Draft Mode
Two modes, two trust zones
Split AI work into two explicit modes. In Draft: a cheap or no-side-effects pass, ideal for trying ten ideas in the time of one. In Live: the production model, real outputs, real cost. The 2026 industry has settled on "Plan Mode" terminology (Cursor, Claude Code) — the underlying move is the same: a deliberate, visible promotion from sketch to commit.
Framing
The problem
Iterating with a production-grade agent burns real tokens and produces real side effects, even when the user is just exploring.
The pattern
Offer two explicit modes: a cheap, sandboxed Draft for exploration and a committed Live for the final pass. Make the mode visible, the cost asymmetric, and promotion a deliberate act.
Why chat breaks here
Chat treats every turn the same — same model, same cost, same persistence. There is no surface for cheap exploration that the user can leave behind.
Risks
If Draft outputs look identical to Live, users may trust them as final; weak mode signalling can blur the boundary.
Avoid when
The task is short and a single-tier model already serves both exploration and commit cheaply.
Use when
Iteration is exploratory and the production pipeline is too expensive to use as a sketchpad.
DOPE evaluation
- Directability
- Iterate as many times as needed in Draft; promote a chosen one to Live
- Observability
- Mode is always visible; sandbox banner makes the trust zone unmissable
- Predictability
- Draft outputs are explicitly marked as cheap-model approximations
- Explainability
- Each mode declares its model, cost profile, and whether side effects apply
In the wild
- Cursor Plan Mode (Cursor) — Shift+Tab toggles a no-edit planning mode where Cursor proposes a plan but cannot write files; user explicitly promotes to Agent mode for execution. Strongest 2026 example.
- Claude Code Plan Mode (Anthropic) — Shift+Tab toggles a no-edit planning mode in the terminal. Claude proposes a plan; user promotes to execute. Same paradigm as Cursor — the 2026 industry settled on Plan vs Agent terminology.
- Replit Agent Sandbox (Replit) — Caveat — captures the side-effect-free preview half but not the cheap-model half. Sandbox-to-Deploy is the explicit promotion act; Effort-Based Pricing makes cost asymmetric.
FAQ
When should I use the Draft Mode pattern?
Iteration is exploratory and the production pipeline is too expensive to use as a sketchpad.
When should I avoid the Draft Mode pattern?
The task is short and a single-tier model already serves both exploration and commit cheaply.
What problem does Draft Mode solve?
Iterating with a production-grade agent burns real tokens and produces real side effects, even when the user is just exploring.
Why is chat the wrong fit for this?
Chat treats every turn the same — same model, same cost, same persistence. There is no surface for cheap exploration that the user can leave behind.
Related patterns
- Often paired with: Compute Budget — Two trust tiers + a hard cap. The drafts stay cheap, the live run is bounded, the workflow is honest about both.
- Often paired with: Parallel Alternatives — Draft Mode is one-after-another exploration; Parallel Alternatives is many-at-once. Same goal, different shape.
- Alternative to: Intent Preview — Preview the plan before commit vs run a cheap draft to see the shape. Both reduce surprise at commit time.