A config analyzer for .grokrules (and .grok/GROK.md).

When xAI shipped Grok Build (May 14, 2026), the team deliberately matched Anthropic's skill and instruction-file format. The official xAI docs say so: Grok also reads CLAUDE.md, AGENTS.md, and the .claude/rules/ family alongside its own .grok/GROK.md. That's why ccmd doesn't need a separate "Grok mode": the patterns that rot a CLAUDE.md (long compound rules, vague qualifiers, ISO dates near the top, prohibitions without a reason) are exactly the patterns that rot a .grokrules.

The rest of this page walks through what the analyzer does to a real paste, with the actual file ccmd flags as a worked example. The whole thing takes about 200 ms in your browser.

1. The input

Here is a representative .grokrules from a real-feeling Next.js payments project. It looks fine on first read. It scores 4 of 12 on the rubric and ships 7 findings.

2. How ccmd decides this is a .grokrules

The detector lives at src/lib/analyzer.ts:45. It inspects the first 300 characters of the pasted text and picks .grokrules when it sees either the literal string .grokrules or a markdown heading like # grok. Otherwise it falls through to CLAUDE.md. The rubric and heuristics that run next are identical, so a wrong label changes only the badge at the top of the report.

3. What the rubric actually checks

Four passes run against the file, in order:

4. The output for the sample above

Same file, run through the analyzer. The actual UI renders a richer panel; the shape of the data is below.

Read top to bottom: L3 and L4both flag the same line for different reasons (an unbounded "Always" and a vague "appropriately"); fix one and the other often goes away. The two missing-why findings on L8 and L9 are the cheapest fix on the page: one line each. The bloat finding on L13 is a 34-word stack paragraph that wants to be three short rules. The cache-bust on L17 alone is worth more than the rest combined: a timestamp that close to the top of a file Grok Build (or Claude Code) caches every session is paying full input-token cost every turn for no behavior change.

5. What the other tools in this corner do (and don't)

6. Fixes that move the score

For the sample file above, the four changes that get the rubric from 4 / 12 to 9 / 12 and cut roughly 60 tokens per turn:

• Move the date out of line 17. Compute "today" at runtime from a hook or strip the line entirely. Saves the cache hit and removes the single highest-severity finding.
• Add a Why: under each Never. "Never use any" becomes "Never use any. Why: silent runtime errors in checkout flow on 2026-03, took 4 hours to trace." One line each, two findings gone.
• Split the 34-word stack paragraph. Three short rules: framework, payments boundary, migration rule. The second half of the original line was getting ignored anyway.
• Drop "appropriately" and "carefully". Replace with concrete tests. "Handle edge cases appropriately" becomes "Cover the four edge cases in cases.test.ts."

Add a Tests-as-Truth line (rule 7), a stack name (rule 9, already half-present), and one self-improvement instruction (rule 12: "when you hit a behavior you don't expect, write the rule to this file before fixing"), and the file scores 9 / 12. The token saving is real but secondary; the larger win is that Grok Build (and Claude Code, and Codex, and Cursor, all reading the same file) stop ignoring half the rules.