Don’t ask chatbot questions, start writing briefs
How to go from chatbot questioning to writing first agents
If I had to start my agentic workflows from absolute zero today while knowing everything I know after writing How to command meetings better than 99% of tech managers and I got two hours back from the worst part of my job. Here’s where they went, I’d do a few things differently.
This article is for managers who are just getting started or who feel stuck and unsure how to get started with agentic workflows to tackle their daily tasks.
1. Create your Digital Double
Note: This point assumes you’ve already used chatbots like ChatGPT; otherwise I highly recommend you kickstart guide Claude Cowork 101 by Ruben Hassid (link) with a how-to on how to create your personal digital-double!
A digital double is your “copy” in the form of an AI agent that is aware of a selected amount of facts, that you provide it.
It’s helpful in kickstarting your first agentic workflows by providing the model (I use Claude Opus 4.6) the crucial context about you.
I recommend exporting a concise summary of your career profile and the top 5 tasks that you executed in the chat.
There’s a tremendously simple way that I used to “onboard” on my personality. I ran the following prompt into ChatGPT:
I’m creating a digital double in Claude. Your task is to assist me in that by:
1. Exporting all that you know about me for Claude to have as much professional context as possible, into a markdown file. The outcome: Claude must ingest my profile as clearly as possible to immediately form an impression about me. Target file: about-me.md
2. Exporting the top 5 tasks that I you have assisted me with. Group each task by the applicable prompt. Provide information about inputs and outputs of said tasks. The outcome: obtain foundation for agents with potential hand-off scenarios. Target file: my-tasks.md.
Limit each file to 1000 tokens.
These two files go into a Claude Project: add them as a new Artifact. Now, every time you prompt Claude, either in Chat or Cowork modes, it will first look these files up.
This is where the magic starts - Claude already knowing who you are, how you work, and what your top tasks look like.
That's the foundation your first agent runs on.
2. Establish a Feedback Loop
Every knowledge worker I’ve watched get sharper at their job does one thing consistently: they take the feedback and incorporate it relentlessly.
After a project, a tough conversation, a failed experiment, they inspect what worked and what didn’t - feedback is key!
The same cycle applies here:
Define what you’re trying to get done;
Build the workflow and run it;
Reflect: did that produce what I actually needed?
Provide feedback → update the context, try again. Repeat.
An agent that doesn’t continuously “learn” about you is useless.
Most of the people I’ve seen use agents completely miss out on closing the loop - the agent will not learn, leaving a ton of information about your work that you could’ve used.
One prompt that does the job:
Review my past 5 prompts. What did you learn about how I work, what I need, or what I want to avoid that isn’t already in my about-me file? Propose additions only: plain bullet points. Exclude what’s nothing already covered.
When Claude misreads your communication style, add a line.
When a workflow returns something too formal, too long, or structured in a way you’d never send - note it.
30s of friction converted into permanent signal.
I debrief in Notion every day. Most of that stays in there…
…but when something teaches me something about how I work (i.e. what I need, what I hate seeing, what “done” actually means to me, etc), I immediately update my about-me.md file.
3. Write a brief, not a question
The single biggest shift in how I use Claude happened when I stopped asking questions and started writing briefs.
A question gets an answer - a brief gets work done, that is specific to you.
When I need Claude to help me prep for a performance review cycle, I don’t type “how do I run a performance review?” I write: here’s what I need, here’s the team context, here’s what done looks like, here’s what I want to avoid. That’s 1 minute of setup. Then I run it.
Key difference: be as precise as possible about the task.
Before you open a new chat, write down 5 things:
What you’re trying to get done;
What context it needs to do it well;
What a good outcome looks like;
What it should stay away from;
(If applicable) what’s weird about it.
Here’s what that looks like as an agent when the performance review season arrives.
I open a new markdown file and add it as a Claude artifact:
# Performance Review Agent
## Task
Draft performance review summaries for each of my direct reports.
## Context
- 6 engineers on the team
- Quarterly 1:1 notes: [link your Notion docs here]
- Promotion ladder criteria: [link here]
## Example
- Summary for John from last year: [link here]
## Output
One paragraph per person. Calibrated against the promotion ladder, ready to paste.
## Avoid
- Generic HR language.
- Creativity; be grounded solely in the data provided.
## What's weird
Some engineers may have less than a full quarter of notes. Work with what's there.This is just a starter, not a finished product - it will require multiple reps on honing it and adapting it, but it will provide the 80% of results for 20% of effort.
Bottom line: Claude runs it in under a minute. Next cycle, the brief is already there - I update the context and run it again.
The brief is the program. Save it, reuse it, hand it to an agent.
The premise of this article: you don’t need perfect to get started with agents.
You don’t even need all three at once. Just start with one.
Run that first chatbot prompt, paste what you get into a new Claude project, and see what happens. Everything else builds from there.
One question before you close this tab: what’s the first task you’d hand to an agent that actually knows you?
— Leszek