Flowly
Log inDownload
Use cases/Creative & Building

Overnight Mini-App Builder

Brain-dump a side-project idea before bed; wake up to a working prototype the agent built and tested while you slept.

Creative & Buildingadvanced~30m setup
Tools
execwrite_fileread_filedelegate
Channels
telegramdesktop
Uses
subagentscron

The category of project that fits in one night — a pricing calculator, a markdown-to-pdf script, a simple dashboard, an Obsidian helper — shouldn't take a weekend. Tell the agent your spec at midnight, sleep, review what it built over coffee.

What it does

  • Captures your spec via Telegram before bed
  • Plans the implementation (architecture, files, tests)
  • Builds it step-by-step with self-tests after each step
  • Commits to a fresh git branch under ~/projects/overnight/<slug>/
  • Sends a morning report: what's done, what failed, link to repo

What you'll need

  • Exec + write_file + read_file — built in
  • A ~/projects/overnight/ parent directory
  • A model with strong code quality (Claude Sonnet recommended)
  • Cron for the morning delivery

Setup

1. Define the night-build protocol

Send to Flowly
When I say "overnight: <spec>": 1. Create branch: ~/projects/overnight/<slug>/ (slug from spec). git init, initial commit empty README. 2. Plan: write a PLAN.md with: - Goal (one paragraph) - Architecture (file structure, key choices) - Build order (steps 1-N, each independently testable) - Test approach Save and confirm to me before proceeding. 3. For each step: - Implement it - Write a test - Run the test - If pass: commit with message "step <n>: <one-line>" - If fail: try one fix; if still failing, commit a TODO marker and continue (don't block all subsequent steps on one bug) 4. Final smoke test: end-to-end run of the built thing. 5. Write OUTCOMES.md: what works, what doesn't, what would need another night. 6. Limit: stop after 4 hours of work or 50 LLM calls, whichever first. Better to leave a working partial than a broken whole.

2. Morning report cron

Send to Flowly
Cron "overnight-report" daily at 7 AM: If a build ran last night (look in ~/projects/overnight/ for new dirs since yesterday), summarise: - Spec - What got built (working features list) - What didn't (TODO list from incomplete steps) - Test pass rate - Path to repo - One-line "is it usable as-is?" verdict Send to Telegram.

3. Try it

Before bed:

Send to Flowly
overnight: A small CLI that takes a YouTube URL, downloads the audio, transcribes via local whisper, and saves a markdown summary to ~/notes/youtube-summaries/. Should handle playlist URLs by processing each video in sequence.

Sleep. 7 AM:

✓ overnight build: yt-summarizer

Built: download flow, transcription pipeline, markdown export, playlist handling Failed: error handling for private videos (TODO in main.py:42) Tests: 8/9 pass

Repo: ~/projects/overnight/yt-summarizer/ Verdict: usable today. Skip private videos.

Tips

  • Keep specs tight. "Build a CRM" fails. "A CLI that takes URL and emits markdown summary" works. Single deliverable, well-defined inputs and outputs.
  • Set hard limits. 4 hours of work is plenty for a mini-app. The agent must STOP rather than spiral into rebuilds.
  • Test-first matters. Without "write a test for each step", the agent claims success on broken code. The tests are what gives you confidence the morning report is honest.
  • Don't run on shared infrastructure. Overnight builds run arbitrary code. Sandbox the dir, don't let exec reach beyond ~/projects/overnight/. Add this as a pre_tool_call hook.
  • Audit the diff. Every morning, scan the commit log before using the result. Catch hallucinated APIs early.
  • It's not for production. This is for prototypes and personal tools. Real production work needs you in the loop.