Field notes

Shorter pieces, one argument each.

Notes written while working through a specific problem: what the question actually was, what went wrong, and what to do differently next week.

Field note · Jul 20, 2026 · 2 min read

A Good AI Answer Should Show Its Work

Citations are an interface decision long before they are a trust feature.

Showing sources is usually argued for as a way to build trust. That is true and secondary. The practical reason is that an answer with visible working can be checked in seconds by the person who has to act on it, and an answer without it has to be re-researched or believed. Design the citation for the reader's next action, not for the footnote convention.

citationsprovenancedesigntrust

Field note · Jun 2, 2026 · 2 min read

The Difference Between Finding Information and Using It

Retrieval is a search problem. Usability is a structure problem. They fail in different places.

Teams debug web-enabled AI features as if there is one pipeline. There are two. Finding the right page is a selection problem, solved with queries, scoping, and ranking. Turning that page into something a model can use is a structure problem, solved with normalization, boundaries, and extraction. Mixing them means every bad answer gets the same fix — a prompt edit — and only half of them improve.

retrievaldebuggingnormalizationagents

Field note · Apr 15, 2026 · 2 min read

Technical Content Is a Product Surface, Not a Distribution Channel

Guides, docs, and templates are where evaluation and implementation actually happen.

Treating content as distribution puts it on a publishing calendar and measures it by traffic. Treating it as a product surface puts it next to onboarding and measures it by whether someone finished the job they came to do. The second framing changes what you write, what you cut, and who reviews it — and it usually costs less, because it stops producing pages nobody needed.

contentdeveloper experiencedocsstrategy

Field note · Nov 5, 2025 · 2 min read

When a Search Result Is Not Enough

A ranked list answers “where might this be?” Products usually need “what does it say, and as of when?”

Search is very good at producing plausible places to look and structurally incapable of confirming that the thing you need is on the page. The gap between a ranked list and a usable answer is retrieval, normalization, and extraction — and most disappointing web-enabled features are built as though that gap does not exist.

searchretrievalextractionagents

Field note · Nov 4, 2025 · 2 min read

Monitoring Is Not Intelligence Until Someone Can Make a Decision

A feed of page changes is raw material. Intelligence is the part where a named person can act.

Most monitoring projects succeed technically and fail organisationally. The crawler runs, the diffs arrive, the channel fills up — and six weeks later nobody reads it, because every message says a page changed and none of them says what it means or who should care. The missing layer is not better detection. It is a short chain from signal to decision, written down before the first alert fires.

monitoringintelligencedecisionsalerts

Field note · Aug 21, 2025 · 3 min read

Your AI Agent Does Not Need “the Internet.” It Needs a Retrieval Plan.

“Give the agent web access” is a capability statement pretending to be a requirement.

Handing an agent a browsing tool answers the question of what it can reach and leaves untouched the questions that determine whether the output is any good: what it is looking for, where it is allowed to look, how current the information must be, what shape the answer takes, and what happens when the evidence is thin. Those five answers are a retrieval plan, and writing one takes about fifteen minutes.

agentsretrievalproductscope