Overview

HumanLayer publishes a set of Claude commands that implement an “Advanced Context Engineering” (ACE) workflow. These commands structure agent work into three phases: research, planning, and implementation.

Source

• Repository: https://github.com/humanlayer/humanlayer/tree/main/.claude/commands
• Video talk: https://www.youtube.com/watch?v=42AzKZRNhsk
• ACE document: https://github.com/humanlayer/advanced-context-engineering-for-coding-agents/blob/main/ace-fca.md

The Three Commands

research_codebase.md

Purpose: Document the codebase as-is without suggesting improvements.

Key principles:

• Spawns parallel sub-agents to research different aspects
• Uses specialized agents: codebase-locator, codebase-analyzer, pattern-finder
• Documents what EXISTS, not what SHOULD BE
• Outputs structured research documents with file:line references
• Syncs findings to a shared thoughts directory

create_plan.md

Purpose: Create detailed implementation plans through interactive iteration.

Key principles:

• Read all mentioned files FULLY before spawning sub-tasks
• Be skeptical - question vague requirements
• Be interactive - get buy-in at each step
• No open questions in final plan - resolve everything first
• Separate success criteria into automated vs manual verification
• Output to thoughts/shared/plans/YYYY-MM-DD-description.md

implement_plan.md

Purpose: Execute approved plans phase by phase.

Key principles:

• Read the plan completely first
• Follow the plan’s intent while adapting to reality
• Update checkboxes as you complete sections
• Pause after each phase for human manual verification
• If mismatch found, STOP and present the issue clearly

Core Philosophy: Frequent Intentional Compaction

From the ACE document:

LLMs are stateless functions. The only thing that affects the quality of your output is the quality of the inputs.

The workflow addresses AI limitations in complex codebases by:

1. Compacting context: Distilling searches, code understanding, and logs into structured artifacts
2. Phased approach: Research → Plan → Implement keeps each phase focused
3. Human checkpoints: Manual verification between phases catches issues early
4. Spec-driven development: Plans become the source of truth, not just the code

Why This Works

Traditional “vibe coding” problems:

• Context window fills up with searches and logs
• Agent loses track of the goal
• Produces “slop” that needs rework

ACE approach:

• Keep context utilization in 40-60% range
• Build high-leverage human review at key points
• Structured artifacts persist knowledge across sessions

Adapting for Fabbro

We store these commands in .agents/commands/ and symlink to .claude/commands/ for agent-agnostic usage. See Agent-Agnostic Commands Configuration.

The commands reference HumanLayer-specific tooling (humanlayer thoughts sync, Linear integration). For fabbro, we’d adapt:

• Use bd for task tracking instead of Linear
• Use git directly for syncing research/plans
• Store artifacts in project-local directories