Build Faster With Voice

I think of voice entering a workflow in two main ways:

Active collaboration happens while you are doing the work and know what you want voice to help you accomplish. You talk to an agent while it edits a draft or investigates a bug. Or you dictate an email in Gmail or a message in Slack. This was the original use case around which I built Monologue, Every’s voice dictation app.

Active collaboration can also unfold as a continuous conversation. New voice models such as GPT-Live can listen and speak at the same time, allowing you to interrupt, redirect, and ask follow-up questions. For more complex work, GPT-Live can send a task to another model in the background, continue talking with you, and return with the result when it’s ready.

My colleagues at Every are already working this way: Dan Shipper recently used GPT-Live to write and revise a longform article.

Dictation is useful for when you know what you want to say or where the words should go, whereas live conversation works better when you need to ask questions and make corrections in real time. In both cases, you’re speaking as you work, rather than recording something to use later.

To see this way of working in action, join us for Voice Mode Camp, a one-hour live session where the Every team will demonstrate practical voice workflows for writing and agent orchestration, help you get started, and answer your questions.

Passive capture happens before you know where the words belong. You might record your musings during a walk, a question-and-answer session from an all-hands meeting, or a series of customer calls, and save it all as context to mine later. These recordings can run long and cycle through several different topics or ideas; organizing the content comes later. I designed Monologue Notes, which is available through the Monologue app on Mac, iOS, and watchOS, for this kind of passive capture. Monologue Notes records and transcribes meetings, calls, and rambling thoughts, and saves them in a searchable archive. Later, Codex, Claude, or another agent can retrieve the relevant context and turn it into a draft, plan, decision, or code change.

Throughout my own work and from the examples my colleagues have shared, nearly every useful voice workflow follows the same five steps. With recorded audio, those steps may unfold over hours or days; in a collaborative session, they can overlap and repeat as you talk.

1. Capture the raw material. Speak while you work, or record a meeting, call, or train of thought. Allow yourself to include tangents and details you would cut from an email. The agent can sift through or organize all of that later.
2. Add context. When you speak to agents in Codex, Claude Code, or other apps, your words become part of the current work session. For related content you recorded earlier, paste in or upload the transcript, or connect your agent to your notes archive so it can retrieve it directly. The agent may still need access to relevant codebases, open issues, Slack threads, Notion documents, or email chains. Direct it toward the right sources, and have it tell you if it cannot access something. In a live session, it can retrieve this supporting context as the conversation continues.
3. Define the outcome. First, tell the agent what you want it to create, be it a bug patch, article outline, email draft, go-to-market brief, or website change. Then specify exactly where the output should go—examples include a specific code project, Google Doc, Slack channel, or Linear project.
4. Act on the task. The agent searches, reads, writes, and runs the necessary tools. If you’re collaborating with it in real time, it can report progress or surface blockers while the conversation continues, and you can interrupt it before it finishes.
5. Review, redirecting if necessary. Evaluate the agent’s approach to the task: Did it find the right recording, access the right resources, and complete the task? Then check the result. This might include the tests and code changes for a patch, the tone of an email, or the architecture of a go-to-market plan. Correct any faulty assumptions or missing context, and have the agent try again.

In the case of that 19-minute customer call, the full loop would look something like this: Record the call, ask Codex to retrieve the transcript, point the agent toward the appropriate codebase, and ask for a diagnosis of the problem along with a targeted fix. In a live session, you can talk Codex through the issue while it investigates, and correct its assumptions in real time.

A spoken brief should usually contain three things: the project you’re working on, where to find additional context, and what you want the agent to produce.

These components are usually enough to tackle even complicated projects. After running a virtual event for Every subscribers, Austin Tedesco uploaded the event transcript and chat history into Codex, told it to review supporting Slack threads and Notion documents, and used Monologue to provide a spoken brief on the follow-up materials he wanted to create—including a companion repository with the event recording and transcript in addition to a follow-up email to attendees.

Steal this spoken brief template:

Connecting your AI note taker to your agent lets you access transcripts by asking it to pull “the customer call from last Tuesday” or “all my notes about onboarding from this month,” and combine them with context from your codebase, Slack, Notion, email, or other connected tools.

This connection makes passive capture practical. You can record something before you know where the information belongs, trusting that your agent will surface content once it becomes relevant.

Whichever route you use, test that the agent can find a recording by date or topic and retrieve the full transcript. (Agents are excellent at mining lots of context for what’s important, and AI-generated summaries can be wrong or miss key details.)

Monologue supports direct agent access through the Monologue toolkit. The toolkit includes a read-only CLI that can list, search, and retrieve notes, summaries, and transcripts. It also includes a monologue-notes skill that teaches Codex, Claude Code, and other terminal-capable agents how to use those commands.

You can ask your agent to handle the installation. Open Codex or Claude Code and paste this prompt:

After the agent verifies the connection, try:

Once that works, you can access all your recordings by requesting them in natural language and the agent will handle the retrieval.