Two months behind a first attempt

A field note from Marc's account of making Rage gear.

After roughly two months of sustained building with agents, Marc asked an agent to make Rage gear. He described the result as working in one attempt and attributed it to the context the agent already had about him and what he liked.

That experience captures something this book wants to teach: a new task can benefit from judgment developed during earlier work. The next request can start with references, preferences, and decisions that would otherwise need explaining again.

The work before the request

The interesting lesson is the preparation that can accumulate across a working relationship. A person makes choices, corrects an interpretation, keeps a useful result, and explains why something misses. When those decisions are retained and made available for a relevant task, the next brief can begin further along.

For a continuing creative collaboration, useful context might include accepted references and why they fit, discarded directions and why they missed, the intended audience, recurring vocabulary, and constraints that remain current. These are examples of what a context brief could contain; they are not a reconstruction of the material used for Marc's gear.

The useful practice is to preserve those decisions in a form the next task can use. An accepted example and a clear reason can be more useful than a long record of conversation without a current direction.

Make the earlier work available

Keep a short, dated context brief with a few selected references. Write down the reasons behind a preference. A saved image can show what was accepted; a sentence about the choice can help a future agent recognize which part to carry forward.

At the next task, identify which references still apply and which constraints have changed. Ask the agent to use those materials, then inspect the result. A new session, another model, or a different tool may have different access. Time spent working together does not by itself establish what context the next run can see.

If the result fits, preserve the useful decision and its scope. If it misses, identify whether the problem was the brief, unavailable context, a changed preference, the procedure, or execution. Correct the relevant record before adding more instructions everywhere.

Taste and procedure have different homes

Knowing what a person likes can help choose a direction. Knowing how to prepare and check a deliverable helps carry that direction through a workflow. Both can matter to a successful attempt.

A preference such as which examples feel right belongs with the approved references and their reasons. A repeatable sequence for inspecting inputs, preparing outputs, and checking them may belong in a skill. Keep the reusable procedure connected to the current brief so that it can serve different people and changing preferences.

What to take into your own work

After a useful session, save one accepted artifact, the reason it worked, any changed preference or constraint, and the next unresolved question. Make it easy to find when a related task arrives. You do not need to preserve every exchange to carry forward a useful lesson.

The ambition is a working relationship in which new requests can benefit from earlier judgment. Marc's account makes that ambition tangible. The practical test is whether the context you preserve helps the next task, with enough visibility to correct it when it does not.

About this account

This is Marc's reported experience, rather than a controlled comparison of the causes of success. The original prompt, loaded context, and wider production steps were not examined for this note. The proposed method above is available to try; private artwork, conversations, and production records remain outside the book.