Which of my rules did Claude Code actually use?

What Claude Code shows you: what loaded

• /context breaks your context down by category. The files under Memory files are the CLAUDE.md and rules files that loaded at launch. A CLAUDE.md in a subfolder loads later, on demand, so it is not listed there.
• The terminal prints a Loaded line with the path when a subfolder CLAUDE.md or a rule with paths loads mid-session. You see the path, not the content.
• The InstructionsLoaded hook fires once for every CLAUDE.md or .claude/rules file that loads, with the file's path, its scope and the reason it loaded: session start, a subfolder, a matching path, an import or a compaction. It cannot block anything, and it does not fire for an AGENTS.md Claude reads directly.
• /memory lists where your memory files live, including ones that do not exist yet, and opens them. It is a map, not a load report.

To keep a record across sessions, log the hook. This appends one line per file that loads:

// ~/.claude/settings.json
{
  "hooks": {
    "InstructionsLoaded": [
      {
        "hooks": [
          {
            "type": "command",
            "command": "jq -c '{file_path, load_reason}' >> ~/.claude/instructions-loaded.log"
          }
        ]
      }
    ]
  }
}

A rule with paths loads only when Claude reads, writes or edits a file that matches. If its Loaded line never appears, the rule never reached the model, and nothing about its wording matters yet.

What it cannot show: which rule was used

Anthropic's own debugging guide draws the line. If /context shows a file loaded and Claude still is not following it, the problem is likely how the instruction is written, not whether it loaded. A CLAUDE.md is context, not enforced configuration.

Skills are the one place Claude Code counts use: /skill-doctor shows how often each skill is invoked, because a skill is a discrete call. CLAUDE.md lines and rules are always-on text, and nothing in Claude Code's docs counts them. /doctor prompt-audit reviews the files themselves, not what happened in a session.

The academic name for the gap is the attribution crisis: agents cite a fraction of what they read. The Always-On Agents survey, covering 400 works, lists “which stored information influenced a specific output” among the field's open problems. Not a solved dashboard. An open problem. A count can still point at the likeliest suspects: stale rules that are served and never cited.

Two numbers you can count: served and cited

ETH Zurich's result is the one to keep in your head. arXiv:2602.11988: a developer-written context file moved success 2.4 points, a gain the authors could not tell from noise, and cost about 20 percent more inference. The file was served on every run. Our inference, not something the study measured: almost none of it can have been load-bearing, or the gain would not be that small. Nobody running that study had a count of which lines were which.

Harbor records both, per rule:

• Served is how many times Harbor handed the rule to an agent.
• Cited is how many times an answer referred back to it.

A rule served a hundred times and never cited is a rule to question, and the list of those is a list of rules you can probably delete. That is the product: subtraction with a denominator, not a retrieval score. The docs show how Harbor counts it.

If a rule must hold, do not leave it to a count

A count tells you what to cut. It does not make a rule stick. For anything that must never happen, such as a destructive command, use a hook or a deny rule that refuses it before it runs. The counts drawn on the home page are sample data and labelled as such. The one result we publish is the controlled test on that page, with its setup stated: one agent and one saved rule, not a team over time. Any other percentage attributed to Harbor did not come from Harbor. Which rules belong where, for a team, is in our Claude Code best practices for teams.