README.md

---
title: DIY Creator
emoji: 🛠️
colorFrom: gray
colorTo: red
sdk: docker
app_port: 7860
pinned: false
---

# DIY Creator 🛠️ — a multi-agent team you can watch work

Name something you want to build. A **team of three agents** hands off to each other to produce a
vetted, step-by-step build guide — and you watch them do it in real time.

> Project **#17** of the *"30 AI Projects in 15 Days"* challenge. Focus: **multi-agent orchestration.**

![Demo: the Researcher gathers a brief, the Writer drafts a guide, the Critic scores it and sends it back to revise](demo.gif)

*Above (local model): 🔍 Researcher gathers the brief → ✍️ Writer drafts a guide → 🧐 Critic scores it **5/10 and sends it back** with concrete fixes → Writer revises → the guide ships (honestly noting any minor items that remain). On a stronger backend it converges to a clean pass.*

## The team

| Agent | Role |
|-------|------|
| 🔍 **Researcher** | Gathers a practical brief: materials (with quantities), tools, techniques, safety hazards, common mistakes, and honest difficulty/time/cost. |
| ✍️ **Writer** | Turns the brief into a clear, start-to-finish guide a beginner could actually follow. |
| 🧐 **Critic** | An **independent reviewer** — scores the draft, flags issues by severity, and **sends it back**. The writer revises; the critic re-checks. The loop runs until the guide passes (or hits the round cap). |

## The lesson: roles, handoffs, and an honest gate

A multi-agent "team" is three things done well:
1. **Role-specialized agents** — each one has a single job and a prompt tuned for it.
2. **Explicit handoffs** — the researcher's brief feeds the writer; the writer's draft feeds the critic.
3. **A critique → revise loop** — nobody ships their own first draft.

Two design choices worth calling out:

- **The critic is independent (the "diamond").** It reviews the writer's work, ideally as a *different*
  model, so the review isn't the author grading itself. Set `DIY_CRITIC_MODEL` to a second local model
  for a sharper, more independent review.
- **The gate is honest-by-construction.** The critic *finds* issues (what LLMs are good at), but the
  pass/revise decision is made in **code** from issue severity — a guide passes only when no high- or
  medium-severity issues remain. The model can't rubber-stamp its own work.

## Run it

```bash
pip install -r requirements.txt
python app.py            # http://127.0.0.1:7860
# Local Ollama by default (free/private). For the cloud model instead:
#   DIY_LLM_BACKEND=claude ANTHROPIC_API_KEY=sk-... python app.py
```

**Config (env):** `DIY_LLM_BACKEND` (`ollama`|`claude`) · `DIY_OLLAMA_URL` · `DIY_OLLAMA_MODEL`
(default `qwen2.5:7b`) · `DIY_CRITIC_MODEL` (optional independent critic) · `ANTHROPIC_API_KEY` /
`ANTHROPIC_MODEL` for the Claude backend.

> Note on models: a small local model self-critiquing is genuinely noisy — it will sometimes ship
> "after review, minor notes remain" rather than a clean pass. That's honest. A stronger model (or a
> `DIY_CRITIC_MODEL` diamond) converges to a clean pass more often.

---
Built by **Robert Lucyk** · [GritAI Solutions](https://gritai.solutions) · part of the 30-in-15 challenge.