The Best Claude Prompts for 2026 — Calibrated for Opus 4 & Sonnet 4.5
Claude responds to prompts differently than ChatGPT. It honors XML structure more strictly, follows multi-step reasoning more reliably, and writes prose that's emotionally legible in ways GPT-5 sometimes flattens. The Originals below are hand-crafted with these differences in mind. Each one is tested on Opus 4 and Sonnet 4.5; each ships with a complete example output. Free, MIT-licensed.
Last reviewed: April 2026 · Tested on Claude Opus 4 + Sonnet 4.5.
Why Claude rewards different prompt design than ChatGPT
Five differences matter when you're crafting Claude prompts:
- XML structure is non-optional, not optional. Claude's tokenizer + post-training make XML tags a strong signal. Markdown headers in long prompts often produce noisier output.
- Long-context reasoning is unusually strong. Opus 4's 1M-token context with prompt caching means you can paste a 200-page document and ask structural questions. ChatGPT can do this but with more drift.
- Tone-following is more reliable. If you tell Claude "be direct, do not use platitudes," it will follow. ChatGPT often defaults back to soft hedging.
- Refusals are more thoughtful. Claude will tell you when a request is genuinely unsafe; it won't refuse for cosmetic reasons. This makes it ideal for grief, recovery, and difficult-life writing prompts.
- Code reasoning depth. For complex codebase-level tasks (refactoring, architecture review), Opus 4 outperforms GPT-5 in our internal tests.
Top 12 Claude Originals (calibrated for Opus 4)
These are the Originals where Claude's particular strengths shine — long-context reasoning, prose precision, tonal consistency, structured analysis.
Best for: by use case
- Fiction writing & long-form prose — Claude's tonal precision is unmatched. See: Romance Meet-Cute Architect, Memoir Vulnerability Calibrator, Character Design Deep Sheet.
- Complex codebase analysis — Opus 4 + 1M context = read entire repos. See: Codebase Archaeologist, Code Review Teacher.
- High-stakes decisions — Claude's nuanced reasoning suits judgment calls. See: Life Decision 5-Scenario Forecaster, Quit-or-Stay Calculator.
- Emotional / difficult writing — Claude doesn't flatten emotion. See: Eulogy Writer, Grief Processing Framework.
- Devil's advocate reasoning — Claude steelmans without hedging. See: Steelman Devil's Advocate, Pre-Mortem for Major Project.
How to use these prompts in Claude
- Choose Opus 4 for complex tasks. Sonnet 4.5 for high-volume / fast-iteration tasks.
- Enable prompt caching if you'll re-run the same prompt structure with different inputs (saves 90% on tokens).
- Use the auto-intake feature. Each Original asks for missing fields rather than guessing.
- For the API: paste the full XML prompt into the system message, then the user message contains your specific input.
- For Claude.ai: paste the full prompt in the message box. Long prompts work fine — Claude doesn't truncate.
FAQ
Do I need Opus 4, or will Sonnet work?
Sonnet 4.5 handles 80-90% of the Originals at near-Opus quality, much faster, much cheaper. Save Opus for: long-form fiction, codebase-level reasoning, multi-step strategic decisions. Each Original lists its recommended model.
How do these compare to ChatGPT prompts?
Some Originals work better in Claude (anything tone-sensitive or long-form). Some work better in ChatGPT (anything requiring rapid iteration or web-search integration). Most are model-agnostic. See: Best ChatGPT Prompts for the parallel list.
Can I use these via Claude API?
Yes. The XML structure is API-friendly. For high-volume use, set up prompt caching — Anthropic discounts cached prompts by 90%. Each Original works as a system prompt with the user's input as the user message.
What about Claude Code or Claude in IDEs?
The coding-related Originals (Codebase Archaeologist, Code Review Teacher) work in Claude Code, Cursor, Zed, and other Claude-powered IDEs. Paste them as system instructions or context.
More from Promptolis
Get new Originals every Friday
2-3 hand-crafted Originals per week. No spam, no upsells, unsubscribe in 1 click.