Prompt & skill library

Setup

Lead with the role — it shapes the model's entire response space before a single instruction lands.

Patterns

Few-shot beats zero-shot 87% of the time on technical tasks. Cost: three examples. Reward: a different model.

Voice

Constrain tone with adjectives, not adverbs. "Direct, peer-to-peer" outperforms "speak clearly".

Format

If you need JSON, ask for JSON Schema first — then have the model fill it.

Reasoning

Add "think step by step" only once. Repeating it confuses the model into meta-thinking about thinking.

Voice

Ban your three most-hated AI tells in every system prompt. "delve", "tapestry", "in the world of".

Setup

Specify the audience before the format. "For a busy engineer" beats "use bullets" alone.

Reasoning

For >5-step tasks, generate the plan first as a separate call, then execute it in a second call.

Patterns

Self-critique adds 38% accuracy at 2.1x token cost. Worth it for any output you'd hand to a human.

Format

Specify line counts, not word counts. Line counts visibly constrain the model; word counts get ignored.

Voice

Add "no hedging" to any draft prompt. Cuts "might", "could", "may" by 60%.

Setup

When stuck, paste a sample of GOOD output before your prompt. The model will pattern-match upward.

Reasoning

Ask "what could be wrong with this answer?" as a follow-up. You get a free quality pass in 200 tokens.

Patterns

For agent loops, cap max iterations explicitly. "Stop after 5 critiques even if not satisfied."

Format

Use markdown tables for any output with >3 columns. The model handles them cleaner than prose.

Voice

Set a "voice anchor": one paragraph of your real writing. Best voice transfer technique in 2025.

Setup

Separate system and user prompts even on platforms that flatten them. The structure aids reasoning.

Reasoning

For ambiguous tasks, ask the model to enumerate 3 interpretations first, then pick one and proceed.

Patterns

Always include one negative example. "Not like this:" + bad sample. Cuts mode collapse by half.

Format

For structured outputs, validate with a schema and re-prompt on failure. Saves a human review pass.

Voice

Specify register, not formality. "Casual peer" vs "polite stranger" produces sharper differentiation.

Setup

Date your prompts. "As of 2026, …" anchors the model in the right knowledge cutoff frame.

Reasoning

When asking for opinions, force the model to commit. "Pick one and defend it" beats "weigh options".

Compliance

For regulated outputs, require citation lines. The model can't hallucinate sources it must cite by URL.