A work map for an agent that does not improvise

A short document that defines role, boundaries, permissions, evidence and a stop line before you let an agent work.

What you get

A work map you can paste before any agent task

Who it is for

For developers, founders and operators who want to hand work to an agent without getting confident improvisation back.

1. Role card

Before a task, write the agent a short card. Without it, the agent tries to guess what matters.

Role: what it is responsible for in this task.
Input: which files, decisions or data it is allowed to use.
Output: what it needs to return at the end.
Boundaries: what not to touch without approval.
Stop line: when it stops and brings a question back to you.

2. Work loop

A good agent does not jump straight from request to execution. It runs through a predictable loop.

Reads the input and summarizes its understanding.
Proposes a short plan.
Makes one small change or takes one action.
Brings evidence: a test, a screenshot, a log or a link.
Summarizes what changed and what is still open.

3. Permissions by risk

Not every agent needs the keys to the whole system.

Read: allowed by default if there is no sensitive data.
Local write: allowed only inside the task folder.
External action: always requires explicit approval.
Money, customers, deletion, sending: a human approves before it runs.

4. Paste-in template

Copy box

You are working as an agent for the following task: [task].

Role:
- [what you are responsible for]

Allowed input:
- [files / sources / decisions]

Boundaries:
- Do not touch [area]
- Do not send / delete / charge without explicit approval

Desired output:
- [what to return at the end]

Completion check:
- [what evidence proves it worked]

If critical information is missing, stop and ask.

How to use this now

Correct use: every time an agent gets a task that carries risk, start from a role card. Less magic, more predictable work.

Explore

Tools