Train a model for the job, not for a demo.

Start by deciding not to train.

Most jobs do not need a custom model. A workflow, a retrieval step, or a prompt on a model that already exists is enough when the path is known and the misses are rare. Training starts when the same step keeps failing on the client's own language, documents, or decisions, and paying a general model to miss it again is the wrong spend.

Write the job before the examples.

The spec says what the model must get right, what it must not do, and which cases will show that. A pile of files is not a training set. The job is written first, the same way any other outcome is written. A prompt is not a substitute for it.

The steps.

• Collect the work the business already does

  The examples come from the documents, tickets, or decisions the client already handles. Invented examples train a model for a job that does not exist.

• Label them against the spec

  A person checks the labels. A wrong label is a wrong lesson. The cases that will be used to judge the model are held back. The model does not see them while it is trained.

• Adapt a model that already exists

  The client's job is taught to a model that already knows language. Training one from nothing is the wrong shape for almost every commission. The adaptation is sized to the step, not to a general assistant.

• Prove it on the cases it has not seen

  The held-out cases are the evaluation. A miss comes back to the examples or the spec. The model does not ship because a demonstration looked finished.

• Put it in the step, inside the boundary

  The adapted model replaces the expensive call on that repeated step. Tools, data, and actions stay limited. A person still accepts the decision the spec says a person accepts.

What a client gains.

• The model fits their work

  It is trained on their documents, their product language, and their decisions. A general demo is not the deliverable.

• The repeated step stops missing the same way

  The cases that failed become examples. The next run is checked against the ones the model has not seen.

• The spend follows that step

  A smaller adapted model can carry the job that a large general model was being paid to attempt. The strong model stays on the step that is still hard.

• The client can tell whether it works

  The evaluation is written. The examples, the held-out cases, and the step it sits in are the outcome they can operate. A person still accepts what must stay with a person.

This belongs to both offers. In engineering, a model can be adapted to a codebase, a test, or a review the pod repeats. In the business, it can be adapted to the documents and the decisions a process already runs. Either way the brief names the job, the examples, and how anyone can tell.