Break development work into phases for Copilot
The poster splits work into phases with explicit tasks so Copilot's agent works through development steps instead of everything at once.
Done withGitHub Copilot
- What they did
- The poster breaks a feature into explicit phases in a quick note, such as building the API, wiring middleware, and adding rate limits. For each phase they run a planning task in parallel with Traycer to get file-level changes, review and edit that plan line by line, then hand the approved plan to Copilot using Sonnet 4 to write the code instead of letting it decide architecture end to end.
- How it went
- They say this gets about 90% of the work done and produces clean code most of the time, while making review easier and limiting random codebase changes. The stated downside is that blind end-to-end agent runs waste time with wrong designs, and the planning step is less useful for very small tasks.
- Worth knowing
- Their main concrete catch is cost: they cite Copilot with Sonnet 4 at about $10 for roughly 300 requests.
Try it yourself with GitHub Copilot
I'm building [project description]. Break the work into [number] phases with explicit tasks in each, and show me the plan first. Then do only Phase 1, stop, and summarize what you did so I can review it before we start the next phase.
Want help getting this working? Join the Vibe Coding Academy ↗ Run by Alex Finn, who also runs this site.
Source: Reddit · Undated
Related use cases
Run a solo company with eight agent departments
A founder set up CEO, CFO, COO, lawyer, accountant, marketing, CTO, and improver agents that share memory and consult each other.
Iterate on tasks with Copilot CLI autopilot and fleet
The author used autopilot to iterate until done and /fleet to split work across parallel subagents, noting overhead on small tasks.
Assign GitHub issues to Copilot coding agent
A developer assigned a GitHub issue to Copilot and watched it work, but it started from the default branch instead of the intended feature branch.
Build a C compiler with a team of parallel agents
Anthropic's engineering write-up on tasking Opus 4.6 agent teams with building a C compiler.
Refactor a million-line repo with 1,393 subagents
A Nous Research developer had Hermes dispatch 1,393 subagents over about 19 hours to clean up its own repo, cutting non-test Python source by 34.4% in a merged PR.
Build a mobile app turning photos and voice into reminders
The poster built Loopi, an app that turns photos, text and voice into reminders, notes and expenses, in about 45 days using Replit and AI agents.
Get 5 like this every morning
The best things people got an AI agent to do, each with the prompt to try it.