Prompt guidance

Prompting ACP is different from prompting a stateless chatbot. ACP can work inside projects, continue threads, persist environment state, inspect files, and publish resources. Strong prompts should take advantage of those platform primitives instead of ignoring them.

Start from the operational goal

The best ACP prompts lead with the concrete outcome you want, not vague meta-instructions. Agents work better when they know what should exist at the end of execution and how success should be judged.

State the deliverable first: app, report, deployment, migration, analysis, or fix.

Describe what “done” means in observable terms.

Prefer clear operational verbs like build, update, deploy, audit, or investigate.

Be explicit about state

ACP can continue threads and reuse persistent computers. Prompts should say whether the agent should keep working from current state, inspect what already exists, or treat the task as a fresh start.

Say whether the agent should continue from the current thread history or start clean.

If a computer already exists, tell the agent to inspect before changing anything.

When branching is safer, explicitly ask the agent to fork instead of modifying the current state.

Name the required artifacts

ACP works well when the expected outputs are concrete. If you want code, docs, configuration changes, file edits, or published resources, say so directly.

Name the files, routes, database schema, UI surfaces, or reports you expect.

Mention whether the result should be persisted in the thread, the computer, or published as a resource.

Include acceptance criteria so the agent can validate completion before stopping.

Specify constraints and boundaries

ACP agents can browse, edit files, use skills, and operate stateful computers. Good prompts make the allowed surface area visible so the agent knows what tools, resources, and constraints it should respect.

Say whether the agent may install dependencies, publish resources, or use the browser.

Name the project, resource, or computer it should stay inside.

State any design, security, compliance, or deployment constraints up front.

Prompt for iterative execution

ACP is strongest when work can continue through threads, projects, and persistent compute. The best prompts do not try to compress an entire workflow into one turn. They set a clear next objective and let the system evolve over multiple steps.

Use projects and mission control when a task should become a multi-step plan.

Use agents and teams when work should be split by role.

Prefer progressive prompts over one giant instruction blob.

ACP prompt examples

Project-oriented prompt

State-aware prompt

Research + artifact prompt