Every

Codex Runs My Inbox Now

3230 summary words 14 min summary Watch video

Start with the signal

14 min read

Summary

At-a-Glance

  • Verdict: Watch fully
  • Core thesis: Codex (an AI agent with in-app browser, file system access, and calendar/email integrations) enables inbox zero by auto-drafting replies, scheduling meetings, and learning user preferences from recorded decisions, turning email and company information feeds into conversational task cards.
  • Why it matters: Demonstrates a working compound-AI workflow where an agent maintains state on the file system, learns from user decisions, and handles cross-app orchestration (email, Slack, Notion, calendar) in natural language—relevant for operator workflows, agent system design, and productizing AI copilots.
  • Best use: Watch the demo to see the card-based review interface, the prompt engineering for tracking decisions, and the end-to-end 'vibe coding' flow; use the transcript to extract the architecture (state on file system, Codex native apps, auto-review mode) and the exact prompt for building an inbox sweep app.

Executive Summary

The speaker achieved inbox zero for 13 consecutive weeks by building a Codex-powered inbox app. Codex is an AI agent platform with an embedded browser that accesses the user's computer, apps, email, calendar, and file system. The speaker 'vibe coded' a custom app that sweeps the inbox several times per day, generates a card for each email summarizing the content and drafting a reply, and allows the user to scroll through cards while conversing with Codex to modify drafts, propose meeting times (Codex reads the calendar), or archive emails. The user queues actions, Codex executes them (sending emails, posting to Slack, etc.), and all decisions are recorded on the file system so Codex compounds learnings over time—improving at surfacing relevant items, hiding noise, and proposing next actions.

Beyond email, the speaker built a 'company feed' that aggregates transcribed meetings from Notion and Slack messages across a 30-person company, presenting them as cards with statuses (needs review, Codex to-do, in progress, done). The user can review internal debates (e.g., product naming, positioning) and instruct Codex to reply in Slack threads they aren't in. The architecture is Codex-native: state lives on the file system, the web app renders in Codex's in-app browser, and prompts evolve based on recorded user preferences. The speaker uses auto-review mode (Codex 5.5, extra-high setting) to build apps end-to-end with minimal manual coding.

The speaker demonstrates live prompt engineering, showing how to ask Codex to build an inbox sweep app from scratch: specify email connection (Quora CLI or Gmail), card-based UI, next-action proposals, conversational interaction, file-system state management, and validation steps. Codex sets its own goal, writes code, and iterates until the app works. The video is promotional for Every (a newsletter/product lab) and Quora (the product), with a promise that Codex-native inbox flows will ship in the next couple of months. The key insight is that this workflow removes procrastination friction—lookup, scheduling, context-switching—by embedding an agent in every app and compounding decision history.

Key Takeaways

  • Claim: The speaker hit inbox zero for 13 weeks straight using a Codex-powered inbox sweep app. | Evidence: At 2 PM the inbox has only 30 emails, and the speaker credits Codex for enabling consistent inbox zero by automating scheduling, research, and reply drafts—tasks the speaker would procrastinate on. | Caveat: The speaker does not quantify baseline email volume, time saved per week, or error rates (e.g., incorrect meeting proposals, misclassified emails). The 13-week streak is anecdotal; no comparison to previous productivity metrics is provided. | Implication: For operators managing high-volume inboxes, embedding an agent that drafts replies and proposes calendar-aware meeting times can convert procrastination triggers (context switching, lookup overhead) into quick review-and-queue workflows. The file-system-based learning loop means the system improves with use, making it a compounding productivity asset. | Timestamp: 00:00
  • Claim: Codex has an in-app browser that lets you do any browser-based work with an agent sitting alongside you, accessing your computer, apps, and files. | Evidence: The speaker demonstrates the Codex interface with a web app running in the embedded browser. Codex can access email (via Quora CLI or Gmail API), read the user's calendar to propose meeting times, post to Slack, pull from Notion, and write state to the local file system. | Caveat: The transcript does not clarify security boundaries (e.g., whether Codex can access arbitrary files or only designated directories), authentication flows for third-party apps, or latency when the agent orchestrates multiple API calls. The demo shows manual 'execute' commands, not fully autonomous execution. | Implication: This architecture enables cross-app orchestration without traditional RPA or integration middleware. For agent system builders, the in-app browser + file system state model avoids fragile cloud state and allows local-first, version-controlled prompt/decision history. For non-technical users, it means building custom workflows via natural language instead of Zapier/Make. | Timestamp: 00:24
  • Claim: All decisions (archive, draft, schedule, etc.) are recorded on the file system, and Codex uses this history to improve prompts over time. | Evidence: The speaker shows the file system storing prompts and decision logs. Codex 'watches what I do' and refines prompts to surface interesting emails, hide noise, and propose the right next action. The speaker calls this 'compounding my learnings.' | Caveat: The speaker does not reveal the prompt evolution mechanism (e.g., retrieval-augmented generation, few-shot learning from past decisions, or manual prompt editing). The quality of compounding depends on labeling consistency and edge-case handling, which are not discussed. | Implication: For AI operators, this suggests a pattern: persist user feedback in a structured format (file system, SQLite, etc.) and feed it back into the agent's context window or fine-tuning pipeline. The local-first approach avoids vendor lock-in and enables versioned prompt history. For product builders, it implies that agent UX should make decision capture lightweight (swipe/queue UI) to maximize training signal. | Timestamp: 04:11
  • Claim: The company feed aggregates meetings (transcribed in Notion) and Slack messages into cards, letting the user review and respond to internal debates without attending every meeting or reading every thread. | Evidence: The speaker shows cards for positioning debates (e.g., 'archive' vs. 'mark as done' terminology) and instructs Codex to post the user's vote to Slack on behalf of them. The feed is separate from the inbox feed and runs in a dedicated Codex thread for a 30-person company. | Caveat: The transcript does not explain how Codex identifies 'debates' or surfaces relevant cards vs. noise (meeting summaries, Slack chatter). There is no mention of permission controls—e.g., whether Codex can post to any Slack channel or only those the user moderates. The demo implies manual review before execution, not autonomous posting. | Implication: For company operators, this pattern reduces meeting overload and Slack fatigue by surfacing decision points that require executive input. The agent acts as a chief-of-staff, triaging internal conversations and drafting responses. For GTM/product teams, it suggests a product opportunity: company-wide feeds with role-based filtering (exec summary vs. IC detail) and approval workflows before posting. | Timestamp: 02:14
  • Claim: Non-technical users can prompt Codex to build a complete inbox sweep app end-to-end by specifying requirements in natural language and using auto-review mode (Codex 5.5, extra-high). | Evidence: The speaker types a prompt asking Codex to build an inbox app with email connection (Quora CLI or Gmail), card UI, next-action proposals, conversational interaction, file-system state, and validation steps. Codex sets a goal ('build end-to-end Codex native inbox feed app') and will autonomously code, test, and iterate. The speaker says 'I can just leave it' and come back to a working app. | Caveat: The demo cuts before showing the final app output, so the actual success rate, debugging iterations, and edge-case handling are unknown. The speaker does not specify how long 'leave it' means (minutes? hours?). The 'extra-high' setting and 'unlimited budget' imply high token cost and possibly multiple model calls, which may not be feasible for all users. | Implication: For AI product builders, this demonstrates that agentic coding tools can compress the prototype-to-working-app timeline for internal tools, especially when requirements are well-structured and the user is willing to iterate. For non-technical operators, it lowers the barrier to custom workflow automation—no need to hire a developer for one-off tools. For investors/strategists, it suggests that no-code/low-code incumbents face disruption from conversational IDEs that generate code, not just workflows. | Timestamp: 05:31

Detailed Brief

Codex Architecture & Workflow

  • Claims: Codex has an embedded browser that allows users to do browser-based work while an agent sits alongside with access to computer, apps, and files.; State is stored on the local file system, not in the cloud, making apps 'Codex native.'; The inbox sweep runs multiple times per day, generating a card for each email with summary and draft reply.; The user scrolls through cards, converses with Codex to modify drafts or propose actions (e.g., 'meet Doug in Brooklyn'), and queues actions for batch execution.
  • Evidence: Demo shows Codex interface with a web app in the in-app browser.; Speaker mentions 'all the state should be on the file system' and shows file system storing prompts.; Demo shows cards for scheduling (Doug, Greg Eisenberg), sponsorships, and intros.; Speaker says 'I can just say OK, execute' and Codex will send emails, do research, etc.
  • Caveats: No discussion of error handling (e.g., wrong meeting time, misclassified email, API rate limits).; Security model not explained: does Codex have full file system access or sandboxed directories?; Batch execution requires manual 'execute' command; not fully autonomous.; No mention of how Codex authenticates to Gmail, Slack, Notion, or calendar APIs.
  • Implications: Local-first state enables version control, data portability, and avoids vendor lock-in.; Card-based UI + conversational queue model reduces decision fatigue and context switching.; Cross-app orchestration (email + calendar + Slack) requires robust API integrations and auth flows.; For operators, this pattern suggests building lightweight review interfaces (cards, queues) for any high-volume workflow (support tickets, sales leads, content moderation).

Company Feed & Information Aggregation

  • Claims: The company feed aggregates transcribed meetings (from Notion) and Slack messages into cards.; Each card has a status: needs review, Codex to-do, in progress, or done.; The user reviews internal debates (e.g., 'archive' vs. 'mark as done') and instructs Codex to post responses to Slack threads they aren't in.; The feed is separate from inbox, managed in a dedicated Codex thread for a 30-person company.
  • Evidence: Demo shows cards for positioning debates (Every plans, Quora terminology).; Speaker says 'tell Brandon and Kieran I vote for mark as done' and Codex will post to Slack.; Speaker mentions meetings and Slacks 'I'm not in,' implying async review of company-wide activity.
  • Caveats: No explanation of how Codex surfaces 'debates' vs. routine updates (spam filtering logic unclear).; Permission model not addressed: can Codex post to any Slack channel? Is there approval before posting?; The demo shows manual review + queue, not autonomous decision-making or posting.; Scalability unclear: does this work for 100-person or 1,000-person companies? Transcription quality and Slack volume may overwhelm card generation.
  • Implications: For executives, this creates a personalized 'chief of staff' feed that surfaces decision points without attending every meeting.; For product teams, it suggests a horizontal opportunity: build role-based feeds (exec summary vs. IC detail) with AI-powered triage.; For GTM, this pattern reduces Slack overload by converting threaded discussions into reviewable cards with proposed responses.; For operators, it demonstrates how to use meeting transcripts + Slack logs as training data for agent prompt refinement.

Learning Loop & Prompt Evolution

  • Claims: All user decisions (archive, draft, schedule) are recorded on the file system.; Codex 'watches what I do' and refines prompts over time to surface relevant emails, hide noise, and propose next actions.; The speaker calls this 'compounding my learnings,' implying that the system improves with each review cycle.; Prompts are stored in the file system, visible to the user.
  • Evidence: Demo shows file system with stored prompts.; Speaker says 'it knows for any given email, what did I say to do' and 'it gets better over time.'; Visual shows prompt files labeled by decision type (archive, interesting, draft).
  • Caveats: The mechanism for prompt evolution is not explained (RAG, few-shot examples, manual edits, or fine-tuning?).; No metrics on improvement (e.g., reduction in false positives, increase in draft acceptance rate).; Quality of compounding depends on labeling consistency—if user decisions are noisy, the system may learn bad patterns.; No discussion of how to reset or override learned patterns if the agent drifts (e.g., starts archiving important emails).
  • Implications: For AI operators, this suggests persisting user feedback in a structured format (JSON, CSV, SQLite) and feeding it back into the agent's context or fine-tuning pipeline.; For product builders, it implies that agent UX should make decision capture lightweight (swipe, upvote, queue) to maximize training signal.; For workflow automation, it demonstrates that local-first state + versioned prompts enable rollback and debugging of agent behavior.; For investing, it suggests that tools with compounding learning loops (vs. stateless APIs) will have stronger moats and retention.

Demo: Building an Inbox App in Codex

  • Claims: The speaker demonstrates prompting Codex to build an inbox sweep app from scratch using natural language.; The prompt specifies: email connection (Quora CLI or Gmail), card UI, next-action proposals, conversational interaction, file-system state, and validation steps.; Codex is set to auto-review mode (5.5, extra-high) with an 'unlimited budget' and will set its own goal, code the app, and validate it.; The speaker says 'I can just leave it' and come back to a working app, emphasizing low manual coding effort.
  • Evidence: Demo shows the speaker typing a prompt into Codex Monologue: 'I want you to help me make an inbox sweep app... build that end to end.'; Visual shows Codex setting a goal: 'build an end-to-end Codex native inbox feed app in my folder.'; Speaker mentions using Codex 5.5 with extra-high setting for this experiment.
  • Caveats: The demo cuts before showing the final app, so success rate, debugging iterations, and time to completion are unknown.; The 'unlimited budget' and 'extra-high' setting imply high token cost, which may not be feasible for all users or experiments.; The prompt is detailed but assumes Codex can infer UI/UX decisions (card layout, swipe mechanics, error states) not explicitly specified.; No mention of how Codex handles API authentication (e.g., OAuth flows for Gmail) or error handling during autonomous coding.
  • Implications: For non-technical operators, this lowers the barrier to custom workflow automation—no need to hire a developer for one-off tools.; For AI product builders, it demonstrates that agentic coding tools can compress prototype-to-working-app timelines, especially for internal tools with well-structured requirements.; For no-code/low-code platforms, this suggests disruption risk from conversational IDEs that generate code instead of workflow graphs.; For investors, it signals that agentic coding (Codex, Cursor, Replit Agent) may shift value from dev tools to AI-native platforms that combine IDE, runtime, and agent orchestration.

Notable Concepts & Terms

  • Codex: AI agent platform with embedded browser, file system access, and integrations (email, calendar, Slack, Notion). Enables users to build and run 'Codex native apps' with state on the local file system, conversational UI, and cross-app orchestration.
  • Codex native app: Web app designed to run in Codex's in-app browser, storing state on the user's file system (not cloud) and rendering UI in the browser while the agent handles backend logic, API calls, and decision execution.
  • Vibe coding: Informal term for building apps by conversationally prompting an AI agent (e.g., Codex) to generate, test, and iterate code with minimal manual coding. Emphasizes low-friction prototyping and natural language requirements over traditional software engineering.
  • Inbox sweep: The automated process where Codex reads all emails in the inbox, generates a summary card for each, drafts a reply, and presents the cards for user review. The user queues actions, and Codex executes them (send, archive, schedule).
  • Company feed: A Codex-powered feed that aggregates meeting transcripts (from Notion) and Slack messages into reviewable cards, surfacing internal debates and decisions for the user to triage and respond to, even if they didn't attend the meeting or read the thread.
  • Compounding learnings: The speaker's term for Codex's learning loop: user decisions (archive, draft, schedule) are recorded on the file system, and Codex refines prompts over time to improve at surfacing relevant emails, hiding noise, and proposing next actions.
  • Auto-review mode: Codex setting (5.5 with 'extra-high') where the agent autonomously codes, tests, and iterates on a goal with minimal user intervention. Implies high token budget and possibly multiple model calls.
  • Quora / Quora CLI: A product (or CLI tool) from the Every team that integrates with email and other apps. The speaker mentions it as an alternative to Gmail for connecting the inbox sweep app. Context suggests it's part of the Codex ecosystem.
  • Every / Monologue: Every is a newsletter/product lab focused on AI and the future of work. Monologue is a product within Every (possibly a Codex chat interface or thread type). The speaker promotes subscribing to Every for updates on Codex-native apps.

Operator Notes / Why Ken Should Care

  • For agent system design: the file-system state + in-app browser architecture enables local-first, version-controlled agent workflows without cloud vendor lock-in. Consider this pattern for internal tools that require cross-app orchestration (email, calendar, Slack).
  • For AI product builders: the card-based review UI + conversational queue model reduces decision fatigue and context switching. This pattern is replicable for high-volume workflows (support tickets, sales leads, content moderation).
  • For workflow automation: the demo shows that agentic coding tools (Codex 5.5, auto-review mode) can compress prototype-to-working-app timelines for internal tools. Evaluate whether your team should adopt conversational IDEs over traditional no-code platforms.
  • For GTM/sales ops: the inbox sweep + calendar integration demonstrates how to remove procrastination friction (lookup, scheduling) by embedding an agent in the workflow. Consider similar patterns for CRM hygiene, lead follow-up, and meeting prep.
  • For investing: the compounding learning loop (decisions recorded → prompts refined → better next actions) suggests that tools with persistent state and feedback loops will have stronger retention and moats than stateless APIs. Watch for local-first AI tools vs. cloud-native incumbents.
  • For content/media ops: the company feed pattern (aggregating meetings + Slack into triaged cards) reduces information overload for executives and content strategists. Consider building similar feeds for editorial calendars, audience feedback, and competitive intel.
  • For business operations: the demo implies high token cost ('unlimited budget,' 'extra-high' setting) for autonomous app building. Evaluate ROI of agentic coding vs. hiring contractors for one-off tools, especially if the app requires ongoing maintenance.

Watch Map

  • 00:00: Intro: inbox zero for 13 weeks straight using Codex
  • 00:24: Codex in-app browser + agent architecture explained
  • 00:56: Demo: inbox sweep cards (Doug scheduling, sponsorships, intros, Greg Eisenberg calendar)
  • 02:14: Company feed: aggregating meetings (Notion) and Slack into cards, responding to internal debates
  • 03:05: Batch execution: user queues actions, Codex sends emails and posts to Slack
  • 04:11: Learning loop: decisions recorded on file system, prompts refined over time ('compounding learnings')
  • 05:11: How to do this yourself: use Quora or build with Codex Monologue
  • 05:31: Live demo: prompting Codex to build inbox sweep app end-to-end (auto-review mode, Codex 5.5 extra-high)
  • 06:52: Call to action: subscribe to Every, check out compound engineering workflow

Source/Metadata

  • Title: Codex Runs My Inbox Now
  • Transcript words: 3273
  • Duration seconds: 433
  • Timestamp note: Timestamps provided from video duration; transcript does not include explicit chapter markers but section breaks are clear from demo flow.
Search this video Resume transcript at 0%
Full transcript 1699 words · 14 min read
0:00

SPEAKER_00

This is my inbox. There's 30 emails in there and it's 2 p.m. That's crazy. That never happens.

0:07

SPEAKER_00

And what's crazy about this is I've hit inbox zero for 13 weeks in a row. That has literally never happened before in my life. And Codex has just completely changed how I'm able to process information that comes into my email and allows me to spend less time on stuff like scheduling and stuff that I would procrastinate on and more time on doing fun creative projects like building more stuff with Codex. I want to show you how I do it. So what happens is I have a long running thread called inbox. And what people don't really know about Codex is it has an in-app browser in it. And what that allows you to do is do all of your work. Any work that you would do in a browser, you can do it in Codex over here. And then you have an agent that you can bring to all your work. So you're going back and forth and collaborating with a super smart agent that has access to your computer, access to all your apps, access to everything it might need to get work done. And it's sitting with you on any app that you want to use. And in this case, this is actually an app I built. So the other cool thing about Codex is it can just make stuff. I just vibe coded this app that lets me do my email with Codex. So let me explain how it works.

0:14

SPEAKER_00

So a few times a day, Codex will sweep through all the emails in my inbox. And for each email, it will give me a little card. The card looks like this. It tells me a little bit about the email. It lets me see the email if I want. And Codex also drafts a reply. And what I can do is I can just say, hey, can we actually, maybe I'd like to meet with Doug near my office. Can we propose somewhere in Brooklyn and see if he's okay with that? And now I'm on to the next email. So new card, new scheduling thing. And now we can just keep going. Here's someone who's asking for sponsorships, and I want to sponsor. Codex will automatically draft a reply to the right person. I can say, this looks good. Now we're on to the next one. Now someone else who I want to make an intro for, I can just say, I already made the intro. You can archive this one. Next card. And here it's Greg Eisenberg. We're scheduling a Codex walkthrough and it already gave me schedules. It has access to my calendar so it can propose times. This is crazy. All the stuff that I would normally procrastinate on because I have to look at my calendar or I have to go look up a bunch of documents and propose a reply, Codex just does it. And I just scroll through my email and talk to it. And then it goes and does it. And what's cool is the way that I built this. It's just a list of feeds, all the information I care about. So this is for my inbox, but there's other stuff. So I have this company feed, which I keep in a separate Codex thread so I can deal with it separately.

0:19

SPEAKER_00

And what the company feed does is it gathers everything that's happened in the company. Remember, it's a 30 person company. There's meetings all the time. There's Slacks everywhere. It gathers all the stuff that's happened in meetings that have been transcribed and put in Notion. It gathers everything from Slack. And then it just gives me a bunch of cards of stuff that I might be into. And so basically each card has a status. So it's stuff I need to look at. It's stuff that Codex needs to do. It's stuff that Codex is currently working on or it's stuff that we've already dealt with. And all I have to do to get through my inbox or any other feed of information I care about is review everything in review and queue it for Codex. And Codex takes care of the rest.

0:23

SPEAKER_00

So I can be like, OK, we're talking about positioning for Every and how we want to do our plans. There's a debate going on internally for Quora. There's a debate going on about whether we use the word archive or mark as done. And I'm kind of into the mark as done category. Can you just tell Brandon and Kieran that I vote for mark as done? Yeah. And now it's just going to go do that. And these are meetings or Slacks that I'm not in. And what I can do is I can just say, OK, execute. And it's going to go take all the work that I just did. I can do the same thing for inbox. It's going to go take all the work that I just did. And it's going to send out the emails. It's going to do the research. It's going to do whatever I want. What's amazing about this is all these decisions get recorded. So it's all recorded on my file system. It knows for any given email, what did I say to do? Did I say this is interesting? Did I say archive it? Did I draft a reply? And it gets better over time. So I compound my learnings. And you can see that these are the prompts that I use. And Codex maintains all this. It just watches what I do. And then it just gets better over time. Every time I do this, I'm getting a prompt that's better at finding things I might be interested in, archiving or hiding things I don't want to see, and then figuring out, OK, what's the right next action for this thing? So it's just incredibly powerful.

0:27

SPEAKER_00

So how can you do this yourself? One, use Quora. Quora.computer. We'll have a flow like this that's just built out. So you can just do any of this inside of Codex. As a Codex native app, we're going to have that coming out in the next couple of months. So Quora.computer or subscribe to Every if you just want to be notified when all this stuff comes out as it progresses.

0:31

SPEAKER_00

But if you want to do this yourself, OK, let's say, how would I do this? Honestly, just go into Codex. And let's see. I'm going to just use Monologue here. OK, I want you to help me make an inbox sweep app. And here's how it should work. It should connect to my email. You can either use the Quora CLI or Gmail. Either one should work. For every email that's in my inbox, I want it to create a card that shows me what the email is about and then proposes a next action, whether that's a draft or any other thing that makes sense. And then it should allow me to just swipe through my email or scroll through my email and just get through my email just by talking to each card. And what I want it to do is I want it to track every single step in the process and it should be Codex native. So what that means is I want the app to be responsible for the state. So all the state should be on the file system. And I want the app responsible for rendering the view, but it should be built to be used inside of Codex in the in-app browser. Can you just build that end to end? I want you to set a goal with an unlimited budget and make sure the goal is really detailed and you have a set of validation steps where you step through it yourself to make sure that it works. So I'm just going to say something like that, which is pretty basic. And there may be some things in there that you don't understand, but that's totally fine. We'll give you this prompt in the show notes so you can just paste that in there yourself. And what I'm going to do is I'm just going to press go and Codex is on auto review mode. I'm on 5.5 with extra high. That's probably what you want to do for this kind of experiment. But what's really great about Codex is it's going to set itself a goal. And the goal, you can see this, it's like build an end-to-end Codex native inbox feed app in my folder. The app should connect to email through the best available local path, preferring the Quora CLI. And I can just leave it. I'll come back and it's going to be something very similar to this app. That's what's so powerful about it. It's taking something that used to require a lot of time and a lot of engineering and making it available to you regardless of your level of technical skill. So that's the video. It's a quick one for you today.

0:36

SPEAKER_00

If you want more, you should check out our compound engineering workflow. We'll put it up here. It's going to tell you everything you need to know about using compound engineering, our engineering philosophy to build stuff like this, whether you're technical or non-technical. And you should subscribe to Every. Every is the only subscription you need to stay at the edge of AI. You can think of us as a frontier AI lab for the future of work. All we do all the time is explore how to take these models and these products and use them to do better creative work, whether that's writing or coding or design. And we put that into our videos, into our newsletter that comes out every day, into our products that we make like Quora, Monologue. You've seen some of them today. So if you love this video, you should definitely subscribe. Every.to/subscribe.

0:42

SPEAKER_00

Any work that you would do in a browser, you can do it in Codex over here. And then you have an agent that you can bring to all your work. So you're going back and forth and collaborating with the super smart agent that has access to your computer, access to all your apps, access to everything it might need to get work done. And it's sitting with you on any app that you want to use. And in this case, this is actually an app I built. So the other cool thing about Codex is it can just make stuff. Like I just vibe coded this app that lets me do my email with Codex. So let me explain how it works.

1:09

SPEAKER_00

So basically a few times a day, Codex will sweep through all the emails in my inbox. And for each email, it will give me a little card. The card looks like this. It tells me a little bit about the email. It lets me see the email if I want. And Codex also drafts a reply. And what I can do is I can just say, hey, can we actually maybe I'd like to meet with Doug near my office. Can we propose somewhere in Brooklyn and see if he's okay with that? And now I'm on to the next email. So new card, new scheduling thing. And now we can just keep going. Here's someone who's asking for sponsorships,

1:39

SPEAKER_00

who wants to sponsor every automatically draft a reply to the right person. I can say, this looks good. Now we're on to the next one. Now someone else who I want to make an intro for, I can just say, I already made the intro. You can archive this one. Next card. And here it's Greg Eisenberg. We're we're scheduling a codex walkthrough and it already gave me schedules that it has access to my calendar so it can propose times. This is it's crazy. Like all the stuff that I would normally procrastinate on because I have to look at my calendar or I have to go look up a bunch of documents and propose a

2:07

SPEAKER_00

reply. Codex just does it. And I just scroll through my email and and talk to it. And then it goes and does it. And what's cool is the way that I built this. It's just a list of feeds, all the information I care about. So this is for my inbox, but there's other stuff. So I have this company feed, which I keep in a separate codex thread so I can deal with it separately. And what the company feed does is it gathers everything that's happened in the company. Remember, it's a 30 person company. There's meetings all the time. There's slacks everywhere. It gathers all the stuff that's happened in meetings that have been transcribed and put in Notion.

2:38

SPEAKER_00

It gathers everything from slack. And then it just gives me a bunch of cards of stuff that I might be into. And so basically each card has a status. So it's stuff I need to look at. It's stuff that codex needs to do. It's stuff that codex is currently working on or it's stuff that we've already dealt with. And all I have to do to get through my inbox or any other feed of information I care about is review everything into review and cue it for codex. And codex takes care of the rest. So I can be like, OK, we're talking about positioning for every and how we want to do our plans. There's a debate going on internally for Quora. There's a debate going on about whether we use

3:10

SPEAKER_00

the word archive or mark as done. And I'm kind of into the mark as done category. Can you just tell Brandon and Kieran that I vote for mark as done? Yeah. And now it's just going to go do that. And these are meetings or slacks that I'm not in. And what I can do is I can just say, OK, execute. And it's going to go take all the work that I just did. I can do the same thing for inbox. It's going to go take all the work that I just did. And it's going to send out the emails. It's going to do the research. It's going to do whatever I want. What's amazing about this is all these decisions get recorded. So it's all recorded on my file system. It knows for any given

3:45

SPEAKER_00

email, what did I say to do? Did I say this is interesting? Did I say archive it? Did I draft a reply? And it gets better over time. So I compound my learnings. And you can see that these are the prompts that I use. And Codex maintains all this. Basically, it just watches what I do. And then it just gets better over time. Every time I do this at getting a prompt that's better at finding things I might be interested in, archiving or hiding things I don't want to see, and then figuring out, OK, what's the right next action for this thing? So it's just incredibly powerful. So how can you do this yourself? One, use Quora. Quora.computer. We'll have a flow like this

4:21

SPEAKER_00

that's just built out. So you can just do any of this inside of Codex. As a Codex native app, we're going to have that coming out in the next couple of months. So Quora.computer or subscribe to every if you just want to be notified when all this stuff comes out as it progresses. But if you want to do this yourself, OK, let's say, how would I do this? Honestly, just go into Codex. And let's see. I'm going to just use monologue here. OK, I want you to help me make a inbox sweep app. And here's how it should work. It should connect to my email. You can either use the Quora CLI or Gmail. Either one should work. For every email that's in my inbox, I want it to

4:53

SPEAKER_00

create a card that shows me what the email is about and then proposes a next action, whether that's a draft or any other thing that makes sense. And then it should allow me to just like swipe through my email or scroll through my email and just get through my email just by talking to each card. And what I want it to do is I want it to track every single step in the process and it should be Codex native. So what that means is I want the app to be responsible for the state. So all the state should be on the file system. And I want the app responsible for rendering the view, but it should be built to be used inside of Codex in the in-app browser. Can you just build

5:31

SPEAKER_00

that end to end? I want you to set a goal with an unlimited budget and make sure the goal is really detailed and you have a set of validation steps where you step through it yourself to make sure that it works. So I'm just going to say something like that, which is pretty basic. And there may be some things in there that you don't understand, but that's totally fine. We'll give you this prompt in the show notes so you can just paste that in there yourself. And what I'm going to do is I'm just going to press go and Codex is on auto review mode. I'm on 5.5 with extra high. That's probably what you want

6:00

SPEAKER_00

to do for this kind of a experiment. But what's really great about Codex is it's going to set itself a goal. And the goal, you can see this, it's like build an end-to-end Codex native inbox feed app in, in my folder. The app should connect to email through the best available local path, preferring the core of CLI. And I can just leave it. I'll come back and it's going to be something very similar to this app. That's what's so powerful about it. It's taking something that used to require a lot of time and a lot of engineering and making it available to you regardless of your level of technical skill. So that's the video. It's a quick one for you today.

6:34

SPEAKER_00

If you want more, you should check out our compound engineering workflow. We'll put it up here. It's going to tell you everything you need to know about using compound engineering, our engineering philosophy to build stuff like this, whether you're technical or non-technical. And you should subscribe to Every. Every is the only subscription you need to stay at the edge of AI. You can think of us as like a frontier AI lab for the future of work. Like all we do all the time is explore how to take these models and these products and use them to do better creative work, whether that's writing or coding or design. And we put that into our videos, into our newsletter

7:04

SPEAKER_00

that comes out every day, into our products that we make like Quora, Monologue. You've seen some of them today. So if you love this video, you should definitely subscribe. Every.to slash subscribe.

Reading tools

Type to find a passage

Appearance
Ask this transcript

Notebook

Export Markdown

Add a note