How to Prompt Claude Opus 5: Five Rules That Cut Scope, Tokens, and Rework

Direct Answer

Claude Opus 5 responds best when a prompt gives it the complete job, a clear stopping point, calibrated output limits, and concrete acceptance criteria. It is more capable of planning and completing long tasks on its own, but it can also widen scope, write too much, spawn unnecessary subagents, and repeat verification work that it already performs internally.

Paul J Lipsky turns Anthropic's dense prompting guide into five approachable rules: give the full job up front; define where the task ends; control the response length; control the deliverable size; and remove generic "double-check everything" instructions. That is a strong beginner framework, with two qualifications: old Opus 4.8 prompts do not suddenly fail, and effort should be selected by testing the task, not merely by the subscription plan.

Watch the Walkthrough

Video credit: Paul J Lipsky. Paul translates the official guide into a non-technical website example using a fictional Northline Coffee brand. The reusable templates below are original JQ AI SYSTEMS adaptations rather than copies of Paul's complete prompt.

Source and Correction Note

The supplied transcript is the source for Paul's model-routing advice, five-rule framework, Northline Coffee prompt, and Claude Design demonstration. Anthropic's official Opus 5 prompting guide, effort guide, model overview, general prompting guidance, and platform release notes were checked on 28 July 2026.

Anthropic's wording is more measured than the video's opening. The company says Opus 5 performs well out of the box on existing Opus 4.8 prompts. The model does not require a completely new prompting language. The practical migration work is removing obsolete scaffolding and tuning the behaviors most likely to waste time: verbosity, scope expansion, excessive delegation, and repeated verification.

Effort also deserves a correction. Paul recommends low for Pro and medium for Max as simple defaults. Anthropic instead recommends starting at the default high effort, then running an effort sweep on representative tasks. Low and medium are often strong and cheaper, but the correct setting follows measured quality, latency, and cost, not the name of the subscription.

Useful Links

The Five Rules, Source-Checked

Rule 1: Give the complete outcome up front

Anthropic says Opus 5 is strongest on difficult multi-file and end-to-end work when it receives the complete task specification and is left to run. This reduces the chance that the model optimizes an early subtask without understanding the final product.

Build a one-page product website for [brand].

Outcome:
Help [audience] understand [offer], choose [product], and complete [conversion].

Use:
- assets in [folder];
- the existing design tokens;
- product facts from [source].

Deliver:
A finished implementation in [destination], with the existing shop flow preserved.

"Complete" does not mean enormous. A simple task can still use a one-line prompt. The rule matters most when the model must coordinate several dependent steps.

Rule 2: Define scope and the end condition

Opus 5 can improve more than the user intended. Give it room for routine judgment while naming the boundaries that matter.

Scope:
- build only the homepage;
- preserve the current logo, colors, product names, routes, and checkout;
- do not add a CMS, authentication, new dependencies, or new products;
- make routine visual decisions yourself;
- ask one question only if different answers would materially change the outcome.

Stop when:
- the homepage is implemented;
- all existing tests pass;
- the specified viewport checks pass;
- no out-of-scope file has changed.

This is better than telling the model exactly how to perform every implementation step. Opus 5 can choose the route; the person defines the destination and guardrails.

Rule 3: Specify the completion response

Anthropic says effort controls thinking volume, not visible response length. If you want a concise answer, request one directly.

When finished, respond with:
1. the outcome in one sentence;
2. up to three bullets covering the main changes;
3. the verification result;
4. any blocker or decision still required.

Do not narrate routine steps or repeat the brief.

Rule 4: Limit the artifact separately

The answer can be concise while the file it creates is bloated. Calibrate the actual deliverable.

Rule 5: Remove redundant verification prompts

Anthropic says Opus 5 self-corrects and verifies its own work without being told. Legacy instructions that demand another full verification pass or a verifier subagent can add latency and tokens without improving quality.

The wrong conclusion is "verification no longer matters." The correct conclusion is do not duplicate vague internal checking. Keep external evidence: tests, screenshots, source links, calculations, lint, review, and acceptance checks.

Choose Effort by Task, Not Plan

Anthropic documents five Opus 5 effort levels: low, medium, high, xhigh, and max. The official starting point is high, followed by an effort sweep on your own evaluations. Low and medium are important because they can retain strong quality with lower token use and latency.

Run the same representative prompt at high, medium, and low. Compare accepted result, corrections, latency, and tokens. Then route the task to the lowest effort that consistently clears the acceptance threshold.

The Anatomy of a Strong Opus 5 Prompt

1. Outcome: the finished result and why it exists.
2. Context: audience, current state, approved facts, assets, and existing system.
3. Scope: what to change and what must remain untouched.
4. Judgment: decisions the model can make without asking.
5. Escalation: questions or consequences that require the person.
6. Artifact limits: size, sections, pages, screens, duration, or complexity.
7. Acceptance criteria: observable checks that establish completion.
8. Completion response: the short summary needed for review.

This structure is more durable than phrases such as "think harder," "act as a genius," or "make it world-class." It gives the model operational information rather than emotional intensity.

Reusable Opus 5 Prompt Template

OUTCOME
Create [DELIVERABLE] that enables [AUDIENCE] to [RESULT].
The finished artifact must exist at [DESTINATION].

CONTEXT
- Current state: [WHAT ALREADY EXISTS]
- Source material: [FILES / LINKS / DATA]
- Approved style or examples: [REFERENCES]
- Important facts: [FACTS THAT MUST REMAIN ACCURATE]

SCOPE
Do:
- [REQUIRED WORK]

Do not:
- [OUT-OF-SCOPE WORK]
- change [PROTECTED ELEMENTS]
- add [UNWANTED DEPENDENCIES / FEATURES / SECTIONS]

JUDGMENT AND QUESTIONS
Make routine, reversible decisions yourself.
Ask only when different answers would materially change the result, cost,
permissions, or external impact.

ARTIFACT LIMITS
- [SECTION / PAGE / WORD / SCREEN / DURATION LIMIT]
- Keep the result no larger than necessary to achieve the outcome.

ACCEPTANCE CRITERIA
- [CHECK 1]
- [CHECK 2]
- [CHECK 3]

Run the relevant deterministic checks once. Do not add redundant verifier
agents or repeat the same verification unless a check fails.

COMPLETION RESPONSE
Lead with the outcome in one sentence.
Then provide no more than three bullets:
- what changed;
- verification result;
- blocker or decision still required.

Verification Without Token-Wasting Theater

Control Subagents Explicitly

Paul focuses on the five most accessible rules, but Anthropic's guide adds an important sixth operational concern: Opus 5 delegates more readily than earlier models. Parallel work is useful only when tracks are genuinely independent and large enough to justify coordination.

Delegate only when the task contains independent, substantial workstreams that
can run in parallel.

Do not delegate:
- work that can be completed directly in a few tool calls;
- routine exploration;
- a second pass whose only purpose is to re-check your own work.

Use the fewest subagents needed. Keep each write scope separate and report the
result from every delegated task.

This keeps a simple website or report from becoming a costly agent swarm.

What Paul's Claude Design Test Proves

Paul gives Opus 5 a complete Northline Coffee homepage brief, attaches logos and product images, selects medium effort, and lets Claude Design build. The result arrives in under three minutes in his session, contains the six requested sections, uses the supplied assets, and returns the requested three-bullet summary.

That demonstrates strong instruction following in one configured environment. It does not establish that medium effort caused the speed, that the same prompt beats another prompting style, or that the design would outperform a baseline commercially. There is no repeated run, cost comparison, blinded review, accessibility audit, responsive test, or conversion data.

A Better Three-Run Evaluation

1. Choose one real task. Use a deliverable you can judge, not a toy prompt designed to look impressive.
2. Freeze the context. Use the same files, assets, tools, model version, and prompt for every run.
3. Run high, medium, and low effort. Do not change the prompt between settings.
4. Score before looking at cost. Grade completeness, correctness, scope control, taste, and acceptance checks.
5. Record rework. Count human correction minutes, not just first-output quality.
6. Calculate accepted-result cost. Include failed runs, delegated agents, retries, and review time.
7. Choose the lowest stable setting. Escalate effort only for the task classes that need it.

Evaluation scorecard
- Outcome completeness: 0-5
- Factual or technical correctness: 0-5
- Scope discipline: 0-5
- Deliverable usefulness: 0-5
- Acceptance checks passed: yes/no
- Human rework minutes:
- Total latency:
- Total tokens or credits:
- Accepted for use: yes/no

Video Chapters

Bottom Line

The best Opus 5 prompts are not longer for the sake of being longer. They are complete where it matters: outcome, context, scope, artifact limits, acceptance criteria, and completion response.

Paul J Lipsky's five-rule translation is useful because it removes common ambiguity. Anthropic's documentation adds the production nuance: old prompts remain a valid starting point, effort needs evaluation, subagents need limits, and self-correction does not replace external evidence.

Give Opus 5 enough information to own the job, enough boundaries to stop at the right place, and enough objective checks to prove the result. Then use the lowest effort level that consistently passes. That is more reliable than piling motivational phrases and repeated verification rituals onto the prompt.

Sources

• Paul J Lipsky: Anthropic Just Revealed How to Prompt Opus 5
• Paul J Lipsky on YouTube and Paul J Lipsky's official website
• Anthropic: Prompting Claude Opus 5
• Anthropic: Effort
• Anthropic: Prompting best practices
• Anthropic: Models overview
• Anthropic: Claude Platform release notes