Best Claude Prompts: A Copy-Paste Library for Opus 5, Sonnet 5 and Haiku 4.5

Anthropic rearranged its model range three times in under a month. Claude Sonnet 5 shipped on June 30, 2026 with introductory API pricing. Claude Fable 5 came back on July 1 after nineteen days offline under US export controls, then left Anthropic's subscription plans on July 19 to become pay-per-use. And on July 24, 2026, Anthropic introduced Claude Opus 5, the successor to Opus 4.8 in the Opus line. If you keep a Claude prompt library, this is the week to bring it up to date: the family-5 models follow instructions more literally, reason adaptively by default, and respond worse to overloaded prompts than their predecessors did.

This guide covers the range as it stands on July 24, 2026, the prompting techniques Anthropic currently documents, guardrails against hallucinated facts, and eight prompts you can paste into a chat today. Everything here works in the Claude app, over the API, and in ilisai, where Claude Opus 5, Claude Sonnet 5 and Claude Haiku 4.5 are available on every plan with credits included — no Anthropic account, no API key.

The Claude range in July 2026

As of July 24, 2026, Anthropic's line-up runs from Claude Haiku 4.5 at the entry level to Claude Fable 5 at the top, with Claude Sonnet 5 as the default recommendation for most work. The newest member is Claude Opus 5, introduced on July 24, 2026 as the successor to Opus 4.8. ilisai serves three of the four — Opus 5, Sonnet 5 and Haiku 4.5 — on every plan; as of the same date, Opus 5 replaces Opus 4.8 in the ilisai catalog.

Prices checked on July 24, 2026 against Anthropic's official model documentation. Two footnotes matter for real costs. Claude Sonnet 5 runs on introductory pricing through August 31, 2026 and moves to its standard rate on September 1. And the family-5 models (plus Opus 4.7 and later) use a new tokenizer that produces roughly 30% more tokens for the same text, so per-token prices do not compare one-to-one with older models. On the plus side, Fable 5, Opus 5 and Sonnet 5 include their full 1M-token context window at flat pricing — a 900k-token request bills at the same per-token rate as a 9k one. On Opus 5, the 1M window is both the default and the maximum.

Fable 5: back online, out of subscriptions on July 19

Claude Fable 5 has been available again since July 1, 2026, after Anthropic disabled it worldwide in June to comply with a US export-control directive. The next change was commercial rather than regulatory: July 19, 2026 was the last day it was included in Anthropic's paid plans, and since July 20 every request is billed through prepaid usage credits.

The timeline, dated:

• June 12, 2026 — the US government orders Anthropic to restrict Fable 5 and Mythos 5 for foreign nationals. Unable to verify nationality per request, Anthropic disables both models for everyone. We covered the episode and what it means for Europe in our analysis of the block.
• June 30, 2026 — Anthropic announces the export controls have been lifted.
• July 1, 2026 — Fable 5 is served again, included in the Pro, Max, Team and premium Enterprise plans at up to 50% of weekly usage limits.
• July 19, 2026 — last day of included access, after Anthropic extended the window from July 7, then July 12 (confirmed by Anthropic's announcements and press coverage from July 12–13, 2026).
• July 20, 2026 — Fable 5 moves to prepaid usage credits at $10 input / $50 output per million tokens. Anthropic says it intends to return the model to subscriptions once it has enough compute, without committing to a date.

Claude Mythos 5, the other model named in the June directive, remains in limited availability: access has been restored for a set of approved US organizations through Anthropic's Glasswing program since June 26, 2026.

The status marker at the top of this page records that Fable 5 has been available again in Europe since July 1, 2026.

To be plain about our own product: Fable 5 is not in the ilisai catalog. The Claude models you can use here are Opus 5, Sonnet 5 and Haiku 4.5 — on the Free plan as much as the Starter and Pro plans, with credits included. The largest jobs may need wallet top-ups. If you want Fable 5 specifically, it runs on Anthropic usage credits since July 20, 2026.

How to write prompts for Claude in 2026

A good Claude prompt still has five parts: a role relevant to the task, the context that frames it, the instruction with its scope spelled out, one or two examples when format matters, and the expected output format. XML tags remain the most reliable way to separate those parts, because Claude reads them as semantic boundaries. What changed in 2026 is the dosage — the family-5 models do better with short, clear instructions than with rulebooks.

1Act as [a role relevant to the task].
2
3<context>
4[Who the result is for, what it must achieve, what constraints apply]
5</context>
6
7<instructions>
8[Exactly what to do, with the scope made explicit]
9</instructions>
10
11<examples>
12[1–3 examples of the expected result, if format matters]
13</examples>
14
15<output_format>
16[The exact structure of the answer]
17</output_format>

What Anthropic recommends for Claude Sonnet 5

Anthropic's prompting guide for Claude Sonnet 5 (accessed July 17, 2026) documents several behavior changes worth knowing:

• It follows instructions literally. Sonnet 5 does not silently generalize a rule from one item to the next. If you want something applied everywhere, say so: "apply this formatting to every section, not just the first one."
• Steer length with positive examples. The model calibrates response length to task complexity on its own. When you need a specific style or length, examples of good output beat lists of prohibitions.
• Drop the "use extended thinking" incantations. Reasoning is adaptive and on by default; on the API, manual thinking budgets are gone and the effort parameter is the lever. In a chat, if you see shallow reasoning on a hard task, add: "this task involves multistep reasoning — think it through carefully before responding."
• Let it check its own work. Sonnet 5 is the most agentic Sonnet so far and reaches for self-verification loops more readily than earlier models before calling a task done. Scaffolding like "summarize progress every three steps" is usually removable now.
• Style is steered in words. The API rejects temperature and other sampling parameters on Sonnet 5; tone and variety are requested in the prompt or system instructions.

What Anthropic recommends for Claude Opus 5

Claude Opus 5 arrived on July 24, 2026, and Anthropic's prompting guide for Claude Opus 5 plus its migration guide (both accessed July 24, 2026) describe a model built to carry large jobs end to end. Reasoning is on by default — the model decides when and how much to think. In a chat, five habits matter:

• Hand over the whole task at once. Opus 5 does its best work on a complete, well-specified brief it can run with, and worse when instructions arrive drip-fed across messages. Front-load the goal, the constraints, the materials and the expected format, then let it work.
• Drop the "double-check your work". It verifies its own output before answering. Verification instructions inherited from older prompts cause over-verification: longer answers and extra effort with no gain in quality.
• Ask for exactly the scope you want. It follows instructions literally and keeps to the scope you state — and it can also widen a task on its own judgment. Say explicitly whether it should stay narrow or go beyond the ask.
• Prompt for brevity. Its default answers and written documents run longer than Opus 4.8's. When you want short output, one explicit line is enough: "keep the response focused and brief."
• For reviews, ask for everything. Have it report every finding with a confidence and severity label, then filter yourself. "Only report the important issues" is obeyed literally, and real findings drop out.

Two more things to know. Prompts written for Opus 4.8 or Sonnet 5 carry over as they are — Opus 5 uses the same tokenizer as Opus 4.8 and, per Anthropic, performs well out of the box on existing Opus 4.8 prompts — so the only rewriting worth doing is deleting those verification lines. And the model ships with reinforced cybersecurity safeguards: it can decline requests it classifies as high-risk security work.

What changes with Claude Fable 5

The prompting guide for Claude Fable 5 (accessed July 17, 2026) describes a model built for tasks that used to be too long or too ambiguous, which changes how you talk to it:

• One brief instruction beats an exhaustive list. Instruction-following improved enough to steer behavior with a sentence. Anthropic warns that prompts and skills written for earlier models are often too prescriptive for Fable 5 and can degrade output — audit what you inherit before reusing it.
• Give the reason, not only the request. The documented template is simple: "I'm working on [the larger task] for [who it's for]. They need [what the output enables]. With that in mind: [request]."
• State the boundaries. Fable 5 can take initiative nobody asked for, like drafting an email you never requested. Set explicit limits: "when I describe a problem, the deliverable is your diagnosis; don't apply changes until I ask."
• Make it ground its progress claims. On long tasks, instruct it to audit each claim against actual results from the session. Anthropic reports this nearly eliminated fabricated status reports in its testing.
• Expect longer turns. A single request on a hard problem can run for many minutes. If it overplans, tell it to act as soon as it has enough information.
• Don't ask it to transcribe its reasoning. Fable 5 runs safety classifiers that can refuse requests to reproduce its internal reasoning in the response. Ask for conclusions and justification instead.

Guardrails: fewer hallucinations, steadier answers

The family-5 models hallucinate less than their predecessors — Anthropic's Sonnet 5 announcement (June 30, 2026) cites lower rates of hallucination and sycophancy than Sonnet 4.6. Less is not zero, so verification techniques remain part of the job whenever you publish facts or decide with the output.

From Anthropic's guide to reducing hallucinations (accessed July 17, 2026):

• Give it permission not to know. A line as simple as "if you don't have enough information, say you can't assess this with confidence" sharply cuts invented answers.
• Verbatim quotes first. For long documents (roughly 20,000 tokens and up), ask it to extract relevant literal quotes before analyzing anything, then to base the analysis only on those quotes.
• Verify and retract. After generating a text with claims in it, ask it to find a supporting quote for each claim and withdraw the ones it can't back, marking the gap.
• Restrict it to the provided material. "Do not use your general knowledge" cuts unwanted blends between your data and its training.

Anthropic's consistency guide adds the structural side: define the exact output format, show one complete example rather than describing the result in the abstract, and chain small consistent prompts instead of one mega-prompt. One migration note from the same documentation: recent models — Fable 5, Opus 5, Opus 4.8, Sonnet 4.6 — no longer support prefilling the assistant turn to force a format; use system instructions or the API's structured outputs instead.

Eight prompts for real work in a chat

Each prompt below is adapted to chat use — writing, analysis, code and research — and applies the techniques above. Paste one, replace the brackets, send. They work in ilisai with any of the three Claude models it serves, in the Claude app, and over the API.

1. From rough notes to a finished text

1Act as a senior communications editor.
2
3<context>
4I need [an email / an announcement / a short report] for [audience].
5Its job: [what the text must achieve].
6Tone: [two or three adjectives, or paste a sample you like].
7</context>
8
9<notes>
10[Paste your rough notes, bullet points, or a voice-note transcript]
11</notes>
12
13<instructions>
14Write the full text from these notes. Use only the facts in <notes>; if something essential is missing (a date, a number, a name), write [MISSING: what] instead of guessing.
15Keep it as short as the content allows.
16</instructions>
17
18<output_format>
191. Subject line or title (3 options).
202. The full text.
213. Anything you marked as missing.
22</output_format>

Why it works. Positive tone guidance instead of prohibitions, and explicit permission to leave gaps rather than fill them — the cheapest anti-hallucination guardrail there is.

2. Editing a draft against a voice

1Act as the in-house editor for [brand or project].
2
3<voice>
4This is how we write (real samples):
5- "[sentence 1]"
6- "[sentence 2]"
7We never publish sentences like:
8- "[counter-example]"
9</voice>
10
11<draft>
12[Paste the draft]
13</draft>
14
15<instructions>
16Edit the draft to match the voice. Apply the edit to the whole text, not only the opening paragraphs.
17Keep every fact and figure as it is; if a claim looks doubtful, mark it [CHECK] rather than rewriting or deleting it.
18Change only what pulls the text away from the voice.
19</instructions>
20
21<output_format>
221. The edited text in full.
232. A short list of the substantive changes and why.
24</output_format>

Why it works. Real voice samples teach style better than abstract adjectives, and the explicit scope ("the whole text, not only the opening paragraphs") plays to Sonnet 5's literal instruction-following.

3. Analyzing metrics without inventions

1Act as a data analyst.
2
3<data>
4[Paste the exported table: dimensions, metrics, period]
5</data>
6
7<context>
8This data comes from [tool] for [product or campaign]. The threshold that matters: [target CPA, conversion rate, churn ceiling…]. I need to decide [the decision] this week.
9</context>
10
11<instructions>
121. Evaluate the numbers against the threshold using only the data in <data>. Do not mix outside benchmarks into this evaluation or the recommendation; if an external benchmark is worth mentioning, put it in a separate note labeled "external reference, verify" that the recommendation must not rest on.
132. Point out anomalies or gaps before drawing conclusions. If something cannot be answered from this data, say exactly what is missing.
143. Recommend a course of action with concrete figures and a two-sentence rationale.
15</instructions>
16
17<output_format>
181. Three-sentence summary.
192. Findings, point by point.
203. Recommendation.
214. Missing data.
22</output_format>

Why it works. Restricting the analysis to the provided material is the documented technique against blended-in benchmarks, and the fixed output format makes runs comparable week to week.

4. Long documents, verbatim quotes first

1Act as an analyst preparing a briefing.
2
3<document>
4[Paste the report, contract, or transcript — or attach the file]
5</document>
6
7<instructions>
8Work in two phases.
9
10Phase 1. Extract the verbatim quotes most relevant to this question: [your question]. Number them. If there are no relevant quotes, write "No relevant quotes found" and stop.
11
12Phase 2. Answer the question using only the numbered quotes, citing them by number. Add nothing that is not in the document. If part of the question remains unanswered by the document, say which part.
13</instructions>

Why it works. This is the quote-grounding technique Anthropic documents for documents past roughly 20,000 tokens: the model has to show where every conclusion comes from.

5. Code review with full coverage

1Act as a code reviewer.
2
3<code>
4[Paste the diff, function, or file]
5</code>
6
7<context>
8This code [what it does] and runs on [runtime / framework / versions]. The change is meant to [intent of the change].
9</context>
10
11<instructions>
12Report every issue you find, including ones you are uncertain about or consider low-severity. Do not filter for importance; I will do that. For each finding give: the line or block, what is wrong, why it matters, a suggested fix, and your confidence (high / medium / low).
13Cover correctness first, then edge cases, then readability.
14</instructions>

Why it works. Sonnet 5 follows severity filters more faithfully than earlier models, which can silently drop real findings. Anthropic's own guidance for review prompts is to ask for coverage and filter afterwards.

6. Debugging with a boundary

1Act as a debugging partner.
2
3<problem>
4Expected: [what should happen]
5Actual: [what happens, with the exact error message or output]
6Environment: [language, framework, versions]
7Recent changes: [what changed before it broke, if known]
8</problem>
9
10<instructions>
11Your deliverable is a diagnosis. Propose the two or three most likely causes ranked by probability, and for each one give a check I can run to confirm or rule it out.
12Do not rewrite my code until I confirm a cause.
13If the information above is not enough to rank causes, ask for the specific missing piece.
14</instructions>

Why it works. It draws the boundary Anthropic recommends for models with initiative: diagnosis as the deliverable, changes only on request — and it forces evidence before action.

7. Research with source discipline

1Act as a research analyst with web access.
2
3<context>
4I am evaluating [decision]. The result goes to [who] to decide [what].
5</context>
6
7<instructions>
8Research [specific question] with the tools available. When you have enough information to move forward, move forward.
9
10Before you finish, audit your own report: every claim must trace back to a source consulted in this session. Flag anything you could not verify instead of presenting it as fact.
11</instructions>
12
13<output_format>
141. Direct answer in 3–4 sentences.
152. Evidence, with the source for each point.
163. What remains unverified, and how to verify it.
17</output_format>

Why it works. It pairs agentic autonomy (move forward without asking on reversible steps) with the evidence audit Anthropic documents against fabricated progress claims.

8. Fact-checking a draft before it ships

1Act as a skeptical fact-checker.
2
3<draft>
4[Paste the text to verify]
5</draft>
6
7<sources>
8[Paste the sources it should rest on]
9</sources>
10
11<instructions>
121. List every verifiable claim in the draft (figures, dates, names, facts).
132. For each claim, find a verbatim quote in <sources> that supports it.
143. Classify each claim: SUPPORTED (with the quote), UNSUPPORTED, or CONTRADICTED (with the contradicting quote).
154. Return the corrected draft: mark unsupported claims [NO SOURCE]. Fix a contradicted claim only when <sources> supports a replacement; if the sources only disprove it, remove it or mark it [NO SOURCE] instead of inventing a correction.
16Plausibility does not count as support; only the provided sources do.
17</instructions>

Why it works. Claim-by-claim verification with retraction is the most direct guardrail Anthropic documents for content with facts in it — the exact failure mode that costs the most in public.

Which Claude model should you pick?

• Claude Sonnet 5 by default. It handles most work at a lower price than Opus, tends to verify its own work before calling a task done, and its introductory API pricing holds through August 31, 2026. Start every writing, analysis or research task here.
• Claude Opus 5 for the frontier pick in ilisai. Introduced on July 24, 2026 as the successor to Opus 4.8 at the same API price, with more headroom than Sonnet 5 for deep reasoning, long tasks and demanding code.
• Claude Haiku 4.5 for volume. Classification, extraction, short answers, anything repetitive where cost per request rules.
• Claude Fable 5 for frontier work — outside ilisai. Long, ambiguous, high-value problems. Included in Claude's paid plans through July 19, 2026; usage credits after that.

The credits behind these models cover the rest of ilisai too, the AI image generator included. And if your prompt library reaches beyond chat, the image and video prompt gallery collects real generations with the exact prompt behind each one.

Frequently asked questions

How do I write a good prompt for Claude?

Give it a relevant role, context (who the result is for and what it must achieve), the instruction with explicit scope, and the output format you expect. With the family-5 models, one short clear instruction outperforms an exhaustive rulebook, and positive examples steer style better than prohibitions.

Can I use Claude Fable 5 in ilisai?

No. The Claude models in ilisai are Claude Opus 5, Claude Sonnet 5 and Claude Haiku 4.5, available on every plan with credits included. Fable 5 runs on Anthropic usage credits since July 20, 2026.

What happens to Claude Fable 5 on July 19, 2026?

It was the last day Fable 5 was included in Anthropic's Pro, Max, Team and premium Enterprise plans, capped at 50% of weekly limits. Since July 20, all Fable 5 use is billed through prepaid usage credits at $10 input / $50 output per million tokens. Anthropic has said it wants to return the model to subscriptions when it has the compute, without committing to a date.

What does the Claude API cost right now?

As of July 24, 2026, per million tokens (input / output): Claude Fable 5 at $10 / $50, Claude Opus 5 at $5 / $25 (the same rate Opus 4.8 had), Claude Sonnet 5 at $2 / $10 (introductory pricing through August 31, 2026, then $3 / $15), and Claude Haiku 4.5 at $1 / $5.

Does Claude work well in languages other than English?

Yes. In Anthropic's multilingual evaluation (MMLU translated by professional human translators, accessed July 17, 2026), Claude Sonnet 4.5 scores 98.2% of its English performance in Spanish and 97.5% in French. Declare the output language explicitly and ask for idiomatic, native-speaker phrasing when the text will be published.

Where can I use Claude without an Anthropic account?

In ilisai: Claude Opus 5, Claude Sonnet 5 and Claude Haiku 4.5 run on every plan — the Free plan included — with credits included and no API key to configure. The largest jobs may need wallet top-ups.

Do my Claude Sonnet 4.5 prompts still work?

Mostly, yes. XML structure, examples, role and context all carry over. Three things deserve a review: heavily prescriptive prompts can degrade output on the family-5 models, "use extended thinking" requests are no longer the right lever, and the more literal instruction-following means scope needs to be spelled out.

Sources

• Anthropic — Introducing Claude Sonnet 5 (June 30, 2026)
• Anthropic — Redeploying Claude Fable 5 (June 30, 2026)
• Anthropic — Introducing Claude Opus 5 (July 24, 2026)
• Anthropic docs — Models overview (accessed July 24, 2026)
• Anthropic docs — Prompting Claude Opus 5 (accessed July 24, 2026)
• Anthropic docs — Migrating to Claude Opus 5 (accessed July 24, 2026)
• Anthropic docs — Pricing (accessed July 17, 2026)
• Anthropic docs — Prompting Claude Sonnet 5 (accessed July 17, 2026)
• Anthropic docs — Prompting Claude Fable 5 (accessed July 17, 2026)
• Anthropic docs — Reduce hallucinations (accessed July 17, 2026)
• Anthropic docs — Increase output consistency (accessed July 17, 2026)
• Anthropic docs — Multilingual support (accessed July 17, 2026)
• BleepingComputer — Claude Fable 5 stays free for paid users until July 19 as Anthropic buys more time (July 12, 2026)