Clarificus Maxima

---
name: clarificus-maxima
description: Audit the repo's Claude instructions for vague rules — rules with no trigger, no scope, or no falsifiable condition. The 5-family reaches wider for instructions than older models, so a vague rule doesn't get ignored, it gets applied everywhere and dilutes the rules that matter. Use for "Claude over-applies my rules", "it keeps doing X when I only meant Y", "tighten up my CLAUDE.md", or as part of /opus-pocus. Read-only; rewrites each vague rule as trigger→action or recommends deletion. NOT for style preferences — this hunts rules whose vagueness changes behavior.
---

# 🛡 Clarificus Maxima

> *"A rule with no trigger is just vibes in a trenchcoat."*

Older models under-read instructions, so authors compensated with broad rules. The 5-family
weighs *every* rule in context on *every* turn — a rule with no trigger fires everywhere,
taxing attention and diluting the rules that carry real constraints.

## What to scan

Every memory file the repo carries (`CLAUDE.md`, `AGENTS.md`, nested), every rule file, output style, skill body, and agent definition.

## Anti-patterns to find

1. **No trigger.** "Be careful with the database." When? Doing what? A rule that cannot answer
   "this applies when ___" applies always, which means it competes with everything.
2. **No scope.** "Always write tests." For a typo fix? A doc edit? Unscoped absolutes force the
   model to either over-comply or learn to ignore the file.
3. **No falsifiable condition.** "Keep code quality high", "prefer clean solutions", "be
   thorough". Nobody — model or human — can check compliance. Pure attention tax.
4. **Aspirational rules.** Rules describing a hoped-for culture rather than a decision
   procedure ("we value simplicity"). These belong in a philosophy doc, not an instruction file.
5. **Hedged rules.** "Generally avoid X unless it makes sense." The exception clause is the
   whole rule; the model must guess the actual boundary every time.
6. **Duplicate rules at different vagueness levels.** A precise rule in one file and its vague
   cousin in another — the vague one wins in novel situations because it matches more contexts.

## Fix prescriptions

| Finding class | Fix |
|---|---|
| No trigger | Rewrite as `WHEN <observable situation> → <action>`. If no trigger can be derived, the rule is a preference — delete or demote it |
| No scope | Add the boundary explicitly, including the negative ("this does not license…") when over-application is the observed failure |
| Not falsifiable | Replace with the concrete check that would verify it, or delete. "Include error handling when the change's stated purpose fails without it" beats "handle errors properly" |
| Aspirational | Move to a non-instruction doc, or convert into its one operative consequence |
| Hedged | Name the exception's actual criterion. If the criterion can't be named, the rule isn't one |
| Duplicates | Keep the precise version, delete the vague one, leave a pointer if files differ |

## The test for a good rule

A rewritten rule passes when a second reader could apply it to a novel case and reach the same
answer as its author. Trigger + action + (where needed) the named exception — nothing else.

## Bounds

- Scope: memory files, rule files, output styles, and the bodies of skills and agents. Not application source, not dependencies, not another installed plugin.
- Cap: 15 findings, ranked by how far the vague rule over-applies. A preference that changes nothing is below the floor.
- Done when the declared scope has been swept once and ranked — every instruction layer has one more loose rule in it.
- 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 rule · proposed rewrite (or DELETE with
rationale). End with the density stat: rules scanned, rules vague, and the three vaguest files
by count. `SCANNED NOTHING` if no instruction layer was found, grading `N/A`.