Harry Prompter

HarryPrompter

Project Story What Inspired Us Traditional agent orchestration systems treat prompts as opaque strings—hidden from users until execution. We wanted to flip this: make the agent plan visible, editable, and iterative before a single line of code is written. What We Learned

LLM output isn't always structured — The initial YAML generation produced prose prompts that UI couldn't parse
The feedback loop is power — Users need to see, tweak, and re-approve plans; not just hit "run" and hope
Objective/Steps/Output as a contract — Structuring prompts with explicit sections enables powerful UI parsing

How We Built It name: "Orchestra" agents: main: prompt: | ## Objective Initialize project structure

## Steps
1. Create files
2. Setup config
## Output Project ready Built with FastAPI (backend), React + Zustand (frontend), OpenAI API (LLM), Git worktrees (isolation). Challenges Faced
Initial broken flow: Users saw prose prompts, fell back to raw text UI — had to send feedback just to get structured parsing
Fix: Updated FASE 6 prompt to require ## Objective/Steps/Output format from the LLM from the start
Git isolation: Using worktrees per-agent to enable parallel execution without conflicts The Vision Build a swarm where every agent's plan is a human-readable, editable contract — not just output, but intent users can verify and refine before execution begins.