A few months ago I had Claude, ChatGPT, Perplexity, and Otter.ai open simultaneously at various points in most workdays. Each for a different thing. Each with its own context, its own conversation history, its own set of half-baked system prompts.
That's a lot of cognitive overhead. And I've gotten it down, for most of my actual work, to one primary interface with a set of structured prompts that do what I need.
I'm not going to tell you to do exactly what I do. But I'll walk through the logic, because the principles generalize.
Start With an Audit: What Do You Actually Use AI For?
Before you can build a workflow, you need to know what you're actually doing. Most people's AI usage is more varied than they realize.
Spend a week noting every time you open an AI tool and what you're doing. You'll probably find something like:
- Research and learning (understanding topics, getting up to speed)
- Writing and editing (drafts, emails, rewrites, editing)
- Thinking (working through decisions, stress-testing ideas)
- Technical help (code, formulas, debugging)
- Summarization (long documents, meeting notes, research papers)
- Generating options (ideas, approaches, variants)
These are different modes. And the prompting approach that works for "help me think through this decision" is different from "rewrite this email to be more concise."
The audit helps you design a workflow that serves your actual use patterns, not a hypothetical use pattern.
The System Prompt: Your Standing Instructions
The highest-leverage thing most people haven't set up is a persistent system prompt.
A system prompt is a set of instructions that applies to every conversation — or every conversation of a given type. It tells the AI who you are, what context it should always have, how you prefer it to communicate, and what it should never do.
Without a system prompt, every conversation starts from zero. The model doesn't know you're a startup founder, not a student. It doesn't know you prefer bullet points over dense paragraphs. It doesn't know you're working in B2B SaaS, so "your industry" means something specific.
A good personal system prompt looks something like this:
About me:
- I'm a [role] at a [type of company]
- My main focus areas are [1-3 things]
- I work mostly on [context about what you do]
How I prefer to work with you:
- Be direct. Skip the preamble and disclaimers.
- If you're uncertain about something, say so — don't fake confidence.
- If I ask for your opinion, give one. Don't hedge into uselessness.
- Format: prefer concise prose over bullet lists unless the content is genuinely list-shaped.
- When I ask you to edit or rewrite something, preserve my voice — don't make it sound like generic professional writing.
What I don't need:
- "Great question!" and similar openers
- Unsolicited caveats about AI limitations (I know)
- Recommendations to "consult a professional" unless actually warranted
This alone changes the baseline quality of most interactions significantly.
Different tools handle this differently. Claude has a "custom instructions" or system prompt feature. ChatGPT has "custom instructions" in settings. If you're using the API, you pass a system prompt directly.
Designing Task-Specific Prompts
Beyond the standing system prompt, the next step is building a small library of prompts for the things you do repeatedly.
The goal is zero friction for common tasks. Instead of thinking "how do I phrase this?" every time you want to do a weekly review, you have a prompt you paste, fill in the blanks, and go.
Examples by task type:
Weekly review:
Run my weekly review with me.
This week I worked on:
[list what you did]
Questions I want to think through:
1. What went well and why?
2. What was harder than it should have been?
3. What am I avoiding?
4. What are the one or two things that would make next week successful?
Ask me follow-up questions to push my thinking. Don't let me give surface-level answers.
Decision journal entry:
I'm trying to decide: [decision]
What I know: [facts]
My current leaning: [what you're inclined to do]
What I'm uncertain about: [your doubts]
Help me think through this by:
1. Asking two clarifying questions I haven't answered
2. Naming the assumption most likely to be wrong
3. Describing what you'd recommend and why
Email that needs a delicate touch:
I need to send an email to [person/relationship].
Situation: [what happened, what I need to accomplish]
What I want them to do/feel: [goal]
What I want to avoid: [specific tone or framing issues]
My current draft (or notes): [paste]
Rewrite or draft this, matching my tone. Then flag anything that could be read the wrong way.
Deep focus prep:
I'm about to spend 90 minutes on [task]. Before I start, help me:
1. Clarify exactly what "done" looks like
2. Identify the one thing most likely to derail this session
3. Give me a suggested sequence of steps to work through
Keep your answers short. I want to start working.
Routing: Knowing Which Mode to Use When
Having a system prompt and a library of prompts doesn't mean every interaction should be a heavy structured prompt. Part of building a good workflow is knowing when to be light and when to invest in structure.
Light mode (quick question, freeform conversation): Good for: exploratory thinking, quick lookups, back-and-forth dialogue, things where you're not sure what you want yet.
Medium mode (specific task with context): Good for: writing tasks, edits, analysis, research questions with a defined answer.
Heavy mode (full brief with role, context, constraints, format): Good for: anything you'll use in a professional context, anything you're going to present, anything where quality matters more than speed.
Most people stay in light mode for everything. They get light-mode results. Invest the few extra minutes in structure for the things that matter.
The Context Compression Problem
One real limitation of current AI tools is context. Long conversations accumulate context that eventually gets dropped as it fills the context window. And starting a new conversation means re-establishing all that background.
A few practical solutions:
Keep a standing context document. A short text file with your name, role, current projects, background, and preferences — the stuff you'd put in a system prompt but that changes over time. When you start a new conversation on a familiar topic, paste it in.
Summarize before switching. At the end of a long, productive session, ask the model: "Give me a summary of the key decisions and conclusions from this conversation that I can paste into a new context to continue." Save that summary.
Segment by project. Use different conversation threads (or custom GPTs / Claude projects) for different projects. Don't mix contexts.
The Honest Trade-offs
I said I replaced 4 tools. That's true but incomplete.
Some things I lost:
Perplexity's citations. Perplexity surfaces sources with every response. I still use it when I specifically need attributable web research, because Claude and ChatGPT don't search the web by default the way Perplexity does.
Otter's automatic transcription. Transcription and real-time meeting notes are a different workflow. I still use specialized tools for these.
Different models' strengths. Claude is better at nuanced writing. ChatGPT's Code Interpreter is still better for complex data analysis. Gemini has deeper Google integration. I use a primary workflow but still reach for others for specific cases.
The consolidation reduced my friction and cognitive overhead significantly. It didn't eliminate all tool-switching.
Be honest with yourself about what you actually need versus what you think you need. The goal isn't minimalism as an aesthetic — it's reducing the overhead that gets between you and getting things done.
The foundations of building effective workflows live in the Intermediate Track, specifically the lessons on system prompts, constrained generation, and how to structure complex instructions.