🦋 Configure

"The whole book is a spellbook. What you carry is a wand."

Every installed skill costs its description in every session, whether or not you use it — that is what "always-on" means. The bodies load on invoke and cost nothing until then, but the descriptions are the router's menu and the router reads the whole menu every time.

For the repo this book was distilled from, all of it earns its place. For a Node service with a forty-line memory file, most of it is a menu of dishes the kitchen cannot cook: five spells have no subject there at all, and they still charge rent.

This spell picks the subset and writes it down.

The mechanism

Claude Code reads a skillOverrides object from settings — skill name to "off":

{
  "skillOverrides": {
    "avada-verbosa": "off",
    "the-pensieve": "off"
  }
}

A skill set "off" is absent from the session's skill listing entirely, so its description is never injected. That is the lever, and it is the only one that reduces always-on cost per skill.

Keys are bare skill names, not plugin:skill — that is the form observed working. If a change does not take effect, try the prefixed form before concluding the mechanism is wrong, and say which one worked.

Three levers, only one of which is this spell

One measurement trap, and it will mislead you

claude plugin details opus-pocus reports the plugin's packaged cost, not your configured one. Measured: with four skills set "off", it reported the identical always-on figure.

It is not lying — it answers "what does this plugin contain", and that is a different question from "what am I loading". Do not quote it as proof a config took effect. This is nihil-revelio's own class in the wild: a number that looks like the answer to the question you asked, and is the answer to a neighbouring one.

To verify a change actually took, start a new session and ask whether a disabled spell appears in the available skills. That is falsifiable; the cost figure is not.

Detection — the same computation as the ritual's triage

This spell and Phase 0.5 of the grand ritual answer the same question at different tempos. Triage decides which passes have a subject this run; configure decides which spells load every session. They share one detector so they cannot drift into disagreeing about the same repo — if a pass is N/A in the triage table, keeping it loaded is paying for a menu item the kitchen cannot cook.

Run the ritual's own Scope and Phase 0.5 detection — memory-file discovery (CLAUDE.md, AGENTS.md, nested, .claude/rules/), hook count, agent count, skill count, memory-file size — and read the profile off the result. Do not restate those rules here; read them from ../opus-pocus/SKILL.md so there is one copy.

Profiles

opus-pocus itself is in every profile: it holds the triage, the bounds and the report format, and without it the individual spells have no shared frame.

The placement rule, so a nineteenth spell has an obvious home: a spell belongs in the smallest profile where it changes an outcome — not merely where it has a subject.

Worked through, for the four that are not obvious:

• legilimens-anthropica — standard. It recalibrates thresholds, and the first threshold that matters is descriptio-reducio's truncation cliff, which arrives in standard. Below that, quietus-capslockius uses raw counts rather than rates, so there is nothing to recalibrate.
• priori-incantato — standard. A handful of hooks is enough to have one wired to an event it never emits, and that is the failure a small setup makes once and never notices.
• nihil-revelio and alohomora — full. Both need machinery you own and can change. A repo whose checks are eslint and pytest and whose only guard is husky cannot act on most of what they prescribe.
• accio-rulebook — full. Splitting a memory file into path-scoped rules pays when the root file is large; "several memory files" is not the same as "one file worth splitting".

protego is in standard and above, and you can switch it off afterwards. The hook is wired by the plugin manifest, not by skillOverrides — so once it is installed and running, turning the skill off keeps the shield and drops the description. Say so when you recommend it; it is the one spell whose job survives its own removal from the menu.

Procedure

1. Detect and recommend. Show the evidence, not just the verdict:

   Repo shape:  2 memory files (6 KB) · 0 hooks · 0 agents · 2 skills
   Recommended: minimal  — 4 spells loaded, the rest off
   Why:         no hooks and no agents, so sorting-hat, muffliato-hookus,
                priori-incantato, nihil-revelio and alohomora have no subject here

2. Estimate the cost both ways. Derive it rather than quoting a number: each skill's always-on cost is roughly its description: length in characters divided by four. Sum the loaded set, sum all of them, show both. State that it is an estimate, and that claude plugin details will keep reporting the packaged figure regardless.
3. Show the diff and ask. The exact JSON that will be added or removed, and the exact file it lands in. Then wait. Never write settings unasked — this file governs every future session in the repo, and a surprise edit to it is the kind of thing that gets a plugin uninstalled.

4. Write, merged, to project scope by default. <repo>/.claude/settings.json, so the profile follows the repo it was chosen for. --user writes ~/.claude/settings.json instead.

   Merge into any existing skillOverrides; never replace the object. A user may already have unrelated skills off, and clobbering those entries is real data loss in a file nobody re-reads. Back the file up first, write, then re-parse the result — a settings file left invalid disables every setting in it, which is a worse outcome than any token saving.

5. Report what changed and when it takes effect. The skill listing is assembled when a session starts, so a change most likely applies to the next session rather than this one. That has not been verified here — say so, and give the check: start a new session and ask whether a disabled spell still appears.
6. Re-running replaces only Opus Pocus entries. reset removes only Opus Pocus entries. Neither touches another plugin's.

Bounds

• Writes exactly one file, and within it only the skillOverrides key, and within that only this plugin's own entries. Another plugin's entry is never added, removed or altered.
• Asks before writing, every time. There is no unattended mode.
• Audits nothing, grades nothing, produces no findings — it reads the repo's shape to pick a profile and stops.
• Done when the profile is written and reported. Not when the configuration is "optimal" — there is no such state to converge on, and hunting one is the spiral this book bounds.
• Never edits a skill's own frontmatter to achieve an effect settings can achieve.
• See ../opus-pocus/references/casting-safety.md for the full rules.

Report format

The detected shape with its evidence, the recommended profile and why, the loaded-versus-all cost estimate with its caveat, the exact diff written, the file it went to, and the takes-effect-next-session note with its check.

SCANNED NOTHING if there is no repo to shape-detect — configure refuses to guess a profile from nothing, because the wrong profile silently hides the spell you needed.