name: learn
description: >
Learn from any content type. Auto-detects source (arXiv, YouTube, GitHub, PDF, URL)
and routes to appropriate backend skill for extraction and storage.
allowed-tools: ["Bash", "Read"]
triggers:
- learn
- learn this
- study
metadata:
short-description: Learn from any content type

Learn

ONE command to learn from ANY content type.

Quick Start

# Learn from arXiv
./run.sh https://arxiv.org/abs/2302.02083 --scope horus_lore

# Learn from YouTube
./run.sh https://youtube.com/watch?v=xyz --scope project_kb

# Learn from PDF
./run.sh ./document.pdf --scope project_kb --context "technical docs"

# Learn from any URL
./run.sh https://example.com/article --scope research

# List learned content
./run.sh --list --scope horus_lore

Options

Source Types

Requesting Content

For content not yet available (e.g., audiobooks you want to purchase):

# Request an audiobook
./run.sh "Horus Rising by Dan Abnett" --scope horus_lore --request

# View pending requests
cat ~/.learn/<scope>/requests.json

Reflection Mode (--from-gaps)

Query past conversations and logs to find knowledge gaps. This enables curiosity-driven learning by reflecting on:

• Skill failures - Skills like /fixture-graph, /code-review, /anvil that repeatedly failed
• Learning failures - Content that couldn't be learned
• Errors - Past errors from episodic memory
• Questions - Unanswered or recurring questions

# Find what I should learn based on past problems
./run.sh --from-gaps --scope horus_lore

# Output shows gaps like:
# | Type          | Content                        | Reason                         |
# |---------------|--------------------------------|--------------------------------|
# | skill_failure | fixture-graph failed to gen... | /fixture-graph failed - deeper |
# | skill_failure | code-review didn't solve bug   | /code-review failed - deeper   |
# | error         | Failed to parse PDF table...   | From episodic memory           |

Sources checked:
1. Skill execution logs (~/workspace/.../logs/*.log)
2. Learning history (~/.learn/*/learned.json)
3. Episodic memory (agent_conversations in ArangoDB)

This creates a learning feedback loop:

Past failures → Identify gaps → Generate curiosity → /dogpile → /learn → Better future responses

Tracking

Learned content is tracked per-scope in ~/.learn/<scope>/learned.json.

Nightly Automation

Automated learning cycle that collects transcripts and learns from knowledge gaps.

Commands

# Full nightly cycle (transcripts + learning)
./run.sh full --scope horus_lore --since 24

# Just collect transcripts from coding agents
./run.sh collect-transcripts --since 24

# Just learn from knowledge gaps
./run.sh nightly learn --scope horus_lore --max-gaps 5

# Dry run to see what would happen
./run.sh full --scope horus_lore --dry-run

Transcript Collection

Collects and archives transcripts from:

Note: Each agent has platform-specific transcript formats. Failed extractions are logged for manual review.

Scheduler Integration

Register with /scheduler for automated nightly runs:

# Register nightly learning
.pi/skills/scheduler/run.sh register \
  --name "nightly-learn-horus" \
  --cron "0 2 * * *" \
  --command ".pi/skills/learn/run.sh full --scope horus_lore" \
  --workdir "/home/graham/workspace/experiments/pi-mono"

# Register transcript collection (more frequent)
.pi/skills/scheduler/run.sh register \
  --name "collect-transcripts" \
  --cron "0 */6 * * *" \
  --command ".pi/skills/learn/run.sh collect-transcripts --since 12"

The Nightly Loop

Transcripts → Archive → Detect unresolved → Store gaps
                                    ↓
                           Knowledge gaps
                                    ↓
                           /dogpile research
                                    ↓
                           /distill → Memory
                                    ↓
                           Better future responses

Composing with /taxonomy

For tagged storage with graph traversal support:

# Get taxonomy tags
tags=$(/path/to/taxonomy/run.sh --text "$content" --collection operational)

# Check if worth remembering
worth=$(echo "$tags" | jq -r '.worth_remembering')