Continuous Learning Loop

"A project that doesn't learn from its own decisions is condemned to re-debate them."

TL;DR

Implements Constitution XVII: every debate and discovery produces reusable insights captured in insights/<domain>.md. Before new debates, consult existing insights. When recurring ambiguity detected, propose constitution amendments. The project compounds knowledge — it never re-debates a settled class of decision. [EXPLICIT]

Procedure

Step 1: Discover

• After any Socratic debate or significant decision, read the debate record
• Identify the three outputs: (1) direct answer, (2) refinements to original question, (3) coverage gaps
• Check insights/ for existing patterns in the same domain
• Determine if this is a new pattern or refinement of existing one

Step 2: Analyze

• Extract the reusable decision pattern (abstract from specific case)
• Classify domain: universal, security, frontend, backend, testing, deployment, governance
• Identify trigger conditions (when should future agents apply this insight?)
• Map to constitutional anchor (which principle does this insight support?)
• Check frequency: has this class of question appeared 3+ times? → constitution amendment candidate

Step 3: Execute

• Write insight to insights/<domain>.md:

  ### Insight: {title}
  - **Origin**: Debate on {topic} ({date})
  - **Pattern**: {reusable decision pattern}
  - **Rationale**: {why this is the right default}
  - **Trigger**: {when to apply this insight}
  - **Constitutional Anchor**: Principle {N}
  - **Status**: active | tentative | superseded
  

• If 3+ debates in same class → draft constitution amendment in .specify/adr/
• Update insights/README.md index with new entry
• Cross-reference debate record with insight

Step 4: Validate

• Insight is abstract enough to apply beyond the original case
• Trigger conditions are specific enough to match future queries
• Constitutional anchor is correct
• No duplicate insight exists (check before writing)
• If amendment proposed: rationale is strong, scope is clear

Quality Criteria

• Three outputs extracted from every debate (answer, refinements, gaps)
• Insight written to correct domain file
• Trigger conditions are actionable
• Constitutional anchor identified
• insights/README.md updated
• Frequency check performed (3+ → amendment)
• No duplicate insights
• Evidence tags applied

Anti-Patterns

Anti-Pattern	Why It's Bad	Do This Instead
Skipping insight extraction	Knowledge lost, same debates repeat	Extract after every significant decision
Over-specific insights	Can't apply to future cases	Abstract the pattern from the specifics
Never amending constitution	Ambiguity keeps recurring	Amend when 3+ debates hit same class
Not consulting insights before debate	Wasted effort re-debating	Always check insights/ first

Related Skills

• socratic-debate — Produces the debates that feed insights
• integrity-chain-validation — Insights improve chain compliance
• repository-organization — Insights files follow indexability rules (XVIII)

Usage

Example invocations:

• "/continuous-learning" — Run the full continuous learning workflow
• "continuous learning on this project" — Apply to current context

Assumptions & Limits

• Assumes access to project artifacts (code, docs, configs) [EXPLICIT]
• Requires English-language output unless otherwise specified [EXPLICIT]
• Does not replace domain expert judgment for final decisions [EXPLICIT]

Edge Cases

Scenario	Handling
Empty or minimal input	Request clarification before proceeding
Conflicting requirements	Flag conflicts explicitly, propose resolution
Out-of-scope request	Redirect to appropriate skill or escalate