context across a human handoff.
Make low confidence observable instead of smoothing it away.

07Field note2026 · 4 min

Drift is a product bug

Production models degrade quietly. Treating drift as a distant operations concern is how AI products become less reliable without anyone noticing.

The escape hatches of an AI product—handoff rates, override rates, human QA queues—are its vital signs. When they move, the world has already changed beneath the model.

Working notes

Monitor inputs and score distributions, not only offline model metrics.
Keep a current human-labeled sample as ground truth.
Set alert thresholds, named owners, and retraining triggers before the incident.

08Analysis2025 · 38 min

Designing for LLM-native discovery

How AI-mediated search changes product discoverability, content structure, and the relationship between authority and retrieval.

As discovery moves from ranked links to synthesised answers, products need information that machines can interpret without sacrificing the clarity humans need to trust it.

Working notes

Structure content around explicit questions and attributable answers.
Treat entity consistency and structured metadata as product infrastructure.
Build authority through evidence that remains useful outside the original page.

Read the published analysis