Backlog agent

Feature ideas show up at the worst moments: in a meeting, on the road, mid-conversation with a user. By the time I'm at a desk half the context is gone, or the idea is already in the backlog under another name.

• Telegram
• Gemini
• TypeScript
• Google Sheets
• Google Docs
• Cloud Run

1. Voice note or textVoice is transcribed automatically
2. UnderstandThe agent checks it's about the product, and asks at most one clarifying question
3. Duplicate checkTwo passes: word match, then meaning. Catches “דארק מוד” vs “Dark Mode”
4. Proposal card✅ / ❌ buttons. I decide
5. WriteA new backlog row with an ID, and a four-part spec in a new tab of the specs doc

The agent understands and proposes, the human approves, and only a deterministic pipeline writes.

The agent itself has no write access. That's the key product decision: in a real backlog, one wrong row costs more than all the time the bot saves.

specs written by the bot

8

minutes of chat memory

30

build phases

9

What broke in production

The same voice note was processed 8–10 times, because the retry wrapped the whole flow instead of just the model call. I also tried a faster, cheaper model for the agent and dropped it: its tool calls weren't reliable. The rule that stuck: correctness before latency.