How to Make Claude Code Your AI Engineering Team

Predictions and Expectations:

• Gary Tan expects that "the way to get agents to do real work is the same way humans have always done it — as a team, with roles, with process, with review."
• Gary Tan predicts that "the barrier to building just collapsed" and states we are in a "completely new era of building software — the agent era."
• Gary Tan believes that "by default, it uses Claude" and characterizes Opus 4.6 as "sort of ADHD CEO," while predicting Codex serves as the "autistic CTO" for when "the going gets tough."
• Gary Tan fears that "supply chain attacks" are "really scary in AI coding right now."
• Gary Tan anticipates that users who rely on Cloud Code "spend 80 to 90% of their time in office hours, plan CEO review and auto plan."
• Gary Tan believes that GStack "gets you to Level 7" of a "software factory," where one can "run multiple conductor windows on different projects...all at the same time."
• Gary Tan expects that with the current capabilities, users "are able to have an idea and then get farther along with it than you ever would be."
• Gary Tan predicts that he can "do 10, 15, 20, sometimes 50 PRs in any given day, depending on the number of meetings I have in that day."
• Gary Tan believes it is the "most incredible time in history to build software."
• Gary Tan predicts that GStack will function as a system where "it lands" automatically when "ready to land," removing the need for a traditional "to-do list."
• Gary Tan expects that the "Conductor" tool will allow him to evaluate and bring in "open source fixes that I'm getting from the community" in "waves."

Timelines and Milestones:

• Gary Tan expects to be "back in about five minutes" after initiating the "Design Shotgun" generation.
• Gary Tan notes that the "Design Shotgun" process "takes about 60 seconds."
• Gary Tan mentions he "started playing with Cloud Code back in January" to establish the timeline of his engagement with the technology.
• Gary Tan states he "built GStack to encode this three weeks ago."
• Gary Tan observes that "a year ago, two years ago, even like three months ago, it's not clear to me that anyone would even try this" regarding browser automation for tax docs.
• Gary Tan notes that building Posterous originally "took two years" with a team of 10 engineers.
• Gary Tan indicates he "studied computer systems engineering at Stanford" and was "employee number 10 at Palantir" during his first decade of career.

Technology and Product Direction:

• Gary Tan plans to use "Office Hours" as a skill "modeled exactly after what we go through at YC as a partner doing Office HOURS with startups" to ask "six forcing questions."
• Gary Tan expects GStack to "encode this three weeks ago" into a system where the model acts as "a team of specialists."
• Gary Tan predicts that "Cloud in Chrome MCP" is "one of worst pieces of software I've ever used" due to "context bloat" and actions taking "two to three seconds" even when working.
• Gary Tan plans to utilize "Codex" to "sort out all of the crazy bugs" that might exist in the code.
• Gary Tan expects the "slash QA and slash browse tool" to allow agents to "take screenshots," "do complex interactions," "run eventually full regression tests," and "assess real browser bug issues."
• Gary Tan predicts that GStack will include a "ship tool" to ensure "your PR is ready to land on main."
• Gary Tan plans to create "multiple open source projects with tens of thousands of stars" using the GStack workflow.
• Gary Tan expects users to be able to "log in, then AI navigates to text docs and downloads PDFs" without "stored credentials."
• Gary Tan predicts that GStack can "auto fill out" missing items like "failure handling, there's no privacy section, 2FA handoff" during the adversarial review.
• Gary Tan expects to have "one or two sessions active for each project" to evaluate community contributions.

Market and Industry Outlook:

• Gary Tan believes that the current era allows for a shift from "document aggregation alone" to a "matchmaking and lead gen for tax preparers" model.
• Gary Tan predicts that businesses could "charge $2 or $5 a month or $2 of $5 maybe an entire year for 1099 into aggregation, but you can actually charge a percentage of the transaction that ends up happening with a tax preparer, which might be 10x more."
• Gary Tan notes that "many banks will email you with new tax documents, but some won't."
• Gary Tan observes that "TurboTax and HRNR Block actually have 1099 import features, and Plaid connects to banks," but questions why "those" aren't solving the specific problem of finding "1099 ints" across all sources.

Company Plans:

• Gary Tan plans to release GStack as "open source" via "github.com slash garytan slash gstack."
• Gary Tan expects to "click on whenever I have an idea... I just click the plus icon in Conductor" to create a new "work tree."
• Gary Tan plans to run "10 to 15 parallel Claude Code sessions all at the same time."
• Gary Tan expects to utilize "28 different commands" within GStack, noting that "some of the people who use it" already employ this workflow.
• Gary Tan plans to run "Office Hours, CEO Review, End Review, Adversarial Review" on any new "work item" before it lands.
• Gary Tan expects to "build all my startup Posterus" quickly using GStack, contrasting it with the "10 million dollars, and 10 engineers" it previously took.
• Gary Tan plans to use "browser automation" to "search the inbox, find all of the 1099s that you need to download" and "send an email to the CPA."
• Gary Tan expects to create "multiple AI versions" of UI designs via "Design Shotgun" and "farms it out to OpenAI Codex."
• Gary Tan plans to use "Conductor" to "make a tax app" that "goes into your Gmail and fish out all of your 1099s."

Financial Guidance:

• No forward-looking financial guidance (specific revenue targets, market cap, or investment amounts) was provided.

Risks and Caveats:

• Gary Tan fears "supply chain attacks" in the AI coding space.
• Gary Tan notes that "without proper setup... the model wanders" and "guessing at that scale is how you get plausible looking code that silently breaks."
• Gary Tan observes that "one in three times" he uses Office Hours, he concludes "This isn't something that makes sense."
• Gary Tan warns that "if it can [auto-fix], it does," but implies reliance on this automation carries risk if the model fails to address complex constraints like "failure handling, there's no privacy section, 2FA handoff."
• Gary Tan notes that "Cloud in Chrome MCP" is a "worst piece of software" with "crazy context bloat" that "often wouldn't even do anything."
• Gary Tan warns that "if you just came in and gave instructions to the model... it will just literally do it" but won't "think about who's the user, what is the business model."

Confidence and Disagreement:

• Gary Tan expresses high confidence that "the way to get agents to do real work is the same way humans have always done it."
• Gary Tan expresses certainty that "the barrier to building just collapsed."
• Gary Tan states a strong opinion that "there's no failure handling" and "no privacy section" in the initial auto-generated review, which he expects the system to auto-fix.
• Gary Tan disagrees with the premise that "it's not happening in the cloud" for his browser automation approach, clarifying "the cloud is just someone else's computer."
• Gary Tan expresses skepticism about using "Codex" for immediate debugging but expects it to sort out "crazy bugs" later.
• Gary Tan confidently predicts he can "do 10, 15, 20, sometimes 50 PRs in any given day."
• Gary Tan hedges his startup idea validation by saying "I'm not saying this is actually a good startup idea" but expects the process to move it "farther along."