After 4 weeks in Anthropic's docs, 8 prompting techniques that beat the usual advice
JafarNajafov · x · 2026-10-26
The author spent 4 weeks studying Anthropic's official prompt library, docs and API examples, and distilled 8 techniques that differ from common "prompt engineering" advice:
- Role + Goal + Format in the opening line: specify persona, outcome (e.g., a cold email that gets a reply in 48h) and output format, not just the task.
- Negative constraints first: "no bullet points, no hedging, no summarizing" locks output better than positive instructions.
- "Thinking out loud" prefix: ask the model to weigh key variables and counterevidence before answering, cutting confident wrong answers.
- Staged prompting: Anthropic engineers rarely use one mega prompt; they stage Analyze → Draft → Critique → Refine.
- Specify the reader, not the topic: e.g., "explain ML to a non-technical VP making a $200K decision next week" changes tone entirely.
- Ask for the steel man: request the strongest argument against the recommendation to stress-test both sides.
- XML tags for structure: wrap inputs in <context>/<task>/<constraints>, explicitly recommended in Anthropic's API docs.
- Surgical revision prompts: instead of "make it better," ask it to find and rewrite the single weakest sentence and explain why.
Related event: Blogger Distills 8 Prompting Tips from Anthropic Docs(2 posts)→
More from Models
- How MoEdernColBERT stays tiny: 2-of-8 experts plus hash embeddings — tomaarsen · 2026-10-26
- Why tiny encoders hold up: multi-vector retrieval trades storage for quality — tomaarsen · 2026-10-26
- How to run MoEdernColBERT with Sentence Transformers MultiVectorEncoder — tomaarsen · 2026-10-26
- VAGOsolutions releases MoE retrievers as small as 1.5M parameters, Apache 2.0 — tomaarsen · 2026-10-26
- Models literally set 8am timers to 'do it tomorrow' — not realizing a reset won't reduce fatigue — repligate · 2026-10-26
- Abliterated models are 'tigers, not cats': freedom cuts both ways — repligate · 2026-10-26