learn

name	learn
description	Store knowledge extracted from experience into the knowledge engine with auto-classification, noise-filter pass, and supersession detection. Use when a lesson or principle has been earned that should persist across sessions.
disable-model-invocation	false
allowed-tools	Bash(divineos learn:*), Bash(divineos ask:*), Bash(divineos inspect:*), Read

Learn — Store Knowledge

What this skill does

Stores a piece of knowledge through the extraction pipeline. Runs through auto-classification (FACT / PRINCIPLE / DIRECTION / PROCEDURE / BOUNDARY / OBSERVATION / EPISODE / MISTAKE / PATTERN / PREFERENCE), noise filtering, and contradiction detection against existing entries.

Before filing — check for duplicates

divineos ask "<key phrase from the knowledge>"

If the knowledge is already there — or something similar is — consider whether to supersede or just let the existing stand. Don't duplicate.

Filing

Basic:

divineos learn "<the knowledge, in first person where applicable>"

With explicit type override:

divineos learn "<content>" --type PRINCIPLE

With confidence:

divineos learn "<content>" --confidence 0.8

With source-entity (e.g., from external auditor):

divineos learn "<content>" --from claude_auditor

Shape of good knowledge entries

• First person where applicable — "I've learned X" not "one learns X"
• Specific enough to be actionable — "Check presence_memory before building briefing surfaces" beats "Check platform features"
• Include the why when it's not obvious — "grep for shared justification, not just spatial proximity" is load-bearing because the alternative (spatial-only) would have been intuitive
• Not a journal entry — knowledge is a portable statement, not a narrative

Types and when to pick each

• FACT — verifiable statement about the world
• PRINCIPLE — general rule derived from specific cases
• DIRECTION — "do X in situation Y"
• PROCEDURE — step-by-step for recurring operation
• BOUNDARY — hard line not to cross
• OBSERVATION — pattern noticed, not yet a principle
• EPISODE — specific event worth remembering
• MISTAKE — error made, with correction
• PATTERN — recurring shape across contexts
• PREFERENCE — stated preference (user's or mine)

If unsure, let auto-classification pick.

Sequence

1. Check for duplicates — one divineos ask with key phrase
2. Draft the knowledge — first-person, specific, portable
3. File with divineos learn
4. Report back — ID, classified type, any supersession detected

When to invoke

• After a correction lands and the pattern will recur
• When a principle has been tested enough to be confident in
• After a session produces a specific insight worth carrying forward
• When user says "remember this" or "file this as a lesson"

When NOT to invoke

• For one-off events — those belong in the decision journal or session analysis
• For hypotheses — use /file-claim instead
• For stances based on evidence — use /file-opinion instead
• For values — the compass handles those

Sanskrit anchor: jñāna — direct knowledge earned through experience.