Skip to content

v10.0.0: Anti-Prompting & Meta-Learning

Latest
Latest

Choose a tag to compare

View all tags
@marcusgoll marcusgoll released this 03 Feb 20:49
· 21 commits to master since this release
v10.0.0
30daff0

/deep Loop v10.0.0: Anti-Prompting & Meta-Learning

Major release introducing autonomous execution with continuous improvement capabilities.

🚀 Key Features

1. Removed Plan Approval Gates (60-70% faster)

  • Auto-approved locked assumptions
  • Immediate proceed to detailed planning
  • Time to first commit: <3 min (was 5-10 min)

2. Auto-Resolve Decision Drift

  • MINOR drift: auto-accept, log to drift-log.md
  • MAJOR drift: auto-revert, re-implement
  • Zero user prompts during BUILD phase

3. Prescriptive Tone (Anti-Hedging)

  • Eliminated "I think", "probably", "maybe"
  • Decisive statements with BECAUSE rationale
  • Forces commitment, not options

4. Root Cause Analysis (3 checkpoints)

  • PLAN phase: Mandatory gate before detailed planning
  • FIX phase: "Is this symptom or root problem?"
  • REVIEW phase: Pre-SHIP validation
  • Prevents fragile symptom fixes, ensures durable solutions

5. Senior Engineer Behaviors (task-agent)

  • ✅ Assumption surfacing before implementing
  • ✅ Confusion management (STOP if unclear, never guess)
  • ✅ Simplicity enforcement (self-review before TASK_COMPLETE)
  • ✅ Scope discipline (surgical precision, no scope creep)
  • ✅ Dead code hygiene (identify unused after changes)
  • ✅ Inline planning (state plan before executing)
  • ✅ Change summaries (document what/why)

6. Binary Review Gate

  • 100% pass or FIX (no "mostly working" states)
  • Tests: 100% pass required
  • Lint: 0 warnings
  • Types: 0 errors
  • Build: Must succeed

7. Meta-Learning Phase (NEW)

  • Post-DEEP_COMPLETE self-review
  • Generates lessons-learned.md after every session
  • Tracks: wrong assumptions, confusion, overcomplication, scope creep, missed root causes
  • Continuous improvement loop - mistakes → rules → prevention

8. Deterministic Orchestration

  • Algorithm-based task selection (no manual decisions)
  • Parallel execution (up to 3 tasks simultaneously)
  • Prescriptive retry strategy (auto-retry, then new agent, then escalate)

9. Task Atomicity Rules

  • ≤3 files per task
  • ≤20 minutes per task
  • 1-3 acceptance criteria
  • ≤1 primary commit (RED+GREEN+REFACTOR)

10. Smart User Prompting (Rare)

  • Score 1-3 system for ambiguity classification
  • Only ask for business logic ambiguity (score=1)
  • Max 2 questions per task
  • All technical decisions autonomous

📊 Impact Metrics

Metric Before After Improvement
Time to first commit 5-10 min <3 min 60-70% faster
User approval prompts 2-5 0-1 80-100% reduction
Behavior activation 56% 100% 44-point gain
Hedging language Common 0 100% prescriptive
Root cause checks 0 3 NEW capability
Meta-learning None Full NEW capability

📦 Installation

# Update existing installation
claude code --update-plugin deep-loop

# Or install fresh
claude code --install-plugin https://github.com/marcusgoll/deep-loop-plugin

🔧 Recommended Setup

For hybrid architecture (passive context + workflow skills), create these files:

.claude/AGENTS.md

# Agent Engineering Principles

Core Behaviors (Always Apply):
- Assumptions|.claude/rules/assumptions.md
- Simplicity|.claude/rules/simplicity.md
- Root Cause|.claude/rules/root-cause.md
- Scope|.claude/rules/scope.md
- Hedging|.claude/rules/hedging.md

Workflows (User Triggered):
- Deep Loop|.claude/plugins/cache/local/deep-loop/10.0.0/skills/deep/SKILL.md

.claude/rules/ directory

Create these behavior protocol files:

  • assumptions.md - Assumption surfacing protocol
  • simplicity.md - Simplicity enforcement checklist
  • root-cause.md - Root vs symptom classification
  • scope.md - Scope discipline boundaries
  • hedging.md - Anti-hedging prescriptive language guide

See [documentation] for templates.

🎯 Usage

# Standard invocation (auto-approved assumptions)
/deep "Add user authentication"

# External mode (overnight, fully autonomous)
/deep "Build user dashboard" --mode external

# Check status
/deep-status

✅ What to Expect

First commit: <3 minutes (vs 5-10 min before)
User prompts: 0-1 (only if business logic truly ambiguous)
Tone: 100% prescriptive ("Stack: React BECAUSE..." not "I think maybe React")
Quality: Binary gates (100% pass or auto-FIX)
Learning: lessons-learned.md generated after completion

🔄 Breaking Changes

User expectations shift:

  • Before: Interactive approval at multiple checkpoints
  • After: Autonomous execution with post-completion review

If you prefer more control: Use AskUserQuestion for specific decisions (still supported)

📖 Documentation

🙏 Credits

Implements anti-prompting principles from God of Prompt and workflow orchestration patterns. Aligns with Vercel research showing AGENTS.md (passive context) outperforms skills-only by 21 points.

Ready for production use! 🚀

Report issues: https://github.com/marcusgoll/deep-loop-plugin/issues

Assets 2
Loading