Give the model a goal instead of a task. "Reframe this document so I can present it to my investment committee" guides the model so much better than "tighten this document", because it names an audience and the decision that audience is about to make. A model working from the goal can tell you the structure is wrong before it starts editing sentences. Name the outcome, name who it is for, and say what a successful version accomplishes.

Plan together before it produces anything. On complex work, tell the model to hold production, ask you questions, state the assumptions it is operating under, and propose an approach you approve before it starts. The models are aggressive, and they like to build. I often tell the model, “Don’t build anything yet; let’s plan it together first.” The planning mode will save you from going down a rabbit hole of revisions, where you discover you briefed the wrong thing.

Make quality observable. "Good," "professional," and "on brand" are reactions, and a model cannot act on a reaction. Use objective rather than subjective language:  the model doesn’t know what “best” means to you until you tell it. Replace "make this look modern" with a style guide. Replace "sound like us" with examples of your writing.

Work with a model that knows you. Custom instructions, memory, an About Me file, CLAUDE.md, AGENTS.md. The documents should hold your preferences, your standards, your goals, your working style, and the decisions already settled so you don’t have to repeat them. The model will start every answer with knowledge of how you like to work, rather than generating generic output. I spent an afternoon on mine, talking through what I want and how I actually work. The instructions function like a cover letter, so edit them often.

Use the ecosystem you are already comfortable in. The one where you know how the files are organized, where the settings live, and what the model does when you hand it a vague instruction. That familiarity does more work than any feature comparison, and it takes months to rebuild elsewhere. Something better will launch while you are reading this, and there is no need to follow every shiny new object.

Bring it the work already sitting on your desk. Something recurring, something with real friction, something that comes out inconsistent. You already know what a good version looks like, so you can judge the output in seconds rather than wondering whether it worked. You will also find out how many exceptions your process has, and how few of them you have ever written down.

Ground it in your own material. Attach the files, the research, the data, the examples, because source quality sets the ceiling. The models are multimodal: you can share images, videos, audio, and files. Some of the best interactions with AI occur through voice: it is often easier and faster to talk it through rather than type it. Share your screen or your browser. We all learn differently, and the model can adapt to the way that is best for you to share.

Turn corrections into a standard the model can read. Make a note every time you correct the model (ask the model to save it as a rejection log). Just the correction, saved where you can find it again. Once you have a stack of them, hand the stack back and ask the model to find the pattern and write the rule it should have been following all along. Keep that rule where it reads before it drafts.  

Ask for evidence that the work is done. Have the model check its output against the sources you gave it, test anything it built, and tell you which parts it is unsure about. Ask it never to guess, and if it is unsure, it should tell you. Then look at what it flagged, and look hardest at whatever came back polished, because the finished-looking document is the one that gets approved without being read. A confident answer and a verified one are different things.

Turn proven work into a reusable system. Start with a task prompt and use it on real work. Once it produces good output several runs in a row, package the instructions and the examples into a template, a skill, or an automation. Reuse gets earned through practice. Package the repeatable steps and leave the ones needing judgment as decision points where a person shows up.