Last reviewed:
What are prompting patterns? Definition and right instinct vs wrong instinct
Prompting patterns are formulation conventions that structure a prompt to orient the model's output. They do not exist as native LLM commands, only as guides for the writer. The distinction matters: no pattern triggers a feature, the effect comes from linguistic precision.
You may have come across graphics listing 80 or 90 "Claude commands". These commands do not exist technically: the Claude model recognises no native commands (the slash commands in Claude Code are features of the tool, not of the model). They are prompting patterns, writing conventions that structure the request without triggering a specific feature on the model side. The distinction matters. A pattern is not an instruction interpreted by the model as a technical directive, it is a linguistic convention applied by the user to clarify intent. The real value of patterns is cognitive: they help the writer structure thinking before prompting (role, context, constraints, expected reasoning), and statistically, better-structured prompts produce better-targeted outputs. No magic on the model side, just writing discipline on the human side. The WYP grid retains 12 recurrent patterns selected for executive utility: assigned role, input context, output constraints, chain-of-thought, few-shot, refuse to invent, sources requested, adversarial critique, reformulation, structured comparison, listed assumptions, pre-mortem. Some patterns age with the models: on current Claude models, Anthropic's documentation recommends relying on built-in thinking (adaptive thinking) rather than a hand-written chain-of-thought, and advises 3 to 5 varied examples, wrapped in XML tags.
Concrete example
Illustrative case. A French mid-cap consulting firm CEO (180 employees) compares in 2026 two prompting approaches on 20 successive requests to Claude (competitive briefings). First try: vague prompts like "give me a summary of market X". Average output rated useful 4/10 on the internal scoring grid. Second try: systematic application of 4 combined patterns (assigned role "senior analyst", one-page context, output constraints three columns plus 300 words, one example of the expected output). Average output 7.5/10 on the same grid. Difference not linked to the model (identical), only to writing discipline. Prompt drafting time: 3 minutes instead of 30 seconds, i.e. 2 minutes 30 additional for an 87.5% gain in the average score (from 4 to 7.5).
FAQ
What is a prompting pattern?
It is a formulation convention that structures a prompt to orient the model's output. A pattern is not a native LLM command nor a technical directive: it is a linguistic convention applied by the writer to clarify intent.
Do the 80 or 90 Claude commands from the graphics really exist?
No, they do not exist technically: the Claude model recognises no native commands (the slash commands in Claude Code are features of the tool, not of the model). They are prompting patterns, writing conventions that structure the request without triggering any feature on the model side.
What are patterns really for?
Their value is cognitive: they help the writer structure thinking before prompting (role, context, constraints, expected reasoning), and statistically, better-structured prompts produce better-targeted outputs. In the example of a consulting mid-cap, moving from vague prompts to four combined patterns raises average quality from 4 to 7.5 out of 10, with the same model.
How many patterns should we adopt?
Five to eight recurring patterns, industrialised in shared templates, not an unmanageable and half-redundant list of 89. Have each pattern practised at least twenty times per user: the discipline of application matters more than the number of patterns known.
How do we avoid over-prompting?
Prioritise patterns with cognitive value (role, context, constraints, examples, refuse to invent) over decorative ones, and document where each pattern is worthwhile and where it adds nothing. Measure the quality gap before and after on a sample of real requests.
See also
Further reading
Prompting best practices, official Anthropic documentation, 2026
Sources
- Prompting best practices, official Anthropic documentation, 2026. https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/claude-prompting-best-practices
- Chain-of-Thought Prompting Elicits Reasoning in Large Language Models, Wei, Wang, Schuurmans, Bosma, Ichter, Xia, Chi, Le, Zhou, arXiv 2201.11903, 2022 (revised January 2023). https://arxiv.org/abs/2201.11903
- Language Models are Few-Shot Learners, Brown et al., arXiv 2005.14165, 2020 (reference GPT-3 paper introducing few-shot learning through prompting). https://arxiv.org/abs/2005.14165
- Prompt engineering guide, OpenAI documentation, 2026. https://developers.openai.com/api/docs/guides/prompt-engineering
- Claude Code slash commands, Anthropic documentation. https://code.claude.com/docs/en/slash-commands