Letting Claude act alone is tempting. But on complex projects, it often leads to silent drift. gns-plan is the skill I built to stay in the loop — before execution, not after.
The Problem with Auto Mode
Anthropic offers opusplan mode: Opus plans, Sonnet executes — automatically, without interruption. On paper, it sounds great.
In practice, this mode removes the human validation step. AI makes decisions alone, and you discover problems once the work is already done.
On complex projects, this drift can be costly. Not necessarily in bugs — but in wrong direction, questionable architecture, choices you wouldn’t have made together.
gns-plan: A Checkpoint Before Execution
I built gns-plan, a personal skill for Claude Code. Its purpose is simple: enforce a validated planning step before anything gets executed.
Here’s what it does in practice:
- Claude analyzes the project and its context
- It asks a few strategic questions — architecture, constraints, dependencies
- It suggests switching to
/planmode — I confirm - We validate the structure together before execution begins
- The plan is saved in
.local-working/with the date and project topic - Tasks are checked off directly in the file as we go
- Claude suggests switching to Sonnet for execution
That last point deserves attention.
Opus to Plan, Sonnet to Execute
Opus is the most powerful model — and the most token-intensive. It’s essential for the planning phase on complex projects. But once the plan is clear and validated, Sonnet is more than enough for execution.
This switch isn’t just about cost. It’s about method: using the right tool at the right moment. Opus during low-activity windows for strategic thinking, Sonnet for implementation.
Anthropic’s opusplan mode makes this switch automatically. gns-plan makes it with you.
The Plan as a Task Manager
An often-underestimated benefit: the saved plan replaces a task manager.
The file lives in .local-working/ inside the project, ignored by git. It’s there between sessions. Just invoke the current plan at the start of a new session to pick up exactly where you left off — without re-explaining the context.
Each checked task in the file is a progress marker, readable at a glance:
- [x] Create the base plugin structure
- [x] Configure the namespace
- [ ] Implement the REST endpoint
- [ ] Write integration tests
A Method, Not Distrust
This approach isn’t distrust of Claude Code. It’s a way of working.
AI that decides alone is fast. AI that validates with you produces durable results.
The difference between the two is human presence at the right moment — before execution starts, not after the fact.