Summaries > AI > Claude > Claude Engineering Lead: "Stop Using Claude Like Itâ€™s 2024"...

Claude Engineering Lead: "Stop Using Claude Like Itâ€™S 2024"

https://www.youtube.com/watch?v=D-SVFneJ_Nk

TLDR AI is making software far cheaper and faster to build, shifting the real challenge from coding to choosing valuable problems, understanding users, and judging results. Felix Reeseberg says people should give capable models access to useful data and tools, guide them through expert workflows, and use independent evidence or reviewers rather than trusting an agent to assess its own work. The biggest opportunities lie in augmenting people, improving neglected software, and building products people genuinely value; staying adaptable matters as AI capabilities keep changing.

Key Insights

Start With a Real User Need

Before building with AI, learn how people work and what they value. The cost and speed of creating software have fallen, but successful products still depend on empathy and solving a meaningful problem. Look for tasks people struggle with, outdated systems they are forced to use, or goals they cannot currently achieve. Build around the value users need—not around a model’s capabilities alone.

Use AI to Explore Ideas Quickly

Treat AI as a rapid prototyping partner: describe an idea, ask it to create variations, and critique the results. Faster prototypes make it easier to explore possibilities, but they also make choosing what deserves attention more important. Favor products whose value becomes clear quickly and avoid adding features without a strong reason. Keep human judgment focused on direction, usefulness, and the quality of the experience.

Augment People by Rethinking Workflows

AI can automate unwanted tasks, but it can also help people do work they previously lacked the skills or tools to perform. For stronger results, redesign the workflow instead of merely inserting AI into an existing step. Break complex work into stages that reflect how capable practitioners approach it, and give the model appropriate tools, such as professional creative software. This makes quality more concrete than asking an AI to produce something “good.”

Give Models Context, Evidence, and Useful Tools

When practical, let an AI inspect relevant source material and use the systems needed to complete the task. Broad context, clear documentation, and well-chosen tools can be more effective than splitting information among many narrowly specialized agents. Start with a simple setup, such as a desktop app with browser access, and add connectors or reusable instructions only when a proven workflow needs them. Revisit those instructions regularly, since guidance written for older models can become counterproductive.

Evaluate the Process, Not Just the Answer

For knowledge work, a plausible final answer is not enough to establish quality. Check whether the AI used reliable sources, appropriate tools, and a sound method, and use hard evidence or performance measures when available. For tasks you can judge directly, such as whether a website looks appealing, apply your own judgment. For work that is harder to verify, seek independent evidence rather than relying on confidence or a lucky guess.

Use Independent Review to Catch Mistakes

AI systems can make confident errors, so ask separate agents or sessions to challenge assumptions, find flaws, and verify claims. A reviewer can sometimes spot problems more easily than the original model can create a strong first draft. A judge agent may help assess conflicting feedback, but verify important criticism against primary sources or other evidence. Avoid trusting a model to evaluate its own work in the same session, where self-assessment can be unreliable.

Build Feedback Loops and Reduce Oversight Bottlenecks

Compare AI feedback with your own review on real documents, transcripts, or other work, then study where the judgments differ. Collect feedback from places where it naturally appears, such as Slack or GitHub, and turn recurring lessons into practical review principles. Share collaborative workflows with teammates so they can assess results together. Use the findings to improve the process while keeping human oversight focused on decisions where mistakes matter most.

Keep Learning as AI Capabilities Change

AI tools and their capabilities can change quickly, so avoid assuming that today’s best workflow will remain best for long. Try new models and approaches, and test them on practical projects rather than relying only on predictions. For engineers and founders, combine technical skills with customer understanding, collaboration, and the ability to deliver useful products. A portfolio of real work can demonstrate that judgment more clearly than memorized APIs or credentials alone.

Questions & Answers

How has AI changed Felix Reeseberg’s role in product development?

He says he no longer writes code and now acts as “a conveyor belt for Claude,” focusing on prototyping, guiding agents, and evaluating their output. AI makes it much faster to build prototypes, so the harder question is deciding what deserves people’s attention.

What makes a good AI product, in Reeseberg’s view?

A good product helps people do something they couldn’t do before, delivers value quickly and clearly, and often succeeds by doing fewer things well. Building it still requires empathy for users and an understanding of what they value.

What is the difference between automation and augmentation?

Automation removes tasks, often saving time or money. Augmentation helps people do work they previously lacked the skills, knowledge, or techniques to perform, which Reeseberg finds especially exciting.

How should AI work be evaluated, especially for writing and other creative tasks?

Evaluate the process as well as the result: consider whether the model used sound methods, reliable sources, and appropriate tools, rather than rewarding a lucky guess. For creative work, break the task into steps that reflect how skilled practitioners work and use professional tools when they matter.

How can people reduce the risk of AI making confident mistakes?

Use independent reviewers or separate agents to challenge assumptions, find errors, and try to disprove the work. A judge agent can assess the criticism and decide whether it is valid. For difficult-to-verify work, rely on more review and hard evidence rather than trusting the original agent to assess itself.

Should people use specialized agents or one general-purpose model?

Reeseberg cautions against assuming specialized agents are always better. A capable general-purpose model can work effectively with broad context, clear documentation, and the right tools; forcing it into narrow roles may limit it. Existing workflows that work well do not necessarily need to be changed.

What is a practical way to get more value from Claude and other AI agents?

Give the model access to relevant raw data, evidence, and tools so it can inspect information and act directly instead of relying on a person to pass details back and forth. Beginners can start with the desktop app’s built-in browser, then add connectors when a useful workflow is established.

How can people improve an AI agent’s judgment over time?

Collect feedback from places where it naturally appears, such as Slack, GitHub, or Telegram; summarize it into principles; and use those principles in future reviews. Comparing an agent’s feedback with a person’s can also reveal differences and help improve the workflow.

What does Reeseberg say makes a promising software business today?

Software is cheaper and faster to build, but successful products still depend on understanding users and delivering value. Opportunities include replacing outdated, clunky business systems and applying powerful AI to meaningful problems that people have not yet imagined solving differently.

What skills should software engineers and job candidates emphasize?

Understanding customers, building useful products, exercising judgment about what is necessary, and combining systems matter alongside technical ability. A portfolio of real projects can demonstrate these qualities more clearly than a degree or school name alone.

What advice does Reeseberg offer for navigating rapid AI progress?

Stay open to new models, tools, and possibilities. AI is advancing faster than many people expect, so products and practical advice may change quickly; trying new things helps people adapt and recognize what has become possible.

Summary of Timestamps

Felix Reeseberg, who helped build Claude Cowork in about a week and a half and leads engineering for Claude at Anthropic, says he no longer writes code. Instead, he acts as a “conveyor belt for Claude,” prototyping ideas, guiding agents, and judging their work. His experience illustrates how AI is shifting some engineering work from writing code to directing and evaluating it.

Because AI makes prototypes faster and cheaper to create, Reeseberg says the key challenge is increasingly deciding what deserves attention. He favors products that help people do things they could not do before, and says strong products often have fewer features and make their value clear quickly.

Reeseberg distinguishes between automation, which removes unwanted or costly tasks, and augmentation, which helps people do work they previously lacked the skills or knowledge to perform. He finds augmentation especially exciting and says the greatest gains often come from rethinking workflows and learning how to assess the results.

A product manager named Pavo reportedly asked Claude to examine a large collection of NASA data and search for a new planet; Reeseberg says NASA was working to confirm the possible discovery. The example shows how users can apply AI to ambitious projects that even its builders may not have anticipated.

A major limitation is that an agent may not reliably judge its own work in the same session. For knowledge work and creative tasks, quality is harder to measure than whether code runs, so a model should be evaluated on its process, sources, and use of tools—not merely whether it happened to produce the right answer.

For writing, a broad request such as “make this good” is less helpful than breaking the task into steps that reflect how skilled professionals work. The speakers emphasize that strong writing and art often come from refined, deliberate processes, and that using professional tools—such as Blender for animation and 3D work—can help more than asking a model to understand quality in the abstract.

One way to catch mistakes is adversarial review: ask separate agents or threads to find flaws, challenge assumptions, and try to disprove the work. Since finding errors can be easier than creating something from scratch, reviewers may still be useful even if they are less capable than the original model.

The speakers discuss using a separate “judge” agent to assess criticism from other agents, check whether feedback is valid, and decide whether to send issues back to the agent that created the work. People can often judge visible results themselves, such as a website’s appearance, but harder-to-verify work may benefit from additional review.

One speaker trusts a single Claude thread to fact-check slide decks by finding and verifying primary sources. They caution against assuming every task needs a specialized agent: a capable general-purpose model may work better when given broad context, useful tools, and clear documentation.

Reusable model instructions, sometimes called “skills,” can become outdated as models improve. The speakers suggest testing whether a skill still helps, since removing old guidance and prompting the model directly may produce better results.

AI is most useful when it can inspect raw data and use tools directly rather than relying on a person to pass information back and forth. Setting up access takes effort, but beginners can start with Claude’s desktop app and built-in browser, then add connectors once a workflow proves valuable.

The speaker describes agents as task-focused projects rather than role-playing personas. Instead of telling Claude to pretend to be a professional, ask what tools and techniques experts use and have it apply those methods.

To improve an agent’s judgment, the speaker recommends collecting feedback from places where it naturally appears, such as Slack, GitHub, or Telegram. Summarizing that feedback into principles can help guide future reviews, although the Telegram process described is still manual.

The speaker says AI has increased productivity, but the team still reviews every output because they do not fully trust any agent. One proposed way to improve this process is to compare Claude’s feedback with a person’s review of the same transcripts or documents, then refine the workflow. Claude in Slack is also suggested for collaborative, team-facing use.

Although software is now much cheaper and faster to build, the fundamentals of a good product have not changed: understand users and deliver value they care about. AI may help very small teams build ambitious companies, but the software most at risk is software people tolerate only because they lack a better option.

The speakers see opportunity in outdated business software used in places such as airports, transportation authorities, and hotels. Visiting workplaces and observing how staff actually use their systems can reveal unmet needs; building the software may be easier than gaining access and building relationships.

The industry has many cheap, low-intelligence models, but the larger opportunity is applying top-tier AI to meaningful problems. The speakers describe a “capability overhang”: people often use powerful models for simple tasks instead of rethinking what could be possible. Closing that gap requires creativity, ambition, and empathy for how people and businesses work.

For software engineers, understanding customers and building products people value may matter more than memorizing APIs or writing elegant code without user impact. A portfolio of real projects can demonstrate technical ability, judgment, customer empathy, and the ability to combine systems. Hiring, the speaker argues, should place greater weight on what candidates have built and contributed than on degrees or schools.

Reeseberg says University of Waterloo engineers often become effective quickly because internships teach practical skills beyond coding, including collaboration, advocating for ideas, and getting products built. The speakers also encourage children to explore widely and believe ambitious goals are possible, since assumptions about what can be done may limit people more than technology does.

Looking ahead, Reeseberg warns that AI is advancing so quickly that predictions and advice can become outdated within weeks. He expects AI products to change dramatically, comparing today’s moment to phones before the iPhone, and recommends staying open to new models and experimenting with new tools.