Learn Taste runs at the end of a session. It reads the conversation, including edits you made yourself, filters real corrections from noise, generalizes each into a rule, finds the reason behind it and decides where the rule should live. It proposes the rules, stops, and writes only what you approve, so the next agent repeats no mistake. It is set to run only when invoked.
What You Get
- A list of proposed rules, each with its reason
- Questions for rules where the reason is unknown
- A conflicts list when a rule clashes with an existing one
- Suggested lint rules or checks for violations that code can detect
Common Use Cases
- Ending a design session by recording the corrections
- Building a DESIGN.md that agents follow
- Turning a repeated visual mistake into a lint rule
What Makes It Unique
Every rule traces to something you said, edited or confirmed, and a reason is never invented: an inferred reason is marked and confirmed first. It prefers a check over a sentence, since a lint rule runs every time and a line in a long file can be skipped.
Under The Hood
A single SKILL.md with no scripts. The model reads the current conversation and, if you name earlier sessions, their Claude Code transcripts under the projects folder in your home directory. It proposes rules, waits, and then writes to DESIGN.md or the chosen surface once you approve.