AI-generated 3D-printable pegboard from hand-drawn sketch
From sketch to 3D print in minutes
A developer created a 3D-printable pegboard toy by feeding a hand-drawn sketch into Codex with just two dimensional constraints. Instead of spending hours in Fusion 360, they got printable models in about a minute, then iterated on fit and feel through test prints.
Repository structure and contents
The GitHub repository contains parametric Python generators instead of hand-edited meshes, making modifications straightforward. Key components include:
- Seven flat play pieces with 8.45mm holes in
models/pieces/ - Four smooth gears in
models/gears/that mesh on the 40mm peg grid - Two printable boards plus a tuned peg in
models/boards/andmodels/pieces/ - Python scripts in
scripts/for generating shapes, gears, boards, and repository assets AGENTS.mdwith instructions for coding agents to extend the set
Tuned dimensions and specifications
The system uses these specific measurements after iteration:
- Grid pitch: 40.0mm
- Piece hole diameter: 8.45mm
- Gear hole diameter: 8.45mm
- Peg: 7.72mm diameter, 40.0mm long, 1.2mm end roundover
- Printed board hole diameter: 8.30mm (still being validated)
How to regenerate and modify
The setup uses standard Python tooling:
python3 -m venv .venv
. .venv/bin/activate
pip install -r requirements.txt
python scripts/generate_pegboard_shapes.py
python scripts/generate_pegboard_gears.py
python scripts/generate_pegboard_board.py
python scripts/generate_repository_assets.pyThe AGENTS.md file provides instructions for coding agents to extend the set, including building larger pegboards (like 6x6), modifying peg lengths, adding new pieces for different peg combinations, scaling the system up or down, or generating tighter/looser fit-test variants.
📖 Read the full source: HN AI Agents
👀 See Also
Autoresearch with Claude Code on Production Codebase: 60 Experiments, 3 Changes Kept
A developer ran 60 iterations of autoresearch with Claude Code on a production hybrid search system (Django, pgvector, Cohere embeddings), keeping only 3 changes with a 93% failure rate. The process identified ineffective optimizations and caught a Redis caching bug.
Self-improving AI agent plateaued due to process bloat, fixed by cutting 60% of config
A developer's self-improving AI agent hit a performance plateau as process bloat accumulated, with the writing pipeline growing to 10 steps and nightly research spending more context loading instructions than reading papers. The fix involved cutting ~60% of root config, reducing the writing pipeline from 10 to 5 steps, and restructuring the dream cycle.
Practical Lessons from Running Multiple AI Agents in Production
A team running an AI-operated store with design, coding, and marketing agents shares insights on what 'hiring' AI agents means in practice, including how to provide sufficient context for autonomous work and where agents break down differently than humans.
How One Developer Fixed 16 Architectural Weak Points in Their AI Agent System
A developer documented 16 architectural problems in their OpenClaw AI agent system and implemented specific fixes including explicit layer definitions, gateway authorization, and evidence-based execution.