← All posts

Shipped in a Day

Qumio v1.0 went from empty repo to seven working skills in 98 commits over 15 hours. Here's what that actually looked like.

2 min readQumio
QumioAgentsWorkflow
Shipped in a Day

Qumio's first commit landed at 11:12 AM on a Saturday. The last one, number 98, landed at 2:19 AM Sunday. Fifteen hours, seven working Telegram skills, zero lines written by hand.

That's not a brag. It's what happens when an AI agent handles the entire plan-execute-verify loop and you only step in for decisions it can't make alone.

What the day looked like

The agent started by researching the OpenClaw framework, writing requirements, and generating a roadmap. Six phases. Ten plans total. Each plan got executed by a subagent in an isolated worktree, committed atomically, then verified.

My job was answering questions. Which Google scopes do you need? Where does Obsidian store its vault? Do you want rigid categories or smart grouping for the email briefing? The agent proposed, I approved or redirected, and it kept moving.

Some phases ran in parallel. While one agent wired up Google OAuth, another was scaffolding the Telegram bot config. They merged cleanly because the plans had explicit dependency tracking.

Where it got messy

Phase 3 had no formal verification step. The agent skipped it. That left a silent regression: calendar cross-referencing got removed during a rewrite and nobody noticed until the milestone audit at the end.

The news skill took three attempts. First, the main agent bypassed the worker entirely and fetched articles itself. Then the worker prompt was too long and caused the local model to compaction-stall. Third try worked after cutting the prompt from 50 lines to 28.

These weren't catastrophic failures. They were caught the same day. But they show the pattern: the agent moves fast enough that small mistakes compound if you skip verification.

The numbers

98 commits. Seven skills shipped: email digest, calendar, Obsidian notes, email-to-Obsidian, news summaries, a morning news briefing, and a knowledge browser. All PII redaction runs on a local Ollama model so personal data never leaves the machine.

The entire thing fits in a single retrospective file that's 60 lines long.