When CJ ran an MVP development service, every new project began with 9 to 10 hours of conversation with ChatGPT, drawing out the product requirements, technical architecture, and task list bit by bit. He later used make.com to chain the prompts into a sequence, saving seven of those 9 to 10 hours. He called that step “magical” in an interview (Starter Story interview[1], 2025-05-21). CodeGuide began with this workflow he used himself: first write the project’s instructions, then have AI coding tools such as Cursor, Claude Code, or Lovable follow them.
What it sells sounds unglamorous: documents. Project plans, PRDs/product requirements, tech stacks, wireframes, and task lists are packaged into a “project knowledge base.” The website says users describe an idea in natural language, answer a few clarifying questions, and receive these documents for AI coding tools to reference (CodeGuide website[2], read on 2026-11-18). Its original role was not to write code for people, but to clarify the project first.
That also explains why CodeGuide looks like a project instruction manual rather than another coding tool. First it breaks down the features, tech stack, architecture, database, pages, and order of tasks; then an AI coding tool writes accordingly. The website’s FAQ puts the problem plainly: AI models can write code but cannot read your mind. A one-line prompt is not enough. AI guesses, producing hallucinations, odd assumptions, and missing features. It is a simple explanation, but it identifies a concrete step: what AI coding tools lack is not generation speed but the project conventions left out of the prompt.
The step it takes over
Getting started is ordinary: describe your idea in natural language, answer a few clarifying questions, and receive the documents. Then a single command attaches the specifications to an IDE or CLI, while a Chrome extension syncs the project blueprint into web-based AI builders. Existing GitHub repositories can also be imported, analyzed, and used to generate specifications for new features that fit the existing architecture (CodeGuide website[2], read on 2026-11-18).
Image: Screenshot from the product website[2] on 2026-11-18, retaining the official interface; not the result of an independent hands-on test.
People have not been replaced at the documentation layer: they must answer the clarifying questions and judge whether the specifications are right. The website has now expanded the product into Software v2: an agent that can analyze repositories, plan documentation, and code autonomously, with a unified interface connecting CLI tools and 200+ models, plus parallel execution of multiple agents. These are capabilities listed by the website. Document quality, the extent of hallucination reduction, actual compatibility with different IDEs, and Sv2’s autonomous coding results have not been independently verified.
A landing page before a product
CJ did not build a complete product first. In the 2025-05-21 interview, he traced the beginning to a tweet about an “ideal Cursor workflow”: 400 reposts and 300 signups. He built a landing page with Bolt and collected 1,800 waitlist email addresses in two weeks; elsewhere in the interview, the figure is “over 1,800” (Starter Story interview[1]).
His second step merits a pause. He sent individual direct messages to beta users who had commented, asking the same question: did they already have a better workflow for solving AI hallucinations? He concluded that nobody had solved it better than the approach he had tried. This was the founder’s own survey, sampled from his own comments section, not a third-party evaluation. To launch quickly when he could not do everything himself, he brought in a friend to handle the backend and customer support while he took the frontend and marketing. The division of work was intended to speed up the launch.
That validation showed that people would leave their email addresses. It did not show they would pay. Revenue figures are a separate matter.
Pricing, and an account with only one page
In May 2025, he offered no trial because the top-tier AI models he used were not cheap. The monthly fee was then $29. He planned to raise it to $39–$49 over the following six months and offered a 40% annual-payment discount (Starter Story interview[1]). That was the position at the time.
The website now has three tiers. Free provides a one-time allocation of 120 credits and completion of 1 project, with no credit card required. Billed annually, Pro costs $288 (shown as $24/month), with 5,000 credits and 15 projects a month. Founder costs $348 (shown as $29/month), and its card lists 6,000 credits and 18 projects a month. Switching to monthly billing gives Pro at $29/month with 5,000 credits/15 projects and Founder at $39/month with 6,000/18. The “pay $29” line at the bottom of the page refers to monthly Pro, not Founder’s allowance (CodeGuide pricing[2], read on 2026-11-18).
On revenue, CJ reported reaching $42,000 MRR within three months of the December 24 launch, a community of 4,800+ AI developers, and an expectation of reaching about $90,000 that month. This is an unaudited founder account; a forecast is not confirmed revenue. In the same interview, he put monthly tool/API costs at about $3,500, with OpenAI’s API the largest item at $2,800. Both revenue and costs come from a single source. Profit remains unclear.
The acquisition evidence is similarly thin. At the time, he did no paid marketing; all traffic came from X, where he posted tutorials daily to drive people to the waitlist. A channel is identifiable, but without waitlist-to-paid conversion, customer acquisition cost, retention, and refund data, it cannot substantiate the revenue on his behalf.
Nor is this the only offering. ChatPRD’s website, read on 2026-11-18, says it turns ideas into specifications a team can deliver, scores documents, generates clickable prototypes from specifications, and opens specifications in v0/Lovable/Bolt/Cursor. Hosted MCP/REST lets Claude Code, Codex, Cursor, and ChatGPT read specifications and ask ChatPRD to write or revise documents (ChatPRD website[3]). Pro costs $179 billed annually (a stated $15/month equivalent, with a credit card required for the 14-day trial); Teams costs $349 per seat annually (a stated $29/seat/month equivalent) (ChatPRD pricing[4]). The two overlap in generating specifications before handing them to coding tools. But there is no firsthand comparison of task fit or quality differences, so these descriptions cannot establish which is better for a particular task.
Back at the documentation layer, CodeGuide has neither made decisions for people nor replaced their coding tools. Its original contribution was to put a sheet of context between a person and AI: how the project should develop, where the data belongs, and the order in which to do the work. Whether that sheet is right remains a human judgment.
The business rests on a simple bet: the cheaper AI makes coding, the more clarifying a project looks like a job in its own right. Whether that holds depends not on how many pages of documentation it generates, but on whether people are still willing to maintain that documentation as the project grows more complex.