I built a useless Gemini chatbot, then added a "teaching layer" of lessons — now it works
Realistic-Middle7168 · reddit · 2026-09-05
A developer built a Gemini support chatbot for a business with Supabase-stored conversations. It sounded great on paper but answered generically and the client hated it. A comment changed everything: every time it errs, append a short lesson to a markdown file the bot reads before answering. Workflow: bot answers → human reviews/approves → improvement saved as a lesson in Supabase → bot reads lessons first next time. The bot went from embarrassingly bad to genuinely useful. Next questions: inject past conversations to learn patterns, tag responses by category — without over-engineering.
More from coding & agent
- Instinct partners with 1Password to make credential-sharing agent-friendly; 32% users already share — mon__lim · 2026-09-05
- Meta models hit 43% share on opencode, up from 3.5% two weeks ago — zephyr_z9 · 2026-09-05
- Who Sudo'd Exposes Full Call Chains Behind macOS Sudo and Touch ID Prompts, a Growing Concern With AI Agents — SIGKITTEN · 2026-09-05
- VLDB talk: Deep Research agents need co-designed retrievers, DR Tulu and AgentIR show — CShorten30 · 2026-09-05
- Watch GPT Sol (ultra) Sculpt a Banana in Blender, Start to Finish (32x Speed) — D3VAUX · 2026-09-05
- Engineers Push Back: AI Labs and the Public Confuse Web Dev With Software Engineering — gerardsans · 2026-09-05