escapeboy/ai-prompts/update-docs

/update-docs — Documentation Updater

Keeps your Claude Code configuration documentation current with the latest Claude API changes, best practices, and framework updates. Searches the web, compares findings to existing docs, and applies targeted updates.

Especially useful for keeping ~/.claude/ system prompts, skill files, and project memories accurate after Claude API updates.

Scope: /update-docs handles external freshness — API / model / pricing / beta-header drift sourced from the web. For internal consistency (dead links, README↔CHANGELOG version parity, terminology, stray Cyrillic, privacy leaks) use `/content-review` instead. The two are complementary, not overlapping — don't duplicate content-review's checks here.

When to Use This Skill (and When NOT to)

Start simple. Reach for this skill only when the update depends on *external* freshness you have to research, or spans enough files that a systematic research → analyze → update pass beats hand-editing.

Usage

/update-docs [action] [target]

Quick Examples

/update-docs                           # Default: analyze (read-only drift check, no edits)
/update-docs research "prompt caching" # Search web for latest info
/update-docs collect https://docs.anthropic.com/en/docs/build-with-claude/prompt-caching
/update-docs analyze                   # Compare findings to existing docs
/update-docs update global-optimization.md
/update-docs update --scope global     # Update all global docs
/update-docs validate                  # Check docs for outdated content

Actions

Five actions, run individually or chained research → analyze → update. Full process steps, primary sources, output formats, scope table, and the per-action editing rules live in [references/actions.md](references/actions.md) — read the row for the action you're running.

Fetching: prefer Lattice for browser/search if available; plain WebFetch is intercepted by the context-mode hook — use ctx_fetch_and_index(url, source) + ctx_search instead (this is expected, not an error).

For the recommended monthly full-refresh workflow, token budget, and troubleshooting, see references/workflow.md.

Boundaries

Always

• After an update, verify: re-read the file (or run a regression grep) to confirm the change landed and didn't break JSON / fence parity before reporting done.
• Lead with authoritative primary sources (platform.claude.com, the bundled claude-api skill) — not community mirrors — for any model / pricing / API fact.
• When editing the claude-code-kit repo, ship the change as the four synced Releasing artifacts (CHANGELOG + README Version History & footer + annotated tag + GitHub Release; see repo CLAUDE.md → "Releasing").

Ask first

• Bulk update --scope global / --scope memories (edits across many files at once) — preview with --dry-run and confirm before applying.
• Publishing a release (pushing the annotated tag + creating the GitHub Release) — shared-state and hard to reverse; sweep for private references first.

Never

• Retro-edit historical records — model IDs, prices, and dates inside CHANGELOG entries, README "Version History", and any .backups/ directory are frozen. Only touch live/current claims.
• Duplicate /content-review's internal-consistency checks — this skill owns external freshness only.
• Treat a blocked WebFetch as a failure — route through Lattice or ctx_fetch_and_index instead.

See Also

• references/actions.md — full per-action detail (process, primary sources, output formats, editing rules)
• references/workflow.md — monthly full-refresh workflow, token budget, troubleshooting
• `/content-review` — complementary: internal consistency (dead links, version parity), where this skill owns external freshness
• `/init-project` — Initial docs setup for a new project
• `/ctx refresh` — Refresh Serena memories specifically
• Research Integration Guide
• Custom Skills Guide — how these skills are authored