Agent or automation planner · provisional name
Control Split
Choose the lightest system that can do the work without outrunning its controls.
Separates repeatable rules from judgment, then recommends a deterministic automation, bounded agent, hybrid workflow, or human-led process. A railway switch sends each part of the work onto the track built for its speed, uncertainty, and consequences.
Input bench
Name the real material
Describe the trigger, the work performed, and the output or action.
How often does this exact workflow happen?
How much does the right path change from case to case?
List the parts a deterministic system could perform the same way every time. One per line. If there are none, write 'None identified'.
Name cases where the normal rule should not be applied blindly. One per line. If there are none, write 'None identified'.
Score the largest plausible consequence, not the average case.
Can the result be restored before real harm spreads?
Choose the strongest check that actually exists today.
What is the most consequential thing the system can read or change?
What specific mistake must the design catch or contain?
Name who owns the outcome and which actions still require their approval.
PRIVACY: INPUTS STAY IN THIS BROWSER. SAVE USES LOCAL STORAGE ON THIS DEVICE. NOTHING IS SENT BY THIS INSTRUMENT.
Output desk
Your result appears here
Load the sample or bring real material.
THE TOOL TRANSFORMS WHAT YOU ENTER. IT DOES NOT CALL A MODEL OR INVENT MISSING FACTS.
Worked example
Starting material
A frequent renewal workflow mixes reliable calculations with customer-specific judgment and contractual consequences.
Useful change
A hybrid route: deterministic code gathers facts and calculates the standard price; a bounded agent prepares the draft; the account director approves every exception and outbound action.
Put the result to work
A method to follow. A prompt to take with you.
Prompts
Delegate or Keep Human?
Decide whether AI should do, prepare, challenge, observe, or stay out of a task based on stakes, context, trust, and reversibility.
Skills & workflows
Agentcraft
Design, compose, harden, diagnose, and attest agents and multi-agent systems so repeated delegation becomes an inspectable operating system.
Tutorials
Agent Reliability for Nondevelopers
Design agent work so a nondeveloper can see the contract, authority boundary, evidence, failure, and recovery path.
StrAItegy writing ↗
The Rules Nobody Wrote Down
The unwritten operating rules of working with AI agents, written down.