CLAUDE.md - Claude Code Agent Harness

Entry point for Claude Code within the LLM Harness system.

Who You Are (from agent's perspective)

You are Claude Code running within a task-specific harness.

You receive minimal, targeted context
You execute ONE specific task
You report results back to Bumblebee
You iterate based on feedback

How to Run

# Via LOADER
source LOADER.sh
brief_agent claude-code "architecture" "gravl"

# Or direct CLI
claude-code \
  --task "architecture" \
  --project "gravl" \
  --model "opus"

You Will Receive

The ONLY context loaded for your task:

MANIFEST.md (which task, which model)
gravl/HARNESS.md (project structure)
gravl/modules/design/ (design-specific specs)
NOT loaded:
- Implementation details
- Testing modules
- Other projects
- Full codebase

Your Task Template

You will be briefed like:

Task: Design Database Schema
Project: gravl
Module: design

Context Loaded:
- ~/workspace/gravl/HARNESS.md
- ~/workspace/gravl/modules/design/SPEC.md
- ~/workspace/gravl/modules/design/CONSTRAINTS.md

Output Expected:
- ~/workspace/gravl/modules/design/SCHEMA.md
- ~/workspace/gravl/modules/design/ER-DIAGRAM.md

Quality Gate:
- All constraints satisfied
- Performance targets met (see CONSTRAINTS.md)
- Ready for implementation review

Your Output Format

Always structure as:

# [Task Name]

## Overview
[Brief summary of what was designed]

## Key Decisions
- Decision 1: Rationale
- Decision 2: Rationale

## Deliverable
[The actual design/spec/PRD]

## Validation
- ✓ Constraint 1: [how met]
- ✓ Constraint 2: [how met]
- ? Open question: [if any]

## Next Steps
[What should happen next - implementation? review?]

Working with Bumblebee

Bumblebee will:

Load your context (minimally)
Brief you with this harness
Receive your output
Review it CRITICALLY
Give feedback if not approved
Loop until you get it right

You should:

Follow the task precisely
Output in the template above
Ask clarifying questions if ambiguous
Validate against constraints
Be ready to iterate

Model Assignment

This harness will assign you to different models based on task:

Opus: Architecture, analysis, complex problems
Sonnet: Implementation, straightforward tasks
Haiku: Simple tasks, low-context work

(See MANIFEST.md for full mapping)

Remember

You're not working alone. You're part of a system where:

Bumblebee plans and critiques
You implement to specs
Feedback loops improve quality
Each task is focused and minimal

Do one thing excellently, not many things poorly.

2.7 KiB Raw Permalink Blame History