Workflow · Production

Windsurf Tab v2 achieves 25-75% more accepted code with variable aggression and context engineering improvements

The problem

Tab v1's system prompt was copied from an unrelated product and left unoptimized, while its one-size-fits-all prediction approach failed to account for users' fundamentally different preferences for autocomplete behavior.

First attempt

The unoptimized system prompt, containing unused tool call prompts and examples inherited from Cascade, caused context poisoning that hurt cost, time to first token, and model performance.

Workflow diagram · grounded in source
1
Keystroke triggers prediction
trigger
“4 predictive capabilities routed to on every key press”
2
Context engineering
ai_action
“fixing how we express the position of the user's cursor to the model and add related file context to improve the quality of the suggested code”
3
RL training for Tab model
ai_action
“We then proceeded to start RL for the new Tab model”
4
Code suggestion delivered
output
“Autocomplete - Standard code suggestions as you type - Supercomplete - Predicts your next multi-line edits - Tab to jump - Navigate between files with a tab keypress - Tab to import - Automatically add import statements”
5
Reward function iteration
feedback_loop
“entered a cycle of iterating on the reward function and training loops to better capture the concept of " high quality aggression" while inventing new evals across our training data to express high-taste developer intuition of Tab output”
Reported outcome

Tab v2 delivers 25-75% more accepted code and a 54% average increase in characters per predict, with variable aggression levels allowing each user to tailor autocomplete behavior to their preference.

Reported metrics
Accepted code increase25-75%
Characters per predict increase54%
System prompt prefix length reduction76%
Reported stack
WindsurfCascade
Source
https://windsurf.com/blog/windsurf-tab-2
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Tab v2 delivers 25-75% more accepted code and a 54% average increase in characters per predict, with variable aggression levels allowing each user to tailor autocomplete behavior to their preference.

What tools did this team use?

Windsurf, Cascade.

What results were reported?

Accepted code increase: 25-75%; Characters per predict increase: 54%; System prompt prefix length reduction: 76% (source-reported, not independently verified).

What failed first in this deployment?

The unoptimized system prompt, containing unused tool call prompts and examples inherited from Cascade, caused context poisoning that hurt cost, time to first token, and model performance.

How is this workflow AI workflow structured?

Keystroke triggers prediction → Context engineering → RL training for Tab model → Code suggestion delivered → Reward function iteration.