Practical lessons for using agents well

An ordinary chat model writes a response. A workflow follows a sequence you designed. An agent pursues a goal by choosing actions, using tools, observing results, and deciding what to do next.

Three different things

• Chat: “Summarise this document.”
• Workflow: extract → classify → format.
• Agent: investigate the situation, choose the next useful action, and continue until the goal is met.

The planning example

Context is working memory you manage

More context is not automatically better. Every extra file or paragraph competes for attention. Good context is selected, labelled, and easy to verify—not a dump of everything nearby.

The high-leverage tricks

• Give the goal, constraints, audience, and definition of done up front.
• Point to relevant files instead of pasting duplicate background.
• Ask for a short plan before expensive work.
• Use checkpoints: research → findings → draft → review.
• Pass structured handoffs: facts, decisions, open questions, next action.
• Stop when the goal is met; do not reward extra activity.

Compress before you continue

At a checkpoint, ask for a small brief. That brief becomes the working context for the next phase and makes resuming later much cheaper.

Create a handoff brief with: goal, confirmed facts, decisions, constraints, open questions, files used, and next action. Keep it under 300 words. Do not invent missing information.

Skills make good behaviour reusable

A skill is a reusable instruction package for recurring work. It describes the job, inputs, constraints, tools, output shape, and verification steps. Codex skills and Claude skills are practical examples.

Tools change the risk

• Read: usually low risk.
• Draft: review before sharing.
• Reversible edit: bounded and inspectable.
• External action: explicit approval.
• Destructive action: confirmation plus verification.

A small planning skill

Purpose: turn a vague work goal into a reviewable plan.
Inputs: goal, audience, deadline, constraints.
Output: assumptions, milestones, risks, open questions.
Verify: every milestone has an owner and next action.
Gate: never contact people or publish without approval.

References

• Prompt engineering guide — structuring instructions and context.
• Claude prompting overview — context and instruction techniques.
• MCP specification — tools, resources, and prompts.
• NIST AI Risk Management Framework — risk-aware AI governance.
• Model Context Protocol — a tool and resource integration model.
• Building effective agents — practical agent design guidance.