Write a Prompt Guide

A skill pack's SKILL.md tells an agent what to do. Its README.md tells a reader what the pack is. Neither tells a user what to type, what context to hand over before the run starts, or which of the pack's known limitations they have to compensate for by hand. That gap is what this skill fills, for one named target pack at a time.

This is not a prompt-engineering tutorial. A generic "be specific, give examples" page helps nobody who already installed the pack. The deliverable is specific to one pack: its real inputs, its real invocation triggers, its real output shape, and its real open bugs.

Evidence basis

The structure below is not invented. It is the intersection of the official prompting documentation of four model/agent providers — Anthropic, OpenAI, Google, and BytePlus — recorded with URLs and per-claim attribution in this repo's SOURCES.md (ANTHROPIC-PROMPTING-01, OPENAI-PROMPTING-01, GOOGLE-PROMPTING-01, BYTEPLUS-PROMPTING-01).

Two techniques appear in all four independently and a third in two of them independently, which is why these three are the non-negotiable spine of every guide this skill produces:

1. Structural separation of instructions, context, examples, and input — Anthropic via XML tags (<instructions>, <context>, <input>), OpenAI and BytePlus via the same four Markdown sections (# Identity, # Instructions, # Examples, # Context), Google via "XML tags or Markdown". Four sources, one technique.
2. Few-shot examples — Anthropic asks for 3–5, relevant and diverse, wrapped in <example>/<examples>; OpenAI calls it "a handful of input/output examples"; BytePlus's general-task template requires at least two worked Question/Output pairs; Google contrasts zero-shot with few-shot.
3. Match instruction precision to the model's reasoning mode — OpenAI: reasoning models want high-level goals, GPT models want "precise instructions that explicitly provide the logic and data required"; BytePlus draws the same line between deep-thinking and non-deep-thinking models. Two independent sources, identical principle.

Anthropic supplies the remaining spine pieces from Claude's specific angle: explain why a constraint matters rather than only stating it, set a role in one sentence, put long/stable material near the top and the query at the end, and — for agentic runs — track state explicitly and chain complex work into steps. Google adds the iteration moves (rephrase, reorder, decompose or chain) and explicit constraints covering both what to do and what not to do. BytePlus adds the discipline that comes before any of it: define success criteria first, then iterate Update → Test → Debug → Evaluate, versioning as you go.

Two items in the source material are API-level and do not apply to a user typing at a Claude Agent Skill in chat, and a guide should say so rather than silently importing them: generation-parameter tuning (temperature/topK/topP, Google) and prompt-caching placement (OpenAI) — the user does not control either from a chat turn. RAG/file-search wiring (OpenAI) is likewise out of scope unless the target pack itself takes file inputs.