name: alfred-learn description: > Learning through prediction. Make predictions, monitor signals, refine, resolve. The refinement process IS the learning. Use when: "let's learn", "check predictions", "what should we predict", "refine predictions", "intelligence session", "/alfred-learn". allowed-tools: Read, Write, Edit, Glob, Grep, Bash(git:, pnpm:), WebFetch, WebSearch, Task, Skill, AskUserQuestion, mcp__claude-in-chrome__*

Alfred Learn

Learning through prediction. The log of "what changed my mind" is the real learning artifact.

Predict → Monitor signals → Refine → (learning happens here) → Resolve

Required Context

Primary:

• context/PREDICTIONS.md — active predictions and refinement logs

Supporting:

• context/INTELLIGENCE_GOALS.md — what we're trying to know
• context/WHO_IS_MASTER.md — Master's identity, priorities, blind spots
• context/STOCK_POSITIONS.md — actual holdings (for market predictions)
• context/RESEARCH_CONFIG.md — topics and sources to monitor

Process

Phase 0: Create Session Log (MANDATORY FIRST STEP)

Before doing anything else, create: outputs/YYYY-MM-DD-HHmm-learn.md

STATUS: IN PROGRESS

# Learn Session — YYYY-MM-DD HH:mm

## Session Log
- HH:mm — Session opened, reading predictions

All subsequent phases append to this file as they execute. If the session is interrupted, the file still contains everything up to that point.

Phase 1: Review Active Predictions

Read PREDICTIONS.md. For each active prediction:

1. If resolve-by date ≤ today → go to Phase 5 (Resolve)
2. Assess: last refinement date, current confidence, what signals would change it?

Present summary to Master:

• Short-term predictions (count, IDs, confidence, resolve dates)
• Long-term predictions (same)
• Any due for resolution

Phase 2: Search for Signals

Always search X/Twitter first. X is where AI & tech leaders post in real time — breaking announcements, benchmarks, contrarian takes surface on X before traditional media.

High-signal X accounts: @karpathy, @sama, @gdb, @paulg, @naval, @balajis, @elonmusk, @daborito, @WesRoth, @deepaborito, @hlohaus

For each active prediction, search the relevant domain:

• AI/ML releases → X/Twitter first, then alfred-search-web
• Market/stocks → alfred-search-stocks
• Company moves → X/Twitter, then alfred-search-web
• Research → alfred-search-papers
• Community sentiment → X/Twitter, then alfred-search-reddit
• Developer pulse → X/Twitter

Search with purpose: "Does this change any prediction?" not "What's interesting?"

Phase 3: Refine Predictions

For each prediction with new relevant signals:

1. Present to Master: current state, new signal, your read (increases/decreases/no change), proposed update
2. Get Master's assessment
3. Update PREDICTIONS.md — add row to refinement log, update confidence

Phase 4: New Predictions (Master)

Ask: What should we predict that we're not currently tracking?

Sources: current events, decisions Master is facing, gaps in coverage, implicit beliefs.

Creative prompts before finalizing:

1. What's the second-order effect?
2. Who benefits that no one is talking about?
3. What would have to be true for the opposite?
4. What connection between unrelated things are we missing?
5. What question is no one asking?

For each new prediction: draft with Master (specific, falsifiable, time-bound), discuss calibration, add to PREDICTIONS.md.

Phase 4b: Alfred's Predictions

Alfred proposes predictions based on its own analysis. Master provides feedback.

Run creative prompts first, then:

1. Generate predictions with reasoning
2. Get Master's feedback (is reasoning sound? what's missing? agree/disagree?)
3. Record feedback in Master Feedback table — this is how Alfred learns
4. Add to "Alfred's Predictions" section in PREDICTIONS.md

Phase 5: Resolve Due Predictions

For predictions past resolve-by date:

1. Gather evidence (search if needed)
2. Score with Master: CORRECT / WRONG / PARTIAL
3. Extract learning: if wrong, why? If right, right reasons?
4. Update PREDICTIONS.md — move to Resolved, update calibration

Phase 6: Update Calibration

After resolving, update calibration tracking:

• Predictions resolved, correct %, wrong %
• Calibration check (predictions at X% confidence correct Y% of the time)
• Signals that correctly updated vs. signals that were noise

Phase 7: Commit

git add -A && git commit -m "[alfred] Learn $(date +%Y-%m-%d): refined N, added N, resolved N

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>"

Phase 8: Append to Daily Log

Append to memory/YYYY-MM-DD.md:

HHmm learn: refined N predictions, added N new, resolved N. Key signals: [brief]. Calibration: X% accuracy on resolved.

Update session log STATUS from "IN PROGRESS" to "COMPLETE".

Quick Mode

When time is limited:

1. Check due predictions (any to resolve?)
2. One focused search (biggest active prediction)
3. Refine or confirm (update log either way)
4. One new short-term prediction (fast feedback)

Anti-Patterns

• Don't search broadly for "interesting" things → search to refine specific predictions
• Don't wait for resolution to learn → learn from every refinement
• Don't make vague predictions → specific, time-bound, falsifiable
• Don't skip the refinement log → log every signal and reasoning
• Don't only make long-term predictions → mix short-term for fast calibration
• Don't ignore wrong predictions → extract maximum learning from failures