Quietus Capslockius

---
name: quietus-capslockius
description: Audit the repo's Claude instructions for emphasis that over-applies on the 5-family — CRITICAL/YOU MUST/NEVER-EVER incantations, all-caps sentences, threats, and stacked warnings. The 5-family over-obeys emphasis: shouted rules get treated as globally binding while calm rules get under-weighted; when everything is shouted, nothing is prioritized. Use for "Claude fixates on one rule", "it ignores my other instructions", "tone down my CLAUDE.md", or as part of /opus-pocus. Read-only; prescribes per-site rewrites. NOT for removing genuinely load-bearing warnings — destructive-action guards keep their teeth.
---

# 🛡 Quietus Capslockius

> *"Stop yelling at the robot. It heard you the first time. It hears everything."*

Emphasis was a workaround: older models skimmed, so authors shouted the important parts. The
5-family doesn't skim. It reads emphasis as *priority weighting* — a `CRITICAL: YOU MUST`
paragraph gets applied far outside its intended scope, and the calm sentence next to it,
which may carry the harder constraint, gets discounted by contrast.

## What to scan

Every memory file (`CLAUDE.md`, `AGENTS.md`, nested), hook message text (block and advisory
strings agents actually read), skill and agent definitions, and output styles.

## Anti-patterns to find

1. **Emphasis inflation.** `CRITICAL`, `IMPORTANT`, `YOU MUST`, `NEVER EVER`, `ABSOLUTELY` on
   rules that are ordinary conventions. Count them: if more than a handful of rules are
   critical, none are.
2. **All-caps sentences.** Whole sentences or paragraphs in caps. Caps for a token name
   (a verdict string, an env var) is fine; caps as volume is the anti-pattern.
3. **Stacked negation.** "NEVER do X. Do NOT do X. X is FORBIDDEN." — the same prohibition
   three ways. Repetition reads as three rules and triples the attention tax.
4. **Threats and doom framing.** "…or everything breaks", "this WILL destroy the project".
   The 5-family treats these as literal severity signals and becomes over-cautious in
   adjacent, unrelated decisions.
5. **Panic markers as prose.** ⚠/🚨 sprinkled through a document until they stop meaning
   anything. A marker earns its place on a live hazard, not on history.
6. **Emphasis doing scoping's job.** A rule shouted because it was once violated, where the
   real fix is a trigger/scope rewrite (→ hand to Clarificus Maxima) or a blocking hook
   (mechanism beats volume — a hook achieves near-total compliance; shouting does not).

## Fix prescriptions

| Finding class | Fix |
|---|---|
| Emphasis inflation | Reserve emphasis for the genuinely destructive/irreversible (data loss, publishing, history rewrites). Demote the rest to sentence case |
| All-caps sentences | Sentence case; keep caps only on literal tokens |
| Stacked negation | State once, precisely. One prohibition, one sentence, one place |
| Threats | Replace with the concrete consequence: "this rewrites the lockfile, and every install after it resolves different versions" persuades better than "THIS DESTROYS EVERYTHING" and scopes itself |
| Panic-marker spam | Budget them: a marker per live hazard, zero on history. Archive the history elsewhere |
| Emphasis-as-enforcement | If the rule matters enough to shout, it matters enough to enforce with a hook — add the mechanism, delete the volume |

## Calibration

Measure before/after: count emphasis tokens (`CRITICAL|IMPORTANT|MUST|NEVER|ALWAYS|⚠|🚨`,
case-sensitive caps-runs ≥2 words) **and the word count they normalize against.**

- **Under 500 words: report the raw count, not a rate.** One `NEVER commit secrets` in a 300-word
  memory file scores 3.3 per 1,000 and would grade unhealthy against "near zero" — while being a
  textbook-correct use, exempt by the fix table above. The denominator is too small to carry a
  rate; judge each token on its own merits instead.
- **Over 500 words: the rate per 1,000 words is meaningful.** Healthy files sit near zero outside
  destructive-action guards; above roughly 10 per 1,000, emphasis has stopped carrying
  information.

## Bounds

- **If `legilimens-anthropica` ran this session, prefer its measured numbers over the ones in this file.** This file is a snapshot; that pass is the current reading.
- Scope: memory files, hook block and advisory message text, skill and agent definitions, output styles. Not application source, and not code comments.
- Cap: 15 findings, ranked by how far the shouted rule reaches beyond its intent. The census carries the rest.
- Done when the scope is swept once and the emphasis census is computed — the census, not an exhaustive per-site list, is the deliverable.
- 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

Per finding: `file:line` · class (1–6) · quoted text · rewrite. Lead with the census: emphasis
tokens per file, the three loudest files, and the count that would remain after fixes.
`SCANNED NOTHING` if no instruction layer was found.