Skip to content

8. Read the summary, run advise, apply a change

A week or so of sessions in the project, with tags on the rejections you made. Three rejections with one tag in one project is the threshold advice uses.

/just-say-so:summary --since 30d

Read three sections. Rejection rate is rejections per 100 turns, counted only from rejections that join to a logged turn. Rules on rejected replies says which Vale rules fired on the replies you rejected, and rejected and clean counts the ones Vale passed, which are the misses only a bullet can cover. What the tags point at lands each recurring tag in one situation: uncovered, judgment, fired and still shipped, silent, or not enough evidence. Records and the summary defines every number.

/just-say-so:packages --refresh

This downloads the 25 writing-rule packages in the Vale registry, about 3 MB, and reduces each to its rule names and messages. Advice can then suggest a rule from a package you do not have. Without it, suggestions for other packages come from one-line descriptions, and the advice says so.

/just-say-so:advise --since 30d

For each situation, the model writes one recommendation with the counts it rests on and the change ready to paste. It names only rules and packages the script printed. For #undefined-term, now covered by your judgment bullet, it will say whether the tag kept recurring after the bullet existed, and suggest rewording or a Vale rule if one in the cache matches. For an uncovered tag, it offers a bullet or a rule.

Pick one recommendation. A .vale.ini line goes in the ini; a bullet goes in .just-say-so-rules.json; a package goes in Packages followed by vale sync. You apply it; the plugin edits no config on its own.

The doctor shows the change: a new style under the reply target, a higher added count on the extra rules line, or a rule at a new level in the raw sections.

Put the commit gate on the repo: the same rules for text that reaches the repo without the agent.