Coding Tools Already Solved the Prompt Problem

Google DeepMind's new Gemini pointer reads what's under your cursor and reacts — the same context-aware trick Cursor, Copilot, and Claude Code have been quietly shipping for two years.

AI assistance is shifting from "tell it what you want" to "show it what you're working on." Google DeepMind unveiled an experimental Gemini-powered mouse pointer this week that reads the screen around your cursor — the same context-aware bet that Cursor, Copilot, and Claude Code have been making inside code editors since 2024. The pointer is news because it brings the pattern to general computing; the more interesting question is what it says about where coding tools go next.

The DeepMind demo, posted by @GoogleDeepMind on May 12, shows Gemini interpreting what's under the cursor — a scribbled note becomes an interactive to-do list, a paused video frame becomes a restaurant booking link. According to MarkTechPost, two demos are live in Google AI Studio as of May 2026: image editing and map navigation, both operable by pointing and speaking. The pitch from DeepMind, per Moneycontrol, is replacing "text-heavy prompts" with intuitive interactions.

This is exactly the bet Cursor and GitHub Copilot already shipped for code. Copilot, per a Second Talent review published this month, acts as a pair programmer that infers intent from the file you're in, the function you're editing, and the cursor's position — no prompt required. The free Windsurf IDE from Codeium takes the same approach, per MakeUseOf. The interaction model isn't novel for developers; the news is that Google is betting it's ready for everyone else.

For builders, the more useful signal is the maturing alternatives layer. Per Kunal Ganglani's May 2026 round-up, Aider, OpenHands, and Cline are now capable enough for complex multi-file work and unlock something the vendor-locked tools cannot — they let you swap Claude, GPT-5, Gemini, or a local Ollama model behind the same context-aware interface. The UX pattern is becoming a commodity. The differentiation is which model you point at, not the pointing itself.

One concrete signal of where the productivity gain lands: tests. Per a gptprompts.ai guide updated this month, Copilot can scaffold a test file in 30 seconds that would take a human 20 minutes — but the same guide warns the gain only holds when paired with a real coverage tool like Istanbul, Coverage.py, or JaCoCo to catch the gaps Copilot inevitably leaves. The AI piece is becoming reliable infrastructure; the human still owns the quality gate.

The bottom line: Pick a coding tool that lets you swap the model underneath — the interface layer is converging fast, but the model layer will keep moving for years.

Sources