advanced-skill-authoringlisted

# Advanced Skill Authoring **REQUIRED BACKGROUND:** `superpowers:writing-skills` — the basics, TDD methodology, and CSO. This skill fills five gaps in that foundation. The patterns here have **not** been adversarially eval-tested with subagents (in violation of the parent skill's own Iron Law). Treat as starting hypotheses; run baselines for anything you depend on. --- ## 1. XML Tags vs Markdown Structure **Decision rule:** Markdown for the document outline. XML tags for content Claude must distinguish from instructions. Claude is trained on XML-tagged inputs. Tags create harder boundaries than `##` headers and resist instruction-bleed. Markdown remains better for navigable structure — Claude Code's own system prompt is markdown-heavy. | Content | Use | Why | |---|---|---| | Section outline | Markdown headers | Greppable, cheap | | Embedded examples | `<example>` | Anthropic's canonical wrapper; Claude won't read the example as an instruction | | Contrastive pairs | Markdown Bad/Good (house style) OR `<example type="bad">` | Either works | | Reasoning scaffold | `<thinking>` / `<answer>` | Trained tags | | Embedded user data, transcripts | `<document>`, `<input>` | Blocks instruction injection from data | | Procedures | Numbered lists | Tags add no value | | Lookup tables | Markdown tables | Scannable, token-cheap | Anthropic's docs assert tags improve parsing but cite no public benchmark; third-party "20-40% improvement" claims are not from Anthropic and shouldn't be