teach

Description and examples

What it does: Updates .vibeflow/ from natural language: corrects a pattern doc, adds a convention, records a decision, or documents a new pattern. Also imports patterns and conventions from an external reference repo via --from. Prefer corrections outside the auto-generated markers so they survive the next analyze.

Examples:

• teach sempre usar camelCase para variáveis de estado — Adds or updates a convention.
• teach o padrão de API mudou, agora validamos com zod — Updates the relevant pattern or conventions.
• teach decidimos usar Redis para cache, não in-memory — Logs the decision (e.g. in decisions.md or conventions).
• teach --from https://github.com/org/platform-patterns — Imports patterns from an external repo.
• teach --from ./my-patterns --name platform — Imports from a local path with a custom alias.

---

Language

Detect the language of the user's current request or conversation. Write ALL output in that same language. Technical terms in English are acceptable regardless of the detected language.

Process the user's current feedback and update project knowledge.

The one rule that governs every edit

When editing an existing generated region, everything you write goes outside the <!-- vibeflow:auto:start/end --> markers — that is what makes a correction survive the next incremental analyze. Read the target file before editing it, and never rewrite what is inside the markers; analyze owns that region. One explicit exception: a new pattern doc created in category (d) is born with its own markers and its initial content inside them, so analyze owns that region from the start.

Before starting

.vibeflow/ has to exist: read .vibeflow/index.md directly by path for orientation. If it isn't there, stop with ".vibeflow/ does not exist. Run analyze first to create the knowledge base." — don't create it by hand.

Then branch on the input: --from in the user's current request goes to Import from external repo; anything else goes to Classify the feedback.