Self-Improvement & Continuous Improvement

ADLC includes a closed-loop self-improvement system that ensures learnings from each session persist and configuration improves over time.

Two Feedback Loops

1. Session Retrospective (per-session)

After each work session, run:

/speckit.retrospective

This triggers:

• Kolb Learning Cycle — 4-stage structured reflection
• B/G/S Questions — 5 deep reflective questions
• Action Items — mandatory, tracked improvements
• Memory Feedback Loop — proposes updates to MEMORY.md and auto-memory

The Memory Feedback Loop closes the gap: insights from reflection flow back into persistent memory, so future sessions start smarter.

2. Configuration Improvement (periodic, HITL-initiated)

When the HITL manager wants to optimize the framework configuration:

/speckit.improve

This triggers a 6-phase cycle:

Phase	Action
0. Context Budget Audit	Calculate token usage across all layers
1. Multi-Layer Analysis	3-agent parallel review (meta-eng + PO + CA)
2. Batched Proposals	Group improvements by severity (P0-P3)
3. Human Review Gate	APPROVE / REJECT / DEFER per batch
4. Implementation	File edits + changelog + version bump
5. Self-Assessment	Scorecard metrics
6. Backlog Update	Track deferred items

How It Works

Session Work
    |
/speckit.retrospective
    |
reflective-learning skill
├── Kolb Cycle (4 stages)
├── B/G/S Questions (5)
├── Action Items (mandatory)
└── Session Context Extraction
    ├── User Preferences
    ├── System Relationships
    └── Knowledge Updates --> MEMORY.md (HITL-approved)
    |
Next Session (improved context)

/speckit.improve (periodic)
    |
6-Phase Cycle (3-agent consensus)
    |
CLAUDE.md / settings / agents / hooks improvements
    |
CLAUDE-CHANGELOG.md + improvement-backlog.md

Key Files

File	Purpose
.claude/skills/learning/reflective-learning.md	Reflection methodology
.claude/commands/speckit.retrospective.md	Session retrospective command
.claude/commands/speckit.improve.md	Configuration improvement command
.claude/CLAUDE-CHANGELOG.md	Configuration version history
.claude/improvement-backlog.md	Deferred improvement tracking

Why Continuous Improvement? (5W1H)

Why (Business Value)

Without structured reflection, each Claude Code session starts from zero context. Learnings are lost, mistakes repeat, and configuration drift accumulates. The self-improvement loop creates compounding returns: each session builds on the last, reducing errors, improving agent coordination, and increasing delivery velocity.

Without	With
Same mistakes repeated across sessions	Pattern recognition prevents recurrence
Configuration drift (stale counts, wrong paths)	Systematic accuracy via /speckit.improve
Ad-hoc memory updates (incomplete, inconsistent)	Structured extraction via Kolb + B/G/S
No feedback loop for manager corrections	Mandatory capture + action items
Agent coordination issues invisible	Reflective observation surfaces friction

What (Deliverables)

• Session retrospectives — Kolb Cycle + B/G/S reflective output with action items
• Configuration improvements — CLAUDE.md, settings.json, agent, hook changes with changelog
• Memory updates — Persistent learnings in MEMORY.md and auto-memory files
• Improvement backlog — Tracked deferred items reviewed at next cycle start

Who (Agents + HITL)

Role	Responsibility
meta-engineering-expert	Leads prompt quality, context budget, anti-pattern detection
product-owner	Validates user value of improvements, acceptance criteria
cloud-architect	Security review, compliance verification, immutable sections
HITL/Manager	Approves/rejects batched proposals, triggers improvement cycles

When (Triggers)

Trigger	Action	Frequency
Session ends	/speckit.retrospective	Every session
Manager feedback received	Mandatory Kolb Cycle on feedback	Immediately
Version release	/speckit.improve cycle	Per release
Drift detected (stale counts, wrong paths)	/speckit.improve targeted batch	As needed
New project onboarded	Bootstrap check + context audit	Once per project

Where (Evidence Paths)

Type	Path
Retrospectives	tmp/<project>/retrospectives/YYYY-MM-DD.md
Feedback logs	tmp/<project>/feedback/manager-YYYY-MM-DD.json
Improvement proposals	tmp/<project>/improvement-cycle/proposals-YYYY-MM-DD.json
Configuration changelog	.claude/CLAUDE-CHANGELOG.md
Deferred items	.claude/improvement-backlog.md

How (Implementation)

1. Capture — Structured evidence collection during session work
2. Reflect — Apply Kolb Cycle (experience → observation → conceptualization → experimentation)
3. Extract — Session Context Extraction (preferences, relationships, knowledge updates)
4. Propose — Batched improvement proposals with BEFORE/AFTER diffs
5. Approve — HITL review gate (APPROVE / REJECT / DEFER)
6. Implement — File edits with version bump and changelog entry
7. Track — Backlog update, scorecard metrics, next cycle review

Quality Gates

• Every retrospective must produce at least 1 action item
• Every retrospective must propose at least 1 knowledge update
• Every improvement cycle requires 3-agent consensus
• HITL approval required before writing to MEMORY.md or CLAUDE.md