Reddit user shares Claude Code workflow that routes each session to Haiku, Sonnet or Opus
A Reddit user described a workflow in which Claude Opus 5.5, at maximum effort, produced a full project plan, split it into self-contained sessions, and rated each session's difficulty to recommend a model (Haiku, Sonnet or Opus) plus an effort level. All of this was written into a single markdown build file. Each work step was then run as a new Claude Code conversation, with the file uploaded first for context and the recommended model and effort set manually. The poster says cheaper models handled boilerplate, renaming and small fixes, harder parts got the heavier model and high effort, and starting each session fresh with the file kept the context clean.
The poster describes it as using the strongest model as a project manager and cheaper models as workers, a concrete, reproducible example of agent orchestration that delegates routine tasks to smaller models. The model identity and the routing quality rest on one unverified user's account, with no independent measurement or figures on cost or time savings.
I tried a workflow on my last project and it went really well, so I'm sharing it.
What I did:
- **Planning with Opus 5.5 on max effort.** I described the whole project and asked it to make a full plan.
- **Splitting into sessions.** I had it break the plan into separate sessions, each a self-contained chunk of work.
- **Rating each session.** For every session it rated how hard the task was, then told me exactly which model (Haiku, Sonnet or Opus) and which effort level that session needed.
- **Making a build .md file.** All of this went into one markdown file covering the plan, the sessions, and the model/effort for each.
How I ran it:
* Each session is a new Claude Code conversation. * I upload the build .md file first so Claude has the full context. * I set the model and effort the plan recommended, then tell it which session to work on.
Why it worked:
* I didn't burn Opus max on simple stuff like boilerplate, renaming or small fixes. Haiku or Sonnet handled those fine. * The hard parts got the heavy model and high effort. * Starting each session fresh with the .md file kept the context clean. I wasn't dragging one giant conversation along. * The sessions were consistent because they all worked from the same plan.
It's basically using the smartest model as a project manager and the cheaper models as the workers.
Has anyone else tried something like this? I'm curious if you split sessions differently or let Claude pick the model on its own.