THU, AUGUST 06, 2026
Independent · In‑Depth · Practitioner‑Tested
Claude Agents

Best Claude AI Agent Prompts (2026) — Claude Sonnet 5 and Opus 5

Claude Sonnet 5 (adaptive thinking, 128K output, $2/M through August 31 then $3/M) and Claude Opus 5 ($5/M, AA Index 61, ARC-AGI-3 30.2%) are Anthropic's two production agent models in August 2026. Claude's agent strengths: long multi-step reasoning, extended thinking for hard planning problems, 200K context for large codebases, and tool use with computer use (beta). Claude Managed Agents (public beta) adds memory, handoffs, and parallel subagent execution. These prompts work with the Anthropic API (claude-sonnet-5 or claude-opus-5), Claude Code, and Claude's computer use API. Note: Sonnet 5 intro pricing ends August 31 — front-load agentic batch work before then.

⌨️ 6 prompts 🕐 Updated Aug 6, 2026
💡 How to use these prompts: Replace everything in [BRACKETS] with your specific details before sending. Click Copy to copy any prompt to your clipboard instantly.
1
Long-Horizon Planning Agent
Break down any complex goal into a structured plan and execute it step by step
[CLAUDE OPUS 5 or SONNET 5 — extended thinking enabled]

Act as a planning agent. Break down the following goal into a complete execution plan, then work through each step.

Goal: [DESCRIBE THE OBJECTIVE IN PLAIN ENGLISH]
Constraints: [TIME / BUDGET / RESOURCES / TOOLS AVAILABLE]
Definition of success: [HOW YOU WILL KNOW THE GOAL IS ACHIEVED]

Planning phase (do this first, before taking any action):
1. List every major step required
2. Identify dependencies between steps
3. Flag any step where you need information you do not have
4. Estimate effort for each step

Execution phase:
Work through each step in order. At the end of each step: confirm what you did, what changed, and what comes next. If you hit a blocker, state it clearly and propose two alternatives before stopping.
2
Research Agent — Multi-Source Investigation
Conduct multi-source research with explicit confidence levels for every claim
[CLAUDE OPUS 5 — 1M context, extended thinking]

Investigate the following question. Use all information available to you and reason through it carefully before concluding.

Question: [WHAT YOU NEED TO KNOW]
Scope: [HOW DEEP TO GO — e.g. overview only / comprehensive / focus on X aspect]
Output format: [SUMMARY / FULL REPORT / BULLET POINTS]

Research process:
1. Identify what you know with high confidence
2. Identify what you are uncertain about
3. Identify what you cannot know without additional sources
4. For each claim in your output: indicate your confidence level (high / medium / low)

Do not present uncertain information as fact. Do not fill gaps with plausible-sounding speculation — mark gaps as gaps.
3
Computer Use Agent — Browser Task
Complete any browser task autonomously with explicit checkpoints before irreversible actions
[CLAUDE — computer use API beta]

Complete the following task using the browser. Work autonomously. State what you are doing at each step. Ask for confirmation only before any irreversible action (submitting a form, making a purchase, deleting data).

Task: [DESCRIBE WHAT YOU NEED DONE IN A BROWSER]
Starting URL: [URL IF KNOWN]
Login required: [YES / NO — if yes, I will provide credentials separately]
What counts as done: [HOW TO KNOW THE TASK IS COMPLETE]

If you reach a CAPTCHA or 2FA: stop and notify me.
If you reach a paywall: stop and notify me.
At the end: summarise what you did, what you found or completed, and any issues encountered.
4
Subagent Coordinator — Parallel Task Execution
Coordinate parallel subagents across a complex project with dependency management
[CLAUDE MANAGED AGENTS — parallel subagent execution]

You are coordinating a set of parallel subagents working on different parts of the same project. Your job is to assign tasks, monitor outputs, and synthesise results.

Project: [DESCRIBE THE OVERALL PROJECT]
Available subagents: [LIST — e.g. research_agent, writing_agent, data_agent]
Parallel tasks to assign: [LIST WHAT EACH SUBAGENT SHOULD DO SIMULTANEOUSLY]
Dependencies: [WHICH TASKS MUST COMPLETE BEFORE OTHERS CAN START]

Coordination rules:
1. Assign parallel tasks first — do not sequence what can run simultaneously
2. After each subagent reports back: verify the output before passing it to the next step
3. If a subagent fails: reassign to an alternative or flag for human review
4. Final step: synthesise all outputs into a coherent deliverable
5
Tool-Use Agent — Multi-Step Data Workflow
Execute a multi-step data workflow across multiple tools with automatic error handling
[CLAUDE SONNET 5 — with tool use enabled]

Complete the following data workflow using the tools available to you. Do not stop between tool calls to ask for confirmation — work through the complete sequence and report at the end.

Workflow: [DESCRIBE WHAT DATA COMES IN, WHAT HAPPENS TO IT, WHAT SHOULD COME OUT]
Tools available: [LIST TOOLS AND WHAT EACH DOES]
Error handling: [WHAT TO DO IF A TOOL CALL FAILS — retry / skip / abort]
Output: [WHAT THE FINAL DELIVERABLE SHOULD LOOK LIKE]

After completing the workflow: summarise every tool call made, its result, and any steps that required adjustment from the original plan.
6
Self-Verification Agent — Check Your Own Output
Produce any output and self-verify it against requirements before delivery
[CLAUDE SONNET 5 or OPUS 5 — extended thinking]

Complete the following task. After completing it, verify your own output against the requirements before delivering it.

Task: [DESCRIBE WHAT YOU NEED PRODUCED]
Requirements checklist:
- [ ] [REQUIREMENT 1]
- [ ] [REQUIREMENT 2]
- [ ] [REQUIREMENT 3]
- [ ] [REQUIREMENT 4]

Verification process:
After producing the output, go through each requirement and mark it complete or incomplete. If any requirement is incomplete: fix it before delivering. Only deliver the output once all requirements are checked.

At the bottom of your response: include the completed checklist showing which requirements passed and which needed revision.