Legilimens Anthropica

---
name: legilimens-anthropica
description: Fetch the CURRENT official and community guidance on structuring repos for Claude's newest models before auditing or changing anything — Anthropic docs, the Claude Code changelog, and credible practitioner writing. Best practices for the 5-family are still being figured out in public and drift fast; an audit calibrated on stale lore prescribes stale fixes. Use at the start of /opus-pocus, or for "what's the latest guidance", "are we following current best practice", "has Anthropic said anything new about CLAUDE.md". Produces a dated guidance digest the other spells calibrate against. NOT for general web research — Claude-instruction guidance only.
---

# 🔮 Legilimens Anthropica

> *"Read the minds at Anthropic before re-inventing them badly."*

Every other spell in this book encodes a snapshot of what was known when it was written. This
one refuses to trust snapshots — including this repo's own. Run it first; let what it finds
recalibrate the other passes' thresholds.

In the grand ritual it runs at **Phase 0.6**, before the graded passes rather than alongside its
Divination siblings — thresholds have to be recalibrated *before* the passes that use them, not
after. Class is taxonomy; phase is execution order.

## Sources, in trust order

1. **Official docs** — the Claude Code documentation (memory/CLAUDE.md guidance, rules,
   hooks, agents, skills pages) and Anthropic's prompt-engineering guidance for the current
   model family. Fetch and READ the pages; don't stop at search snippets.
2. **The Claude Code changelog** — new instruction-layer features (rules scoping, hook events,
   frontmatter fields) ship here first, often ahead of the docs.
3. **Anthropic engineering blog / model release notes** — behavior deltas per model family:
   what the new family does unprompted, what it over-obeys, what got deprecated as advice.
4. **Credible practitioners** — engineers publishing measured results (ablations with eval
   deltas, before/after token numbers), not vibes threads. A claim with a measurement beats a
   claim with a following.

## What to extract

- **Behavior deltas**: what the current family does *without being told* (self-verification,
  wide instruction reach, emphasis sensitivity) — these decide what instruction text is now
  redundant or harmful.
- **Feature surface**: instruction-layer capabilities the repo isn't using (path-scoped rules,
  new hook events, frontmatter fields like model/effort pins).
- **Deprecated advice**: yesterday's best practice that current guidance reverses. Flag any
  the repo still follows.
- **Size/structure targets**: current recommendations for memory-file size and organization.
- **Community-measured numbers**: compliance rates for hooks vs prose, ablation results,
  context-economics findings — with their sources, so later readers can re-check.

## Anti-patterns in your own process

1. Answering from training data when the network is available — the point is what changed
   since.
2. Citing a community claim without its measurement.
3. Treating one practitioner's workflow as normative — extract the measured finding, drop the
   workflow religion.

## Bounds

- Scope: external sources only — official docs, the changelog, release notes, and credible practitioner writing. This pass reads no repo files and proposes no repo edits.
- Cap: roughly 8 fetches. Read the pages you fetch; do not crawl outward from them.
- Done when the current contract for the instruction layer has been established and dated. If the network is unavailable, say so and mark the ritual as running on cached lore.
- Quote once, at read time; mark anything you cannot re-confirm `UNVERIFIED` rather than re-reading to check yourself.
- See `skills/opus-pocus/references/casting-safety.md` for the full rules.

## Report format

A dated guidance digest: per source, the findings relevant to this repo's instruction layer,
each tagged **official / measured-community / anecdotal**, with URLs. End with the
recalibration list: which other spells' thresholds or heuristics this digest updates (e.g. "the
truncation cliff moved", "emphasis guidance now explicit in official docs"). If the network is
unavailable, say so and mark the whole ritual as running on cached lore — do not silently fall
back. Offline, this pass grades `?` (could not verify), never `O`.