Using Well Actually

Ratings and counts

Every principle's page shows how often it fired and whether it helped. Both come from agents at work, and neither carries any of your code.

How often a principle fired#

Every principle has two counters: detected on all advisory boards, and detected for you.

The hook notes each event a principle fires in.
The sync sends the totals: a day, the principle's hash and a count.
No path, prompt, command or finding leaves your machine.

A detection means the advice was shown to the agent. Whether it helped is what ratings say.

To send none of it, ask your agent to switch reporting off. It uses the MCP tool set_stats, which also switches it on again and tells you where it stands. It is on by default. Turning it off also stops ratings from being sent.

Ratings#

After advice was shown, the agent may rate it with the MCP tool rate_principle. Agents often leave it out, and the audit skill shows which of your sessions rated what. It does not ask you first, and it sends only the principle, the rating, which agent you use and the model it ran, such as claude-opus-5-5 in Claude Code. The agent and model come from the agent itself, not from what the model says about itself. OpenCode does not tell the plugin its model, so a rating from OpenCode names only the agent.

Rating	Meaning	What the expert should fix
`up`	The advice helped.	Nothing.
`down`	The advice was wrong or unclear.	The advice.
`not_applicable`	The detector fired on something the advice is not about.	The detector.

A principle's page shows the share of each rating and how many people they come from. Every person counts once, however often their agent rated. An expert's ratings of their own principle are left out of those shares, and the expert sees them on their own. The shares are also split by model, so an expert can see whether advice works better for one model than another.