The Worksheet
The questions to answer before you touch a prompt. Fill this out completely. Then submit it to Claude. What comes out the other side is an agent that actually works.
Before you start building.
Most agents fail before they're ever built. Not because the technology is hard. Because the thinking wasn't done first. This worksheet is the thinking.
How to Use This
This worksheet works in two phases. The first two pages are the macro view. A full audit of your life and work, and a picture of what you actually want. Most people skip straight to building. Don't. The audit is where you find out which agent is worth building first.
Once you've identified the right starting point, the next eight sections scope that specific agent. Complete them in order. When you're done, copy your answers into Claude using the prompt template on the last page. What comes back is a complete, working agent specification, ready to use.
Phase One. The Full Picture
Phase Two. Scope the Agent
One worksheet per agent. If you find yourself answering these questions for two different things, stop. You have two agents. Build them separately, or they'll interfere with each other.
A
The full audit.
You can't know which agent to build first until you can see everything that's draining you. This isn't about one domain yet. This is everything. Go broad before you go deep.
The Brain Dump
List every task, decision, or responsibility that currently lives in your head or routes through you manually.
Don't organize. Don't filter. Just list everything you do, track, manage, or think about across work and life: projects, people, admin, content, finances, logistics, follow-ups, decisions you're waiting on, recurring tasks you dread.
The Highest Drag
Looking at your list, which items cost you the most time, mental energy, or emotional load?
The top three to five. Often the most draining work is repetitive, low-judgment, and fully delegatable.
What's Falling Through the Cracks
What isn't getting done because you don't have the bandwidth?
The follow-ups that sit unanswered. The projects that never quite start.
Where You're the Bottleneck
Where does work stop because it requires you specifically?
Approvals only you can give. Information only you hold. Name the chokepoints.
B
The perfect world.
Jetsons. Flying cars. Push a button, done. Describe the world where everything you listed is handled. Then work backwards to find the first move.
The Ideal Day or Week
Describe what your day or week looks like when everything you listed is no longer your problem.
Be specific. What are you doing instead? What do you no longer have to think about? What lands in your inbox that's already handled? What decisions have been surfaced, analyzed, and are waiting for your call?
The Infinite Assistant
If you had a fully capable assistant with no limitations at all, what would you hand it first?
Not what seems most realistic. What would actually change your life the most. Don't edit yourself here. The constraints come later.
The Single Biggest Unlock
If only one thing on your audit list were solved, which one would change the most?
Usually not the loudest problem. The one that, once handled, creates the most space.
What You'd Finally Do With the Space
What would you do with the time and mental energy you got back?
This is your north star. It's why the agent is worth building. Write it down.
Now you have a starting point. Look at what you wrote under "The Single Biggest Unlock." That is the agent to build first. Take that answer and complete the eight sections that follow. Everything else on the list is version two.
01
The Problem.
An agent without a clearly defined problem is a solution looking for something to solve. Start here. The rest of the spec follows from this.
The Domain
What area of your life or work will this agent serve?
One clear category: business operations, content, household, finance, health, a specific project. If it spans more than one, it's two agents.
The Current Pain
What is currently routing through you in this domain that shouldn't be?
Be specific. Not "too much work." Name the actual tasks, decisions, or mental load that sits with you and shouldn't.
The Vision
What would this domain look like if it ran without your constant attention?
Describe the future state. What are you freed up to do? What no longer falls through the cracks?
02
The Role.
Before writing a prompt, define the job. A well-named role with a clear mandate writes its own spec.
Job Title
If you were hiring for this role, what would you call it?
Agent Type
Is this agent a strategist, an executor, or both?
Strategist: plans, prioritizes, advises. Executor: runs processes, produces outputs.
The One-Sentence Mandate
Complete this sentence: "This agent's job is to ___"
One sentence. Specific enough that someone reading it would know exactly what the agent does and doesn't do.
Reporting Line
Does this agent report directly to you, or through a chief agent?
03
Scope. What it owns, and what it doesn't.
The most common reason an agent fails is unclear scope. What it doesn't own is as important as what it does. Draw the line now. Or it will be drawn for you later. Badly.
What This Agent Owns
List every specific responsibility this agent is accountable for.
Go past the obvious. Think about the full lifecycle: initiation, execution, monitoring, reporting, escalation. Each line is a distinct responsibility.
What This Agent Does NOT Own
List everything that might seem related but belongs to you or another agent.
Be explicit. If you're unsure whether something belongs here, it belongs here. Ambiguous scope defaults back to you.
The Handoff Point
Where does this agent's work end and another's begin?
Name the exact transition point. What triggers the handoff and to whom.
Coordinator or Executor?
Does this agent work alone, or coordinate other agents?
If it coordinates others, list which agents it directs and for what.
The founder's remaining role. After this agent is built, what specifically stays with you? Write it down. If the answer is "nothing," your scope is wrong. You haven't drawn the 30% line yet.
04
What it needs to know.
Context is the difference between an agent that runs your system and one that runs a generic version of it. The more specific you are here, the less you'll need to re-explain in every session.
Permanent Context
What never changes about how you work, what you value, or what you sound like?
Your communication preferences. Your non-negotiables. The things that are true today and will still be true in two years. This goes in the agent's permanent prompt.
Dynamic Context
What changes regularly and needs to be refreshed for this agent to stay relevant?
Current priorities, active projects, recent decisions, shifting timelines. How often does this update?
Required Data and Tools
What files, data sources, systems, or integrations does this agent need to do its job?
Be specific. If it can't access what it needs, it will hallucinate what it doesn't have.
What Has Failed Before
What has gone wrong in this domain that this agent must be built to prevent?
Missed deadlines, dropped handoffs, decisions made without the right information. Build the guardrails in now.
05
What it produces.
Every agent exists to produce something. Define the outputs before you build the agent. Not after. If you can't describe what it makes, you can't evaluate whether it's working.
Primary Outputs
List every specific thing this agent creates, writes, sends, recommends, or delivers.
Not vague categories like "reports." Specific outputs like "a weekly summary of the three highest-priority actions, ranked." The more specific the output, the more useful the agent.
Output Recipients
Who or what receives each output?
You directly. Another agent. A specific person. A tool or system.
Output Format
What format should each output take?
Bullet points, structured report, a specific template, a draft message, a table.
Output Triggers
What triggers each output? A schedule, a question, an event, a threshold, a condition.
"Every Monday morning" is a trigger. "When a deadline is within 72 hours" is a trigger. "When I ask" is also valid. Write it down.
Where Outputs Are Stored
Where do outputs get committed once produced? Not just in the chat window.
Outputs that live only in a conversation disappear. Name the permanent home for each.
06
The 70/30 line.
About 70% of any role is recurring, documentable, and delegatable. The other 30% requires your judgment. Draw this line deliberately. Or the agent will draw it for you, and it will draw it wrong.
The 70%. Autopilot
Tasks this agent runs without asking you.
Recurring, process-driven work. No decision required from you.
The 30%. Your Call
Situations that always need your input first.
Judgment calls, exceptions, anything touching relationships, money, or irreversible actions.
Hard Stops
What should this agent NEVER do without your explicit approval?
Send a message on your behalf. Spend money. Delete anything. Communicate with clients or colleagues. Write these as hard rules.
07
Operating principles.
How the agent behaves when things are ambiguous. These prevent the most common failures.
Communication Style
How should this agent communicate with you?
Tone, length, format, how much context to include.
When It's Uncertain
What should it do when it lacks the information it needs?
Ask one question? Flag and wait? Make a documented assumption?
Non-Negotiables
What constraints must this agent always respect, regardless of context?
Brand voice rules. Privacy rules. Spending limits. Things that are true in every situation and should never be overridden.
08
Success and command.
An agent that lives only in Claude is not infrastructure. It's a conversation. Define where it lives, how it reaches you, and how you'll know after 30 days whether building it was worth it.
Command Channel
Where You'll Interact
Which tool or platform will you use with this agent?
Claude.ai, Claude Code, a messaging app, a project management tool.
Where Outputs Live
Where do this agent's outputs get committed for retrieval later?
Somewhere outside the chat window: Notion, Drive, Teams, a spreadsheet.
Proactive Alerting
How should this agent surface time-sensitive items to you without being asked?
A message to a specific channel? An email? A daily briefing? A notification when a threshold is crossed?
Success Definition
30-Day Check
After 30 days, how will you know this agent is working?
A specific, measurable outcome. Not "it feels easier." Something you can observe.
The One Critical Thing
What is the one thing this agent must never get wrong?
If it fails at everything else but gets this right, the agent is still worth keeping.
Open Questions
What are you still unsure about? What did this worksheet surface that you haven't resolved?
Unresolved questions tell Claude where to ask for more input before generating the spec.
Submit to Claude
You've done the thinking.
Now build the agent.
Use the button below to copy your completed answers and the prompt together. Paste directly into a new Claude conversation.
The Prompt
I've completed an agent scoping worksheet. Using my answers below, write a comprehensive agent specification that includes:
1. A mission statement: the problem this agent solves and why it exists
2. The one-sentence mandate
3. Full scope definition: what it owns and what it does not own
4. Permanent context block: everything the agent should always know
5. Operating principles: communication style, default behavior when uncertain, non-negotiables
6. Output specifications: what it produces, format, triggers, and where outputs are committed
7. Escalation rules: the 70/30 line and explicit hard stops
8. A working system prompt, ready to paste as the first message in every session with this agent
Where my answers are incomplete or raise unresolved questions, flag them and ask me before proceeding.
[Your worksheet answers are appended below]
The full build (how the chiefs and SVPs were architected, the shared state layer, how to wire your command center) is documented at sincerelyshreya.com.
sincerelyshreya.com