Engineering productivitypattern

Plan mode plus hidden Claude Code model aliases for long sessions

Problem solved: Long, multi-step engineering tasks lose context or under-plan unless the right model and plan mode are selected up front.

TierBMedium — Solid coverage with concrete examples
EffortS · Days
Forengineer · founder

What

Two non-obvious Claude Code model aliases plus plan mode collectively give a long-session discipline: opusplan uses Opus for planning and Sonnet for execution; sonnet[1m] opens a 1M-token context window; plan mode (now auto-triggering in CC 2.1+) fans out 10+ parallel sub-agents on complex codebase scans.

Why it matters

Manual model switching wastes tokens and human attention. The aliases bake the planning-then-execution split into one command, and plan mode's parallel scan reportedly hits 100% on tasks (e.g. collecting 200+ API endpoints across a project) where a single-agent run from a different vendor failed.

End-to-end

  1. In Claude Code, run /model opusplan for any task that requires a plan before code edits — Opus drafts the plan, Sonnet executes each step.
  2. For tasks that genuinely need the full context (large refactor reading many files), switch to /model sonnet[1m].
  3. Trust auto-plan-mode in CC 2.1+ — it self-toggles when complexity warrants; no need to type /plan or "ultrathink".
  4. For multi-day work, keep one session and resume daily; auto-summarization preserves the thread.
  5. Treat plan-mode's spawned sub-agents as part of your token budget — they accelerate scan but multiply usage.

Prompts

/model opusplan
/model sonnet[1m]

Gotchas

  • sonnet[1m] doubles pricing past 200K tokens — verify the task actually needs more than 200K context before defaulting to it.

Tools

  • Claude Code 2.1+ (terminal — aliases not exposed in the IDE plugin)
  • Active Claude subscription or API key
← All casesB-133