Building a Daily AI Workflow That Doesn't Suck

Most people's "AI workflow" is opening ChatGPT when they remember to, asking it something, then forgetting it exists for three days. That's not a workflow. It's a habit gap with a $20-a-month subscription attached to it.

A real daily workflow has defaults you don't have to think about, lives inside the apps where the work actually happens, and holds up on weeks when you're slammed. This is the setup I've landed on after about a year of iterating, plus the principles behind it.

Pick a default model and stop shopping

You probably have a favorite. Use it. The minutes you spend wondering whether GPT-5 or Claude or Gemini would handle this prompt better are minutes you're not getting your work done.

I use Claude as the default for writing and analysis, ChatGPT for image generation and quick lookups, and a local Llama model for anything sensitive I don't want leaving my machine. I do not A/B test which model writes a better Slack reply. I have a Slack reply to send.

Pick yours and move on. You can change later. The cost of switching defaults is small. The cost of indecision compounds every day you spend bouncing between tabs.

Put it where the work is

The single biggest predictor of whether you'll actually use AI daily is whether it lives inside the apps you already open. Not a separate tab. Not a bookmarked website. The same surface as the work.

For me that means:

• A Raycast hotkey (Option+Space) that fires my current selection straight at Claude
• A Notion AI block sitting at the top of every meeting note template
• The Claude Code extension inside my editor for any non-trivial code change

If your AI tool requires you to switch windows to use it, you'll stop using it inside two weeks. The friction is too high. Treat any AI workflow that starts with "first I open the website" as broken.

Save five prompts you actually reuse

Most people write the same three prompts every week without noticing. Summarize this. Rewrite for a more senior audience. Pull out the action items.

Save them. Not in a Notion doc you'll never open again. Wire them into a launcher (Raycast, Alfred, a text expander) and trigger them with two keystrokes. I keep five core templates: meeting summary, email rewrite, code commit message, blog draft critique, and a Q&A brain dump that turns rough thoughts into clean prose.

This is the highest-ROI thing you can do. The first time you save eight minutes by skipping the "act as a professional copy editor with 20 years of experience" preamble, you'll wonder why you spent two years retyping it.

Define the rule for when you don't use AI

This one matters more than the rules for when you do. Without it, you start running everything through a model, including the things AI makes worse.

My rule: I don't use AI for emails to people I have a real relationship with, decisions where the wrong answer has a consequence I'd have to apologize for, or any writing where my voice is the point. That last category covers most of what goes on TokenByte, ironically.

Your rule will look different. Write it down anyway. A workflow without a "don't use this here" boundary turns a powerful tool into something that smooths the edges off everything, including the things that should stay sharp.

Review weekly, not perfectly

Once a week, look at what you used AI for that worked and what didn't. Not a journal entry. Five minutes on a Friday.

Two questions. What was the best use I got out of it this week? Which prompt am I going to save as a template? You'll notice patterns inside a month. After two, you'll have a system that fits you specifically, instead of a generic "10 ChatGPT prompts" listicle written by somebody who's never used the tool for actual work.

The boring infrastructure is the point

The reason most AI workflows fail is that people treat the model as magic and skip the boring scaffolding: shortcuts, templates, defaults, written rules. The scaffolding is what turns a powerful tool into a daily habit. The model is the engine. The workflow is the car.

Agents are coming, and they'll absorb a lot of this. But agents don't fix bad inputs. The people who'll get the most out of agents next year are the ones who already have clean workflows now, because they already know what they want repeatedly enough to delegate it.

Build the workflow before the workflow builds itself.