Work samples

Written by a person. Last read by a person on 2026-09-07, 36 days ago. That is older than the 30-day threshold. Ask for this to be re-checked

These are deliverables, not descriptions of deliverables. Each set is the kind of document the job actually produces, written to be used rather than to be admired.

Every page states who it is for before it starts, records when a person last read it, and names which of its claims a build checks rather than asserting they are true.

Start here

Technical curriculum Lab (interactive)

For a technical enablement, curriculum or developer-education hire.

One topic chosen by a needs analysis and a ranked inventory of 1,830 pages, and the hands-on lab it produced: an explorable model of a Claude workload that a rep can use on a call and an engineer can read as code, with every figure traced to its source.

End-to-End AI Enablement 16 documents

Start here if you are deciding whether someone can fund, buy, scale and defend a program rather than only build one.

A strategy, an executive deck, a project plan, a statement of work, budget and vendor analysis, a first 90 days, a literature review, and the objections a sponsor will actually raise.

Retrieval-augmented archive Live demo

For a role building or evaluating retrieval-augmented generation, not just talking about it.

A real RAG system over sixteen public-domain magazine issues: hybrid retrieval, grounded answers with citations, honest refusal, and the rate limiting and health-gated deploys that keep a small, self-funded demo cheap and always up.

Build and debug AI agents Practice (interactive)

For learners building AI agents on the Claude API, and the facilitators who guide them.

Practice tasks whose help fades by level, from worked versions to building and explaining unaided: one on what each design choice costs in requests and tokens, one on diagnosing a broken agent from its evidence. With the needs analysis and the research behind each design choice.

Where the round trips come from Lesson (interactive)

For learners who can make an agent work but cannot yet say what it costs to run, and the facilitators who teach them.

A five-minute model the reader operates rather than reads: change the number of customers, the shape of the tools and how results go back, and watch the requests to the model and the tokens move, counted by the same scripted run the tests check.

The needs analysis and cited principles behind it

Spot the bug Lesson (interactive)

For learners who write an agent's tool loop themselves, and facilitators who want a five-minute warm-up for a group.

A tool loop that runs and still costs its owner money: the reader marks the lines that add round trips, break the prompt cache or get the next request rejected, some lines are decoys, and every explanation cites the Anthropic documentation it rests on.

The needs analysis and cited principles behind it

Developer documentation 4 documents

For an engineering or developer relations hire.

One API workflow documented end to end: a quickstart, a reference, retry behavior, and troubleshooting that starts from what the reader can see rather than from what went wrong.

Support content

For a support or education role.

Help written from the reader's situation. It opens by asking which problem you have, because someone who is stuck cannot yet name the thing they need.

Concept explanations

For anyone assessing teaching, not just writing.

The same idea explained more than one way, because a reader who learned it elsewhere needs the words lined up with what they already know.

Accessibility statement

For a team with a procurement or compliance question.

What this site does against WCAG 2.1 AA, separating the claims a build checks from the ones a person is vouching for.

Published rubrics 24 documents

For anyone who wants to check the work rather than take it on trust.

The rubrics the set is graded against, published so a claim can be argued with.

How this set stays true

For anyone who has watched documentation rot.

The argument behind the machinery. Facts live in one place, prose references them, and the build fails when a page repeats a value it should have looked up.