How to Prompt Opus 5 for Marketing: Anthropic's Rules and 20 Copy-Paste Templates

Sources:

• YouTube: "How to Prompt Claude Opus 5" (knowledge work prompting guide for Opus 5, five rules with live demo)
• Article: "20 Claude Opus 5 Prompts Tested" (buildfast.withAI, July 27, 2026, 20 copy-paste templates with rubric)
• Anthropic official prompting guide: platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-opus-5

Opus 5 is better than Opus 4.8 in almost every category and sometimes better than Fable 5. If you prompt it the way you have been prompting previous Claude models, you will not get good results. Anthropic agrees with this. Their prompting guide for Opus 5 tells you to delete instructions from your existing prompts, not add them.

Four behaviours make Opus 5 different from all previous models:

1. It self-verifies without being asked: prompts containing "double-check your answer" or "re-verify before responding" now actively make output worse, not better. Opus 5 already does this and the instruction compounds into over-verification.
2. It delegates to subagents more readily: it coordinates well but three agents on a two-file task costs three times as much for the same result.
3. It talks more by default: written outputs run longer than any prior Opus, both in chat and in files written to disk.
4. It expands task scope on its own judgment: it will helpfully rewrite your architecture while fixing a typo unless you explicitly say what not to touch.

The single most valuable edit to an existing Opus 4.8 prompt library: deletion. Remove verification instructions. Remove legacy harness scaffolding. Add explicit scope constraints.

Model and effort selection: the decision you make first

Model routing from the video (simplified):

Plan	Default	Complex tasks
Claude Pro ($20)	Sonnet	Switch to Opus 5
Claude Max ($100+)	Opus 5	Stay on Opus 5, rarely need Fable

Fable 5 is now rarely necessary. Use it only when you are genuinely unhappy with Opus 5 results and want to compare. Haiku is not recommended for any real work.

Effort level: higher effort is not always better. Effort controls how much the model thinks, not the quality of the result. Anthropic states that low and medium effort are surprisingly strong on Opus 5.

Effort	When to use
Low	Pro plan, drafting, formatting, simple tasks
Medium	Max plan default for most GTM work
High	Complex analysis, multi-source research, agentic work
Extra high/Max	Hardest coding, frontier reasoning, only when high is not enough

Critical distinction: effort controls thinking depth, NOT visible response length. If your responses are too long, add a length instruction to the prompt. Dropping effort to shorten output is the most common and expensive mistake in Opus 5 migrations.

Sources: YouTube video (5 rules with live demo) + Anthropic prompting guide

These five rules apply everywhere you use Opus 5: Claude chat, Cowork, Claude Code, Claude Design, the API.

Rule 1: Give it the full job at the start

Old habit: break tasks into smaller steps. Ask Claude to look at files first, then build a brand guide, then plan the website.

Opus 5 habit: hand it the entire job from the start and let it run. Front-loading the complete specification is the documented condition for Opus 5's best performance. Every clarification turn re-reads context you already paid for, and the model was going to make a reasonable call anyway.

Before:

"Look at these images and create a brand guide for us."
// [next turn] "Now create a plan for what the website will look like."
// [next turn] "Now build the homepage."

After:

"Build a one-page website for Northline Coffee, a small brand that sells coffee beans online.
The goal is to help people understand the brand, choose a coffee, and place their first order.
Use the logo, colours, product details, and photos in the brand asset folder.
Build the finished website in the project folder."

Result from the live demo: completed in under 3 minutes because it knew the destination from the start. Did not ask follow-up questions because the prompt already answered them.

Rule 2: Tell it when the job ends and what not to touch

Opus 5 expands scope on its own judgment. It will helpfully restructure things you did not ask it to restructure. Fix this with an explicit scope contract: what to build, what not to touch, when to stop.

"Build only the homepage.
Link the buy buttons to the shop page that already exists.
Keep the current logo, colours, and product name.
Make small design choices yourself.
Only ask me a question if the answer would change the whole website."

The "out of scope" list is the highest-leverage addition to any Opus 5 prompt. If the model still expands scope after you add it, the list was too short, not ignored.

For marketing campaigns specifically: name what channels, formats, and assets are NOT in scope. Otherwise Opus 5 will produce the full multi-channel campaign when you only wanted the email sequence.

Rule 3: Limit the response length explicitly

Opus 5 talks more by default than any prior Opus. Outputs in chat run long. Files written to disk run long. Lowering effort will not fix this. Length is a prompt-side control.

"When you finish, tell me where the website is and give me
no more than three short bullets explaining what you made.
Then tell me only what's truly important to know about how the website works."

From the live demo: the model returned exactly three bullets. Nothing else. The user could read the summary instead of dismissing a wall of text.

For system prompts and long agent sessions, add the concision reminder near the end as well as the beginning:

"Keep responses focused, brief, and concise. Keep disclaimers and caveats short.
When asked to explain something, give a high-level summary unless depth is specifically requested."
// Near the end of a long system prompt:
"Keep outputs reasonably concise."

Rule 4: Set the size of the work product, not just the response

A short response does not mean a small deliverable. Separately constrain the size of what Opus 5 builds.

"Keep the website to six parts:
The top section, three featured coffees, why Northline is different,
customer reviews, common questions, and one final buy button.
Keep the writing short. The main headline should be eight words or less.
Each part should be easy to scan and have no more than 70 words."

You do not need to be this precise. Options:

• "Only have six sections. You decide what they are."
• "The whole website should be readable in under 90 seconds."
• "The email should be under 100 words."
• "The brief should fit on one page."

The point is giving Opus 5 a size boundary, not just a topic. Without one, it will build something bigger than you need.

Rule 5: Remove verification instructions from existing prompts

This is the rule that inverts years of advice. Opus 5 already verifies and fixes its own work as it goes. If you force it to do the whole check again at the end, it may use more time and tokens without improving the result.

Delete from your existing prompts:

// DELETE THESE
"After you finish, check every file twice."
"Find every possible mistake and fix all of them."
"Use a subagent to verify your work."
"Double-check your answer before responding."
"Include a final verification step."
"Re-verify before responding."

You can still tell Claude what good looks like. You just do not need to ask it to repeat the check it has already done.

The failure mode is silent: nothing errors, you just pay more tokens for the same result. That is why this is the highest-value single change in any Opus 4.8 prompt migration.

Source: "20 Claude Opus 5 Prompts Tested" (buildfast.withAI, July 27, 2026)

Production-ready templates for GTM and marketing work. Square brackets are your fill-ins. All templates are sourced from or validated against Anthropic's prompting guide.

Marketing content and writing templates

# THE VOICE LOCK (long-form content in a specific voice)
"Write [DELIVERABLE] on [TOPIC].

Voice specification:
- Audience: [who, and what they already know]
- Register: [e.g. direct, peer-level, no corporate hedging]
- Sentence rhythm: [e.g. mostly short, one long sentence per paragraph maximum]
- Banned: [phrases, clichés, structures you never want to see]
- Reference sample: [PASTE 150 WORDS OF THE TARGET VOICE]

Match the reference sample's voice, not its content.
Length: [N] words, and treat that as a ceiling."
# Why: a pasted voice sample outperforms any amount of adjective-stacking.
# Watch for: 'treat that as a ceiling' matters. A bare word count reads as a target.

---

# THE AUDIENCE LADDER (one message, three readers)
"Take [SOURCE CONTENT] and produce three versions:

1. For [EXPERT AUDIENCE]: assume full domain fluency, no definitions, lead with the finding
2. For [ADJACENT AUDIENCE]: knows the field but not this area; define jargon inline
3. For [EXECUTIVE AUDIENCE]: 120 words, decision-relevant only, no methodology

Same facts in all three. Do not soften the conclusion for the executive version."
# Why: producing all three in one pass keeps facts consistent across versions.
# Watch for: check version 3 has not lost a caveat version 1 treated as essential.

---

# THE CONCISION ENFORCER (system prompt fragment for any marketing agent)
"Keep responses focused, brief, and concise.
Keep disclaimers and caveats short, and spend most of the response on the main answer.
When asked to explain something, give a high-level summary unless
an in-depth explanation is specifically requested.

[near the end of a long system prompt, add:]
<tone_preference>
Keep outputs reasonably concise.
</tone_preference>"
# Why: Anthropic's own recommended pattern. The duplicated reminder near the end
# of a long system prompt is the part most people skip.
# Watch for: for files written to disk, add a separate instruction on document length.

Research, analysis, and strategy templates

# THE ADVERSARIAL SELF-CRITIQUE (for campaigns, strategies, or decisions)
"Here is my position: [POSITION].

Do three passes, clearly separated:

1. STEELMAN: the strongest version of my argument, stronger than I made it,
   using the best available evidence.

2. RED TEAM: the most serious objections. Not stylistic quibbles:
   the ones that would actually change the conclusion.

3. VERDICT: does the position survive? If it needs modification, give me
   the modified version. If it fails, say so plainly.

Do not soften the red team pass to be agreeable."
# Why: three-pass structure prevents the model collapsing into agreement.
# Opus 5 will hold a position under pressure rather than flattering you.
# Watch for: if the red team reads like polite hedging, the position was too vague to attack.

---

# THE DECISION MEMO (structured recommendation with visible uncertainty)
"Write a decision memo on [DECISION].

Structure:
- Recommendation, in one sentence, at the top
- The three factors that actually drive the decision
- The strongest case for the option you did NOT pick
- What you are uncertain about, and what evidence would change your recommendation
- What we should do in the next two weeks regardless of which option we pick

Keep it under [N] words. No executive-summary boilerplate."
# Why: answer-first structure suits a model that opens with process narration.
# The uncertainty section converts hedging from scattered prose into actionable list.
# Watch for: the word cap is load-bearing. Without it, Opus 5 memos run long.

---

# THE SOURCE-TIERED RESEARCH SYNTHESIS (competitive intel, market research)
"Research [TOPIC] and synthesise what you find.

Tier every source you use:
- Tier 1: primary sources (filings, papers, official docs, raw data)
- Tier 2: reputable secondary reporting
- Tier 3: commentary, forums, marketing

Structure the synthesis so Tier 1 claims carry the argument
and Tier 3 claims are labelled as opinion.
Where sources conflict, say so explicitly rather than picking one silently.
End with the three things you could not verify."
# Why: 'three things you could not verify' is the highest-value line.
# Converts silent gaps into an explicit list. Opus 5 populates it honestly.
# Watch for: give it real web search access. Without tools, source tiering produces
# confident-sounding fiction.

Agentic GTM workflow templates

# THE SCOPE CONTRACT (any autonomous run where you are not watching every step)
"Deliver what was asked, at the scope intended.
Make routine judgment calls yourself, and check in only when different readings
of the request would lead to materially different work.
If the request seems mistaken or a better approach exists, say so in a sentence
and continue with the task as asked rather than quietly narrowing, widening,
or transforming it.
Finish the whole task, and stop short of actions that are clearly beyond what was asked."
# Why: Anthropic's own recommended wording. Threads the needle:
# full completion, no expansion.
# Watch for: if the agent still over-reaches, the cause is usually a vague task
# description upstream, not a weak scope contract.

---

# THE DELEGATION BUDGET (when Opus 5 can spawn subagents)
"Delegate to a subagent only for large tasks that are genuinely independent
and parallelizable, such as a wide multi-file investigation.
Do not delegate work you can finish yourself in a handful of tool calls,
and do not use subagents to verify or double-check your own work.
If one subagent can complete the task, use one rather than several."
# Why: Opus 5 delegates more readily than any prior Claude.
# Verification delegation is where agent budgets quietly triple.
# Watch for: the 'do not use subagents to verify' clause is the expensive one.

---

# THE CHIEF-OF-STAFF LOOP (multi-hour autonomous marketing work)
"You are running [OBJECTIVE] over the next [TIME WINDOW].

Operating rules:
- Maintain a running state file at [PATH]: current status, decisions made,
  open questions, next actions
- Update it after every meaningful step, not every action
- Escalate to me only for judgment calls, irreversible actions,
  or when blocked for more than [N] attempts
- If you correct an earlier assumption, write the correction into the state file
  and move on

At each checkpoint: what changed, what you decided unilaterally,
and what needs my call."
# Why: the state-file pattern separates agents that survive long horizons from
# ones that drift. Opus 5 treats context as a living document.
# Watch for: set N explicitly. 'Escalate when blocked' without a threshold
# produces an agent that never asks or always asks.

---

# THE NARRATION CADENCE (for user-facing agent products)
"Before your first tool call, say in one sentence what you're about to do.
While working, give a brief update only when you find something important
or change direction.
When you finish, lead with the outcome: your first sentence should answer
'what happened' or 'what did you find,' with supporting detail after
for readers who want it."
# Why: Opus 5 narrates more than prior models.
# Positive examples beat prohibitions here.
# Watch for: if this is not enough, paste two sample updates in the style you
# want rather than adding more rules.

Data, slides, and document templates

# THE DECK FROM RAW NOTES (meeting notes to presentation)
"Turn these notes into a [N]-slide deck for [AUDIENCE].

Rules:
- One idea per slide; the slide title states the idea as a claim, not a topic
- Body: maximum three bullets, maximum twelve words each
- Follow this template/style: [describe or attach it]
- Include a 'so what' slide after every section
- Speaker notes: what I say out loud, not a restatement of the slide

Flag any slide where the notes did not give you enough substance
rather than padding it."
# Why: title-as-claim is the highest-leverage constraint in deck generation.
# The padding flag exploits Opus 5's willingness to admit gaps rather than
# fill them with filler.
# Watch for: iterate in the same session rather than starting fresh.
# Build-then-revise is where testers reported the biggest gains.

---

# THE NUMBERS AUDIT (pipeline data, campaign results, financial models)
"Audit the numbers in [SOURCE].

For each figure that matters:
- Restate the calculation in words
- Recompute it independently and show the result
- Flag any discrepancy between stated and recomputed values
- Flag any figure that depends on an assumption not stated in the document

Present as a table: figure, stated value, your value, status, note.

Do not fix anything yet. Audit only."
# Why: independent recomputation catches transcription errors a read-through misses.
# 'Audit only' is deliberate: let it fix things in the same pass
# and you lose the ability to check its arithmetic against yours.

---

# THE MULTI-SHEET MODEL (revenue models, campaign trackers)
"Build a [MODEL TYPE] spreadsheet.

Structure:
- Sheet 1 'Inputs': every assumption, one per row, labelled and unit-tagged.
  Nothing hardcoded anywhere else.
- Sheet 2 'Calculations': formulas referencing Inputs only
- Sheet 3 'Outputs': the summary view, formatted for reading
- Sheet 4 'Checks': at least [N] validation formulas that go TRUE
  when the model is internally consistent

Every formula must reference cells, never repeat a literal value.
Tell me which assumptions the output is most sensitive to."
# Why: the 'Checks' sheet makes a generated model auditable instead of merely plausible.
# Watch for: open the file and change one input.
# If downstream numbers do not move, something was hardcoded despite the instruction.

Source: buildfast.withAI article (July 27, 2026) + Anthropic prompting guide

Five habits that were harmless or helpful on Opus 4.8 and now cost you quality, money, or both.

The five prompt anti-patterns to retire

1. Verification loops: "double-check your answer", "re-verify before responding", "use a subagent to verify" — these compound with behaviour Opus 5 already performs and add token cost for no quality gain. Delete them.
2. Conservative review instructions: "only report high-severity issues", "be conservative", "just flag the main problems" — Opus 5 follows these literally and under-reports. It is a high-precision reviewer; its extra findings are mostly genuine. Ask for everything and filter afterwards.
3. Using effort to control length: effort governs thinking depth, NOT visible response length. Dropping from high to medium to save tokens will not shorten your output. Write a length instruction instead.
4. Disabling thinking: thinking: { type: "disabled" } now returns a 400 error. Do not pass this parameter at all. Two artifacts appear when thinking is disabled at effort high or above: the model writes tool calls into visible text instead of emitting real tool-use blocks, and internal XML tags can leak into the response. Anthropic's recommendation: keep thinking on and lower effort instead.
5. Drip-feeding requirements across turns: giving Opus 5 one piece of context at a time is the expensive path. Front-load the full specification. Every clarification turn re-reads context you already paid for.

If you need to run with thinking disabled and cannot avoid it:

# This mitigates the tool-call and XML-leak artifacts
"When you use a tool, you may say a brief sentence first.
If no tool can express what the user asked for, say so instead of guessing.
Do not include internal or system XML tags in your response."
# Note: naming the tags specifically works worse than this general form.

One honest caveat

From the buildfast.withAI article, not buried: Anthropic's headline Frontier-Bench figure used Opus 4.8 as a fallback whenever a safety classifier refused a request, and the substitution rate was not published. The number is provisional.

Direct consequence for GTM teams: on Claude.ai, Claude Code, and Cowork, requests flagged by safety classifiers fall back to Opus 4.8 by default. A prompt that trips a classifier may be quietly answered by a different model than the one you selected. If your prompts touch security research, competitive intelligence, or any dual-use area, verify which model actually responded.

Also from Dan Shipper's early access testing (quoted in the Agent Native video): Opus 5 argues with instructions and does not play well with existing skills and plugins on first use. If you hit this, delete existing skills and rebuild them from scratch for Opus 5. The effort of rebuilding is real. The failure mode is silent: nothing errors, the model just costs more for worse results.

All key rules, templates, and settings in one place.

The 5 rules: one-line version

# Rule 1: Give it the full job at the start
# Do not break tasks into steps. Front-load the entire specification.

# Rule 2: Tell it when the job ends and what not to touch
# Explicit scope: what to build + what NOT to build + when to stop.

# Rule 3: Limit response length in the prompt
# 'No more than 3 bullets.' 'Keep it under N words.'
# Effort does NOT control visible length.

# Rule 4: Set the size of the work product separately
# 'Six sections only.' 'Readable in under 90 seconds.' 'One page.'

# Rule 5: Remove verification instructions
# DELETE: 'double-check your answer', 're-verify', 'use a subagent to verify'
# Opus 5 already does this. These lines add cost, not quality.

Settings quick reference

# Model
claude-opus-5

# Effort starting points
# Low:   Pro plan, drafting, simple formatting, batch output
# Medium: Max plan default for most GTM work (RECOMMENDED DEFAULT)
# High:   Complex research, competitive analysis, long agent runs
# Max:    Frontier reasoning only when high is not enough

# DO NOT use effort to control response length
# DO NOT pass: thinking: { type: "disabled" }  # Returns 400 error
# DO omit the thinking parameter entirely if you do not need to configure effort

# Fast mode
# 2.5x speed, 2x price
# Use when a human is watching a spinner. Not for batch work.

# Fallback behaviour
# On Claude.ai, Claude Code, and Cowork:
# classifier-flagged requests fall back to Opus 4.8 automatically
# Enable automatic fallbacks on API via:
# anthropic-beta: "automatic-fallbacks-2026-07-24"

The 5 prompt patterns that separate good Opus 5 prompts from bad

From the buildfast.withAI analysis of all 20 templates:

1. Front-load the entire specification: hand it the complete task and let it run. Steering is the expensive path.
2. Delete, do not add: strip verification instructions, legacy scaffolding, and any harness that adds a separate verification pass.
3. Constrain scope explicitly, not implicitly: "fix the login bug" is an invitation. "fix the login bug; do not touch the session module, do not add tests to unrelated files" is a contract.
4. Ask for everything, filter in a second pass: any prompt containing "only", "just", or "be conservative" is likely producing less than you wanted. Split into an exhaustive pass then a filtering pass.
5. Control length in words, cost in effort: two separate dials. Effort is your budget. The prompt is your editor. Confusing the two is how teams pay more for output they still have to trim.

Template index: which template for which GTM task

Task	Template	Effort
LinkedIn posts, emails, campaign copy	The Voice Lock	Low
Investor deck, QBR, pitch deck	The Deck From Raw Notes	Medium
Messaging for multiple audiences	The Audience Ladder	Medium
Competitive intelligence	The Source-Tiered Research Synthesis	High
Campaign strategy or positioning decision	The Decision Memo	Medium
Stress-testing a campaign idea	The Adversarial Self-Critique	Medium
Pipeline or campaign results audit	The Numbers Audit	Medium
Revenue model or campaign tracker	The Multi-Sheet Model	Medium
Long-horizon agent (outbound, content, research)	The Chief-of-Staff Loop	High
Any agent that might over-reach	The Scope Contract	Medium
Agent that spawns subagents	The Delegation Budget	Medium
User-facing agent product	The Narration Cadence	Medium
Any system prompt that needs shorter outputs	The Concision Enforcer	Any