howiclaude efficiency rubric — v1.0

"Efficiency" is a transparent **0–100 composite** of five **0–20 sub-scores**. Each is computed from what your config actually shows — or, where noted, from something you self-attested in the interview (those drive *self-reported* badges, marked as such on the page). A higher score is **not** "better engineer"; it's "more of the harness-optimization behaviors the community talks about." This rubric is versioned — every analyzed page is stamped `rubricVersion: 1.0`.

### The five sub-scores (each 0–20)

- **Context discipline** — +8 if `CLAUDE_CODE_AUTO_COMPACT_WINDOW` is set at all; +4 more if it's tuned away from the default; +4 if you self-attest compacting before ~40% context; +4 if a Stop/PreCompact hook or a `/compact`-style command exists. - **Automation depth** *(weighted, never raw counts)* — hooks: +3 for the first, +2 for the second, +1 each for #3–#6 (cap +9); slash commands with a non-trivial body: +1 each (cap +5); subagents: +1 each (cap +4); skills: +0.5 each (cap +2). Twenty unused commands is not a 20. - **CLAUDE.md specificity** — Claude grades the file 0–20 against a published grading prompt: concrete rules **with rationale**, path-scoped rules, a `rules/` directory → high; vague platitudes → low; empty/absent → 0. - **Parallelism** — +6 if `worktree.*` is configured or worktrees are used; +6 if you self-attest 2 concurrent instances, +10 for 3, +14 for 4+; +6 for a `/loop`/overnight-agent pattern (cap 20). - **Permission hygiene** — start at 20; −20 if `--dangerously-skip-permissions` is the norm; −6 if `permissions.allow` is effectively `*`-ish (overbroad); a documented `sandbox` config recovers up to +6 of a deduction.

**Composite** = the sum of the five (0–100).

### The percentile band

For v1 the band — Top 1% / Top 5% / Top 10% / Top 25% / Top 50% / Below median — is a **fixed score→band lookup**, calibrated on the catalogued seed setups, *because there isn't a real population yet*. It becomes a true percentile once enough real pages exist. We will never imply precision the data doesn't support.

### Archetype

`The Vanilla Purist` if your config is within a small edit-distance of the stock defaults (a flex for minimalists). Otherwise `The Dictator` if you self-attest dictating more than typing. Otherwise it's named by your dominant *naming* sub-score — Context discipline → `Context Surgeon`; Automation depth → `The Maximalist`; CLAUDE.md specificity → `The Architect` (if you also self-attest plan-mode-first) else `The Documenter`; Parallelism → `The Parallel Demon`. (Good permission hygiene is expected, not an archetype.) If none of those four apply at all, you're `The Minimalist Monk` (or `The Maximalist` if your composite is high anyway). The archetype list isn't closed — re-running the prompt to change yours is fine; it nudges better practice.

Earnings estimate methodology

The "estimated earnings" on a manage page is a *rough* per-click number, applied per gear category. The point is "you could earn something from your rig", not a precise prediction. Actual earnings depend on conversion rates, the specific items, the affiliate program's commission tier at the time, and (for Amazon) the buyer's session.

The categories and their rough cents-per-click rates are listed below. Unknown items fall back to the default rate. SaaS referrals are higher than physical gear because they're typically flat bounties on signup, where physical gear is a small percentage of the sale price after a low conversion rate.

These numbers are version-locked with the launch (`RUBRIC_VERSION` in `src/lib/scoring.ts`); if we change them, we bump the version and note it here.