Skip to content
Claude CodegreenfieldWeb18 files · +1319 −560

Plan and scaffold a Gradio hackathon app for tuition teachers

Session from tutordesk-agent-traces on Hugging Face by naazimsnh02 · MIT · condensed by Coders Talk

48msession
0interventions
GOAL

Evaluate a hackathon PRD for winning potential given sponsor/track constraints, then refine the plan and scaffold a local, agentic Gradio app (TutorDesk AI) built around a ≤32B model.

  1. Prompt · +0
    Check the hackathon rules and our PRD, research current state, and give feedback on whether this idea can win.
  2. Agent did · +0m
    Researched the hackathon and current small-model landscape, then gave feedback that the PRD reads like a product roadmap rather than a hackathon entry and needs aggressive scope cuts to win.
  3. Prompt · +10m
    We also need to cover sponsor bonus awards (Modal, OpenBMB, NVIDIA Nemotron, Llama Champion) — work out which ones fit our plan.
  4. Agent did · +11m
    Verified current specs for OpenBMB and NVIDIA sponsor models, then mapped out which sponsor badges conflict (e.g. local-only vs Modal cloud) and which single model family the team should commit to.
  5. Prompt · +34m
    Clarified agent-trace badge meaning, decided to go fully local (no Modal), and asked to check if the new MiniCPM-V 4.6 model (thought to be ~1B params) could work, plus project the app as agentic.
  6. Agent did · +34m
    Confirmed MiniCPM-V 4.6 is ~1.3B parameters (Qwen3.5-0.8B backbone + vision), unlocking the Tiny Titan badge, and updated the strategy to lead with this model plus an agent-trace publish plan.
  7. Prompt · +36m
    Update the PRD for the new plan and create CLAUDE.md, implementation_plan.md and progress.md to track implementation.
  8. Agent did · +37m
    Rewrote PRD.md to v2.0 scoped to a single 'Weekly Teaching Pack' flagship feature on MiniCPM-V 4.6, and created CLAUDE.md, implementation_plan.md and progress.md documenting the build schedule and award strategy.
  9. Prompt · +45m
    Yes, scaffold the code.
  10. Agent did · +45m
    Scaffolded the full project: requirements.txt, config.py model swap point, curriculum data, llm/client.py llama.cpp wrapper, agent/tools+orchestrator with plan-act-verify loop and trace, pack renderer, and the Gradio app.py.
  11. Fail · +48m
    End-to-end smoke test crashed with a UnicodeEncodeError from an emoji in a print statement on Windows' cp1252 console.
  12. Outcome · +48m
    Fixed encoding by forcing UTF-8 output; verified the agent pipeline runs end-to-end in mock mode (4 sections, 2 answers, 2 verify steps) with PDF export gracefully falling back to HTML, and updated progress.md to reflect the completed scaffold.
VERDICT

I'd start next time by pasting in the full sponsor/bonus badge list alongside the PRD in the first prompt, since that reshaped the whole plan later and cost a lot of back-and-forth. The correction that did the most work was insisting on verifying the MiniCPM-V 4.6 parameter count myself instead of trusting assumptions — that single fact unlocked the Tiny Titan badge and reset the whole model strategy. The scaffold ran cleanly on the first real test once we worked around a trivial Windows console encoding issue.

I tried this · 0

Discussion · 0

How Builds work