Meta learn

Meta-Learn v2.0.0

The compound intelligence layer. Every skill session teaches the system something. This skill ensures learnings compound across skills, hypotheses get confirmed/contradicted, patterns get routed to the right place, and your tenth session is measurably smarter than your first.

New in v2.0.0:

• Auto-trigger detection (recognizes when skill session ends)
• Smart routing (cross-skill pattern detection, domain classification)
• Hypothesis tracking (open questions from prior sessions confirmed or contradicted)
• Confidence scoring (patterns ranked by evidence strength)
• Self-improvement trigger detection (patterns that should become guardrails)
• Approval gates (when to ask user vs. auto-write to knowledge base)

---

Trigger

Auto-Trigger Mode (Primary):

• After any PMM skill session with output, meta-learn runs automatically
• Detects: skill name, session start/end time, outputs produced
• Does NOT trigger: mid-session reviews, skill file quality checks (use meta-review), output quality checks (use meta-verify)

Manual Trigger (Secondary):

• User says: "run learn", "/learn", "capture learnings", "log this session"
• User provides: session summary + skill name + outputs

No Trigger:

• Sessions with no output ("we just talked through it")
• Skill startup/pre-flight (not a closed session yet)
• Skill iterations (only after final output)

---

Inputs

Auto-Trigger Detection:

• Skill name (extracted from execution context)
• Session start/end timestamps
• Session outputs (documents, briefs, decisions)
• User feedback on outputs (if provided)

Manual Trigger Inputs:

• Session content (description or paste)
• Skill name (if not obvious)
• User's three extraction answers (below)

Context Files (Pre-Flight Load):

• /config/routing.yml — domain map, cross-skill signal rules
• /config/hypothesis-tracking.yml — open hypotheses, confirmation criteria
• /config/approval-gates.yml — when user approval needed
• /knowledge/INDEX.md — existing domains and file locations
• /knowledge/global/rules.md — confirmed cross-skill rules
• /knowledge/global/hypotheses.md — open questions
• /sessions/log.md — append session summary (automatic write)

NOT loaded: /foundation/brain.md (context-agnostic design)

---

Pre-Flight (Step 0)

1. Load config files (routing.yml, hypothesis-tracking.yml, approval-gates.yml)
2. Load knowledge index (INDEX.md — where does this pattern go?)
3. Load global rules (what cross-skill patterns already exist?)
4. Load open hypotheses (what questions from prior sessions could this confirm?)
5. Detect skill name (from execution context or user input)
6. Confirm session closed (outputs produced and session end detected)
7. Gate pass: If no session content, ask user for it before proceeding

---

Steps

Step 1: Auto-Detect Skill Completion (30 sec)

What: Determine if a skill session has actually ended.

How:

• Check execution context for skill name, start time, end time
• Verify outputs exist (document, brief, decisions, etc.)
• If outputs missing: ask user "Did the skill session complete?"
• If outputs present: proceed

Output: Confirmed skill name + session duration + list of outputs produced

Example:

Detected: meta-review session (15 min)
Outputs: 3 quality scores, 1 fix recommendation
Session: complete ✓

---

Step 2: Ask the Three Extraction Questions

What: Capture what the user learned that the skill didn't explicitly surface.

How: Ask all three in one message. Wait for complete answers before proceeding.

"Before we close, three quick questions:

1. What surprised you? Something unexpected — a risk, a pattern, an insight the skill surfaced that you hadn't considered before.

2. What was wrong or off? Any recommendation you pushed back on, disagreed with, or felt was missing the mark. Be specific: what and why?

3. What was missing? A gap — context the skill didn't have, a question it didn't ask, an output it didn't produce that would have been valuable."

Gate: If all three answered as "nothing", proceed to Step 6 (clean close). If any have content, proceed to Step 3.

---

Step 3: Extract Patterns & Classify (2 min)

What: Turn answers into specific, falsifiable patterns.

How:

For each answer that isn't "nothing":

1. Restate specifically — "You said [exact quote]. Let me sharpen this..."
2. Make falsifiable — If vague ("better communication"), reframe as testable:
   • ❌ Vague: "We should communicate better about milestones"
   • ✅ Falsifiable: "When timeline changes >10 days, notify stakeholders within 24h OR deals slip"
3. Classify as:
   • New Pattern — First time observed
   • Confirmed Hypothesis — Matches open hypothesis (see Step 4)
   • Contradicts Hypothesis — Disproves open hypothesis
   • Skill Gap — Patterns about what the skill missed (route to meta/gaps)
   • Cross-Skill Signal — Appears in 2+ skill domains (route to global/rules)

Output format:

Pattern 1: [Falsifiable statement]
  Surprise/Wrong/Missing: [Which question it came from]
  Classification: New Pattern / Hypothesis Confirmation / Cross-Skill Signal / Skill Gap
  Confidence: High / Medium / Low (based on evidence)
  Source: [Skill name + domain]

Example:

Pattern 1: "When champion sentiment is declining (2+ signals in Gong), deal velocity slows 40%"
  From: Wrong (skill didn't flag declining champion as deal risk)
  Classification: Cross-Skill Signal (seen in GTM-strategy + pre-mortem + gaccs-brief)
  Confidence: High (4 occurrences, spans 3 domains)
  Recommendation: Add "champion sentiment" monitor to /knowledge/global/deal-signals

---

Step 4: Check Open Hypotheses (1 min)

What: Does this session confirm or contradict any open hypothesis?

How:

1. Load /knowledge/global/hypotheses.md
2. For each pattern from Step 3, scan for matches
3. If match found: note confirmation status
   • ✅ Confirms — Additional evidence for hypothesis
   • ❌ Contradicts — Evidence against hypothesis
   • ⚠️ Partially confirms — Evidence for one aspect, not others
4. If no match: note as new hypothesis candidate

Output:

Hypothesis Check:
  Open H1: "Beachhead ICP wins faster with champion alignment"
    Status: CONFIRMED (pre-mortem session observed 3 deals lost due to weak champion)
    Evidence strength: ↑↑ (now 6 total confirmations)
    Confidence: HIGH (was MEDIUM)
    Action: Move to /knowledge/global/rules.md (threshold: 5+ confirmations ✓)

  Open H2: "Pricing elasticity varies by segment"
    Status: NO DATA THIS SESSION (pre-mortem doesn't touch pricing)
    Confidence: Still MEDIUM
    Action: Continue tracking, need more evidence

---

Step 5: Route Pattern to Correct Knowledge File (1 min)

What: Decide where this pattern belongs.

How:

1. Load /config/routing.yml — domain map

2. For each pattern, determine primary domain:

   • Single-domain (Pre-mortem-specific) → /knowledge/pre-mortem/[pattern-name].md
   • Cross-domain (appears in 2+ skills) → /knowledge/global/[pattern-name].md
   • Skill gap (something the skill missed) → /knowledge/meta/skill-gaps/[skill-name].md

3. If domain folder doesn't exist: create it (system builds incrementally)

Logic from config/routing.yml:

domains:
  beachhead:
    skills: [beachhead-segment, positioning-messaging]
    patterns_file: /knowledge/beachhead/
    cross_skill_threshold: 2  # If seen in 2+ domains, escalate to global
  gtm:
    skills: [go-to-market-strategy, gaccs-brief]
    patterns_file: /knowledge/gtm/
    cross_skill_threshold: 2
  sales:
    skills: [competitive-battlecard, retro]
    patterns_file: /knowledge/sales/
    cross_skill_threshold: 2
  global:
    patterns_file: /knowledge/global/  # Cross-skill, high-impact
    rules_file: /knowledge/global/rules.md
    hypotheses_file: /knowledge/global/hypotheses.md

Output:

Routing Decision:
  Pattern 1 (champion sentiment → deal velocity): 
    Domains touched: GTM + Sales + Pre-mortem = 3 skills
    Decision: ROUTE TO GLOBAL (cross_skill_threshold=2, met)
    File: /knowledge/global/deal-signals.md
    
  Pattern 2 (beachhead ICP pain ≠ beachhead pain):
    Domains touched: Beachhead-segment only = 1 skill
    Decision: ROUTE TO DOMAIN (beachhead)
    File: /knowledge/beachhead/ica-pain-alignment.md

---

Step 6: Detect Self-Improvement Triggers (1 min)

What: Identify patterns that should become guardrails or skill improvements.

How:

1. Check if pattern appears in /context/meta-patterns.md (guardrails)
2. If pattern has 2+ occurrences across sessions → FLAG FOR GUARDRAIL PROMOTION
3. If pattern identifies a skill gap → FLAG FOR SKILL UPDATE
4. If pattern contradicts existing rule → FLAG FOR RULE REWRITE

Output:

Self-Improvement Flags:

GUARDRAIL PROMOTION:
  Pattern: "Champion misalignment kills deals"
  Current status: Guardrail exists (added 3 weeks ago, triggered 4x)
  Recommendation: PROMOTE from MEDIUM to HIGH confidence
  Action: Update /context/meta-patterns.md confidence level

SKILL GAP FOUND:
  Pattern: "Pre-mortem doesn't ask about champion alignment pre-execution"
  Affected skill: meta-review
  Recommendation: Add question to pre-mortem Step 8
  Action: Route to meta-review improvement queue

HYPOTHESIS → RULE:
  Pattern: "Beachhead pain must match buyer pain" confirmed 6x
  Recommendation: Graduate from hypothesis to confirmed rule
  Action: Move to /knowledge/global/rules.md

---

Step 7: Proposal + Approval Gate (2 min)

What: Show user exactly what will be written, request approval (if needed).

How:

1. Load /config/approval-gates.yml — approval rules
2. Determine: Auto-write or ask approval?
   • Auto-write: Routine pattern to existing file (no threshold crossed)
   • Ask approval: New file, contradicts existing rule, or guardrail promotion
3. If approval needed: show exact text that will be written
4. Wait for explicit approval before writing

Example approval request:

APPROVAL NEEDED:

I'm about to write this pattern to /knowledge/global/deal-signals.md:

---
# Champion Sentiment as Deal Velocity Predictor
**Pattern:** When champion sentiment is declining (2+ signals in Gong), 
deal velocity slows 40% on average.
**Evidence:** 4 confirmations (GTM-brief + pre-mortem + gaccs-brief + retro)
**Confidence:** HIGH
**Recommendation:** Add "champion sentiment" monitor to all deal forecasts
**Action threshold:** If declining, escalate to champion management conversation within 48h
---

OK to write? (yes / no / edit)

Approval gates from config:

approval_gates:
  auto_write:
    - adding to existing file (same pattern type)
    - routine session log entry
    - updating hypothesis confidence (same status)
  
  require_approval:
    - new file creation
    - contradicts existing rule
    - guardrail promotion (MEDIUM → HIGH)
    - hypothesis → rule graduation
    - cross-skill signal (first time in global)

---

Step 8: Write to Knowledge Base (1 min)

What: Persist patterns to knowledge files.

How:

1. If user approved (or auto-write gate passed): write
2. If user said "no": don't write, offer to refine
3. Append mode for logs, overwrite for specific patterns
4. Format: Markdown, include evidence, date, source skill, confidence

Files written:

• /knowledge/[domain]/[pattern].md — Pattern documentation
• /knowledge/global/rules.md — Append confirmed cross-skill rules
• /knowledge/global/hypotheses.md — Update hypothesis status + confidence
• /context/meta-patterns.md — Update guardrail confidence if triggered
• /sessions/log.md — Append session summary (automatic, no approval needed)

Format example:

# Pattern: Champion Sentiment → Deal Velocity
**Date:** 2026-07-21  
**Skill Source:** meta-review, go-to-market-strategy, gaccs-brief  
**Confidence:** HIGH (4 confirmations)  
**Evidence:** 4 sessions across 3 skill domains flagged this  

## Pattern Statement
When champion sentiment is declining (2+ signals from Gong), deal velocity slows 40%.

## Why This Matters
Deal risk isn't just about price or product fit. Champion engagement is a leading indicator
of deal health. Declining sentiment predicts slower close by 4-6 weeks.

## Actionable Recommendation
When Gong flags declining champion sentiment:
1. Schedule champion alignment call within 48h
2. Review deal assumptions with champion
3. Escalate to sales leadership if sentiment doesn't recover in 1 week

## Sources
- Session 1 (2026-07-15, go-to-market-strategy): Sales team noted champion disengagement
- Session 2 (2026-07-18, pre-mortem): Pre-mortem identified champion as kill-risk
- Session 3 (2026-07-20, gaccs-brief): Messaging assumption: champion is champion ✓

**Next update:** TBD (hypothesis tracking will update if more confirmations)

---

Step 9: Update Global Hypothesis File (1 min)

What: Record which hypotheses were confirmed/contradicted this session.

How:

1. Update /knowledge/global/hypotheses.md with:
   • Hypothesis name
   • Confirmation status (CONFIRMED / CONTRADICTED / PARTIAL)
   • New evidence count
   • Confidence level change
   • Recommendation (keep tracking / promote to rule / close)

Example entry:

## H3: Champion Alignment Predicts Deal Velocity
**Status:** CONFIRMED (added 1 more confirmation)  
**Evidence count:** 6 total  
**Confidence:** HIGH (was MEDIUM)  
**Last updated:** 2026-07-21  
**Recommendation:** GRADUATE TO RULE (threshold met: 5+ confirmations ✓)  
**Action:** Move to /knowledge/global/rules.md on next synthesis cycle

---

Step 10: Log Session (1 min)

What: Record this session in the historical log.

How:

1. Append to /sessions/log.md
2. Include: date, skill, duration, patterns extracted, hypothesis updates, writes approved
3. Format: CSV or markdown table for easy trending

Example:

| Date | Skill | Duration | Patterns | Hypotheses Updated | Writes | Quality Score |
|------|-------|----------|----------|-------------------|--------|---------------|
| 2026-07-21 | go-to-market-strategy | 18 min | 3 new | H3 confirmed | ✓ global/deal-signals | 87/100 |
| 2026-07-20 | pre-mortem | 25 min | 1 new, 2 confirm | H1, H3 confirmed | ✓ beachhead/ica-pain | 91/100 |

Auto-logged (no approval): Every session generates a log entry regardless of patterns.

---

Step 11: Detect Compounding (30 sec)

What: Identify if this session's learnings compound with prior sessions.

How:

1. Check: Does this pattern appear in prior session logs (last 30 days)?
2. If yes: Note as "pattern strengthened" (confidence increases)
3. Count: How many times has this specific pattern been confirmed?
4. Threshold check: If confirmed 3+ times → flag for guardrail promotion

Output:

Compounding Check:
  Pattern: "Champion alignment → deal velocity"
  Prior observations: 2 (in last 30 days)
  Today: 1 new confirmation
  Total now: 3 confirmations
  Threshold: 2 (for cross-skill awareness), 5 (for guardrail promotion)
  Status: CROSS-SKILL RULE READY (next synthesis cycle, promote confidence)

---

Step 12: Close & Surface Next Action (1 min)

What: Confirm all writes, show what changed, suggest next step.

How:

1. Summarize: Patterns extracted, files written, hypotheses updated
2. Show impact: How many files changed? Which guardrails promoted?
3. Suggest next: "This pattern might affect [next skill session]"
4. Link to meta-synthesis: "Your 24h synthesis cycle will compound this learning"

Example:

✓ Session complete.

Captured:
  - 3 new patterns extracted
  - 1 hypothesis confirmed (H3 champion-alignment)
  - 2 files written (global/deal-signals.md, beachhead/ica-pain-alignment.md)
  - 1 guardrail promoted (champion-alignment: MEDIUM → HIGH)

Impact:
  - Your next GTM strategy session will load this champion-alignment guardrail
  - Your next beachhead session will avoid the ICP-pain mismatch
  - Next 24h meta-synthesis will see 6 total confirmations of H3 → graduates to rule

Suggested next action:
  "Before your next deal review, check: is your champion aligned on positioning?"

---

Operating Rules

1. Extract first, decide second. Always ask the three questions before classifying patterns. Don't assume.

2. Route by evidence, not intuition. Use /config/routing.yml domain map. If pattern touches 2+ skill domains, it's global (not domain-specific).

3. Hypothesis confirmation is stricter than pattern extraction. Need explicit match to existing hypothesis, not inference.

4. Approval gates prevent bad writes. Never write a pattern that contradicts an existing rule without user approval.

5. Compounding is the goal. If pattern appears 2+ times, flag it. If 5+ times, it becomes a guardrail. This is how the system gets smarter.

6. Skill gaps route to meta/improvements. If the skill missed something, don't add it to domain knowledge — flag it for skill review instead.

7. Context-agnostic extraction. Patterns come from session content alone. Do NOT load /foundation/brain.md and infer context.

8. Never force patterns. If user says "nothing surprised me", accept that. Log the session and close cleanly.

9. Logging is automatic. Every session gets logged to /sessions/log.md regardless of patterns. This is the historical record.

10. Confidence compounds. Track how many times each pattern has been confirmed. Confidence increases with repetition (MEDIUM → HIGH after 3 confirmations, becomes guardrail after 5).

---

Quality Gate

Before finalizing any write, verify:

• All three extraction questions answered
• Patterns are specific and falsifiable (not vague)
• Routing decision matches domain map (config/routing.yml)
• Hypothesis check completed (confirmed/contradicted/no-match)
• Approval gate assessed (auto-write or user approval?)
• No contradictions to existing rules without flagging for review
• Cross-skill signal correctly identified (if 2+ domains)
• Session log entry created (automatic)
• Compounding detected (if pattern appeared before)
• User explicitly approved (if approval gate required)

---

Related Files

• /config/routing.yml — Domain map, cross-skill signal rules
• /config/hypothesis-tracking.yml — Open hypotheses, confirmation criteria
• /config/approval-gates.yml — When to ask user vs. auto-write
• /knowledge/INDEX.md — Domain folder locations
• /knowledge/global/rules.md — Confirmed cross-skill rules
• /knowledge/global/hypotheses.md — Open questions
• /context/meta-patterns.md — Current guardrails
• /sessions/log.md — Historical session log