Anthropic releases new AI usage guide! Four stages + three tools to teach you how to build an AI-native company with Claude

Anthropic releases the official manual. The manual analyzes how AI can rewrite the entrepreneurial process of startups in four stages. At the same time, the tools help small teams improve work efficiency and remind to prevent agentic technical debt.

The biggest help AI provides to entrepreneurs is not just "helping you write code." A larger change is: many tasks that previously required teams, consultants, or outsourcing can now be first handled by AI for the initial round.

For example, you can ask AI to help organize market data, compare competitors, write interview questions, create simple prototypes, check code for security risks, or even organize customer feedback. This allows early-stage entrepreneurs not to rush to hire a full team from the start, and to verify more quickly: does this problem really exist? Is there someone willing to use this product? What is the next gap to fill?

Anthropic's official manual, 《The Founder's Playbook: Building an AI-Native Startup》, released on May 14, answers this question: if a company treats AI as a foundational tool from day one, how will the entrepreneurial process change?

This 36-page manual breaks down the startup journey from idea to growth into four stages: Idea, MVP, Launch, Scale. Each stage explains the goals, common failure reasons, which Claude tools to use, and practical exercises to follow.

In simple terms, this is an "AI era startup process chart."

Founder roles, shifting from doing everything to being a command system

In the past, entrepreneurship heavily depended on division of labor. Technical founders handled coding, business founders handled presentations, fundraising, and customer acquisition; if lacking certain skills, they would find co-founders, outsource, or hire staff.

But Anthropic believes AI is rewriting this division. Founders are no longer just executors of single tasks but more like coordinators: assigning research, writing, development, testing, customer service, and operational processes to different AI tools for assistance.

The manual even describes this change as "a 10-person team can still compete to become a unicorn." This doesn't mean 10 people will necessarily create a company worth $1 billion, but AI enables small teams to handle workloads that previously required large teams.

The manual also mentions several cases. The Anything platform uses Claude to help 1.5 million users turn ideas into operational software products; Carta Healthcare uses Claude to process 22k surgical cases annually, reducing data extraction time by 66%.

Although these figures are self-reported by Anthropic, they convey the direction: AI is not only for chatting but is also entering core product development and operations.

Image source: 《Digital Age》

Three Claude tools, each suited for different tasks

The manual clearly categorizes three ways to use Claude.

• Chat: suitable for short tasks, such as asking questions, summarizing a sentence, or quickly checking an idea.
• Claude Cowork: suitable for longer knowledge work, such as organizing a batch of interviews, weekly KPI summaries, or connecting folders and tools.
• Claude Code: suitable for engineering tasks, such as reading code, fixing bugs, building prototypes, running tests, or security checks.

In other words, not all tasks should be dumped into the same chat window. Anthropic recommends: assess the task's nature first, then choose the tool. For quick discussions, use Chat; for processing large amounts of data, use Cowork; for coding, use Claude Code.

Stage 1: Idea — Confirm if the problem truly exists

The Idea stage is the concept phase. At this point, the most important thing is not to immediately build a product but to verify whether the problem you want to solve actually exists. The manual calls this problem-solution fit, meaning "whether the problem and solution align."

Anthropic warns that AI coding too quickly can introduce new risks: founders might rush to build a product before confirming the demand. The manual cites industry statistics indicating that 42% of startups fail because they create products no one needs; with the proliferation of agentic coding tools, this mistake could happen even faster.

In this stage, AI is best used for three things:

1. Help turn vague ideas into testable hypotheses. For example, instead of saying "contract review is too slow," specify "mid-sized legal teams take over 3 days to review each contract because red flags are scattered across emails." This clarifies who to ask and what to verify.
2. Help estimate market size. The manual mentions TAM, SAM, and SOM, which can be simply understood as: how big is the entire market, how much of it can you serve, and how much is most likely to be captured in the short term.
3. Ask Claude to play the devil's advocate. You can request it to find reasons why your idea might fail or to write a critique of your competitors that could beat you, helping you avoid only seeing evidence that supports your view.

At this point, it's suitable to use Claude Code to create a lightweight prototype. The manual emphasizes that this prototype is not a formal product but a tool to let target users "touch and feel" and help verify your direction.

Stage 2: MVP — Build the minimal version, but avoid technical debt

MVP stands for Minimum Viable Product. It means creating a version with the fewest features necessary to verify demand, not building a complete product from the start.

In the AI era, MVP development can be much faster. Non-technical founders can also use Claude Code to produce a working prototype. But Anthropic emphasizes that speed isn't the only variable.

The biggest risk is agentic technical debt, meaning "agentic-style technical debt." Simply put, if you keep letting AI guess your architecture, rules, and trade-offs from scratch each time, the code will quickly become: each module works independently but lacks overall logical consistency.

Therefore, the manual recommends creating a CLAUDE.md file before coding begins. This document records the project architecture, naming conventions, dependencies, and practices to avoid. Each time you start a new Claude Code session, have it read this file to prevent starting from scratch every time.

At the same time, clearly define the product scope: what this version will do, what it won't, and conditions for adding features later. This may seem basic, but in the context of AI rapidly adding features, it becomes even more critical.

Before launch, the manual also suggests using Claude for a security review, such as login authentication, API data exposure, input injection risks, or known vulnerabilities in dependencies.

However, Anthropic reminds that AI security scans are only auxiliary and cannot replace formal security or compliance reviews. Claude Code Security is currently in limited beta, and not all users can access it.

Stage 3: Launch — Don’t let founders become bottlenecks

The Launch stage involves product release and engaging more users. The challenge shifts from "can we build it" to "can it operate stably."

Anthropic points out that many founders are hands-on during the MVP phase, which is advantageous, but during Launch, it can become a bottleneck. Customer support issues, feedback, product decisions, and sales data all bottlenecked on the founder make it hard for the company to grow.

At this stage, Claude Cowork can be used to review operational workflows: which tasks can be fully automated? Which require human input but not necessarily the founder? Which decisions still need the founder’s personal judgment?

Claude Code can be used to audit the architecture, prioritize technical debt left from the MVP phase, and prepare compliance checks like SOC 2, GDPR, HIPAA, or organize procurement documents and patch lists for enterprise customers.

Stage 4: Scale — The moat is not code but accumulated assets

Scale is the expansion phase, where the product moves from early users to larger markets. At this stage, Anthropic believes that the true moat of AI-native startups is not the code itself but what has been accumulated over time.

These include: understanding of specific industries, customer usage data, product workflows, and deep integration with other tools. Because code can be generated faster by AI, simply "having features" is no longer rare; what’s truly hard to copy is the long-term accumulated domain knowledge and customer workflows.

The manual suggests founders organize their industry knowledge, common pitfalls, customer terminology, and success stories into Claude Projects, Memory, or Skills, allowing Claude to gradually understand your company and market.

Additionally, you can use Claude Code to turn user behavior data into feedback loops for product improvement, creating a "data flywheel" narrative. This is important for investors or enterprise clients because it answers: why can't competitors with more resources copy you overnight?

Before reading this manual, be aware of three things

First, this manual is essentially Anthropic’s product document. Its methodology has reference value, but it assumes you will use Claude, Claude Cowork, and Claude Code. In practice, you can extract the concepts and adapt them with tools like Cursor, Replit, Devin, or others.

Second, some figures are self-reported by Anthropic or cited from industry data. For example, "42% of startups fail because they build products no one needs" is not Anthropic’s own research; the 1.5 million users of Anything and Carta Healthcare’s 66% reduction in data extraction time are also vendor case studies, not third-party verified results.

Third, some capabilities discussed in the manual still have entry barriers. For example, Claude Code Security is currently in closed testing, not available to all users.

Conclusion: AI will not decide the direction for founders

The most valuable aspect of this manual is not how many Claude features it lists but how it reorganizes the startup process.

As AI makes "creating things" faster, entrepreneurs should be most cautious about: not making mistakes too quickly. Verify the problem first, then build the MVP; clarify scope before accelerating development; break down operational workflows first, then let tools assist; finally, turn accumulated knowledge, data, and workflows into a moat.

In other words, AI will not decide the direction for founders, but it can amplify the consequences of wrong directions. Used well, it’s an amplifier for small teams; used poorly, it only accelerates and reinforces the creation of products no one needs.