How Notion uses Codex to one-shot specs, build AI Voice Input for the web, and multiply engineering power across small teams.
Notion is a company that has long prided itself on the flexibility of its building blocks. But behind the scenes, the engineering team is increasingly turning to AI to accelerate how those blocks are assembled—not just for writing code, but for translating high-level product instincts into shippable features.
The results have been dramatic. In a recent internal showcase, the team demonstrated how they used OpenAI’s Codex to turn a rough product spec into a working implementation in a single pass—a process they call “one-shotting” the spec. This isn’t about replacing engineers; it’s about removing the friction between thinking and doing.
From spec to code in one shot
The core workflow is deceptively simple. Instead of writing a detailed technical design document and then handing it off to a developer, Notion engineers now write a concise, intent-focused spec—often just a few paragraphs describing the user problem, the desired behavior, and the constraints. They then feed that spec directly into Codex, which generates the initial implementation.
“The key is that we’ve stopped treating the spec as a contract and started treating it as a prompt,” explains one senior engineer. “Codex is good enough to handle the boilerplate and the obvious patterns. Our job is to provide the sharp edges—the non-obvious product decisions and the edge cases.”
This approach has cut the time from idea to prototype from days to hours. For a small team, that multiplication of engineering power is transformative.
Case study: AI Voice Input for the web
One of the most visible examples of this workflow is the recent launch of AI Voice Input for Notion’s web app. The feature allows users to dictate notes, tasks, and even full documents directly in the browser, with the AI handling transcription, punctuation, and basic formatting.
The project started with a simple spec: “Let users speak naturally, and have Notion turn that into clean, structured content.” The engineering team used Codex to generate the initial audio capture logic, the Web Speech API integration, and the fallback handling for unsupported browsers.
But the real magic happened in the iteration loop. Instead of writing unit tests manually, the team asked Codex to generate test cases based on the spec’s edge cases—like handling long pauses, background noise, or switching languages mid-sentence. The AI even suggested a graceful degradation strategy when the user’s microphone permission is denied.
The result? A feature that would typically require a dedicated front-end engineer for two weeks was built and shipped in under three days, with the same engineer handling the review and polish.
Multiplying power across small teams
Notion’s engineering organization is deliberately lean. Many teams consist of just two or three engineers responsible for a full product area. In this environment, Codex acts as a force multiplier, not just for writing code, but for:
- Generating boilerplate for new API endpoints, database migrations, and UI components.
- Refactoring legacy code to match new patterns, with the AI proposing the diff and the engineer reviewing it.
- Writing documentation and inline comments that explain the “why” behind complex logic.
- Creating realistic mock data for testing and demos, saving hours of manual setup.
One team leader noted, “It’s like having a junior engineer who never sleeps and never complains. But you still need to know what to ask for. The senior engineers are the ones who write the good specs.”
What this means for the future
Notion is careful to point out that Codex is not a replacement for human judgment. The AI still struggles with deep architectural trade-offs, cross-system dependencies, and product taste. But by handling the 90% of work that is mechanical, it frees engineers to focus on the 10% that is genuinely creative.
“The best spec is not the one with the most detail. It’s the one that clearly communicates the intent, so the AI can fill in the blanks without asking questions.” — Notion Engineering Lead
For small teams, this shift is existential. It means a two-person team can ship what used to require five people. It means the gap between “I have an idea” and “it’s live in production” is now measured in hours, not sprints.
As Codex and similar tools improve, Notion expects this pattern to become the default. The spec becomes the product, and the engineer becomes the editor—curating, correcting, and elevating what the AI produces. The result is not less engineering, but more engineering per human.