LLM field notes

Working notes on language models

A practical guide to what language models actually produce, how I structure prompts, and where the output still needs a human hand. Less hype, more tested workflow.

Every note here comes from my own experiments. I record what worked, what failed, and what I would change next time.

Workflow

How I test a prompt

Start with intent

I write down what the story or artwork needs before touching the model. A clear goal makes every output easier to judge.

Draft a prompt

I build a short prompt with role, context, and constraints. I keep the first version plain so I can see what each line changes.

Run and compare

I generate several passes and note the differences in tone, accuracy, and detail. Patterns show up faster when I log the results.

Refine the narrative

I edit the best output by hand, keeping the parts that serve the story and rewriting the rest in my own voice.

Lessons learned

Limits that shape the output

Three constraints keep showing up in my testing. Each one changes how I plan a session, and each one is easy to overlook when a response reads smoothly.

Token budgets

Long prompts consume space before the model writes a word. I trim background detail first and keep the instructions that shape the scene.

Context windows

Models lose earlier details in long sessions. I restate character facts and story rules at intervals instead of trusting the thread.

Output quality

Fluent text is not the same as useful text. I judge each draft against the scene goal and cut lines that only sound polished.