00:57 - 01:29

The suggested method is to interview first, build a spec second, and code last.

12:23 - 13:24

Creating an in-depth plan before editing code allows for faster implementation and better results.

15:22 - 16:22

Using Claude to create a plan before writing code helps verify the intended actions.

18:03 - 20:06

Planning mode allows the user to align on a strategy before the model executes changes.

16:54 - 17:54

Product development is increasingly about the idea rather than the process of making it.

00:59 - 01:31

Starting in plan mode and iterating multiple times is a critical step for quality.

18:25 - 20:25

Invest in a good plan file before building to get much better results.

04:12 - 06:12

The broad strokes of the workflow include researching what exists before creating a plan.

14:40 - 15:40

The creator of Claude Code runs five instances in parallel for maximum efficiency.

04:15 - 06:16

Developers can run multiple agent threads simultaneously to scale their engineering work.

05:34 - 07:35

Advanced users run five agents in parallel across separate terminals to accomplish more.

00:55 - 01:55

Using git work trees allows developers to run different instances of Claude on one repo.

00:19 - 02:19

Engineers often work across multiple tabs, letting Claude think while they move tasks.

00:24 - 00:54

The expert workflow involves keeping five Claude Code terminals open in parallel.

17:12 - 18:13

Customers and employees often run four clouds at once to orchestrate complex projects.

00:12 - 02:12

Using like a lot of credits [music] like.

00:25 - 01:29

The goal was to build a tool that lives universally in the terminal.

01:10 - 02:10

The terminal is identified as the most universal and simple interface for development.

08:13 - 09:13

The product is intentionally simple to fit into any workload or terminal environment.

00:02 - 02:02

Quad code has access to everything that.

02:17 - 03:17

The tool works over remote SSH or Tmux, regardless of the environment.

05:44 - 06:44

The terminal tool is favored by power users for its speed and color rendering.

04:48 - 06:49

The CLI is described as the place where the most intelligent work happens.

01:47 - 02:05

Using Claude exclusively in the terminal allows the computer to run much faster.

18:40 - 19:41

The CLAUDE.md file provides additional context about the project code base.

03:48 - 04:48

The CLAUDE.md file acts as the agent's memory for team and project preferences.

11:07 - 12:07

Instructions in the CLAUDE.md file are automatically plopped into the agent's context.

19:39 - 20:39

Checking a CLAUDE.md file into the root helps share memory across the team.

12:33 - 13:35

The model performs better when given project-specific context via CLAUDE.md files.

03:03 - 03:34

The whole team contributes to project-specific rules in the CLAUDE.md file.

17:25 - 18:25

CLAUDE.md files provide critical knowledge to agents that have never seen your code.

00:23 - 02:23

Claude's memory includes project structure and code style conventions for the team.

10:50 - 11:50

Giving an agent ways to verify its work can drastically increase the quality of changes.

00:41 - 01:11

Boris recommends always giving the model a way to verify its own work.

16:22 - 17:24

Automated tests allow Claude to iterate on changes until they pass reliably.

11:27 - 13:28

Verification tools like Playwright are used to confirm that UI changes work correctly.

11:38 - 12:38

The trick is to give the agent feedback tools so it can iterate to a better result.

20:01 - 22:02

Automated verification ensures the feature actually works before the user checks it.

00:24 - 01:24

Agents building your business, building.

00:39 - 01:40

- A lot has happened in the past 12 months,.

08:04 - 09:04

Power users deploy specialized sub-agents for tasks like code simplification and verification.

03:49 - 05:50

Sub-agents allow developers to create instances of Claude focused on a single task.

04:54 - 05:25

Sub-agents automate common workflows for most PRs without bloating the main context.

01:01 - 03:01

In the future, roles like frontend and backend engineers will be handled by agents.

12:33 - 13:33

Agents are described as another way to extend functionality through forked context windows.

14:10 - 16:10

Sub-agents perform specific work for the main orchestrator, such as research or testing.

01:56 - 03:57

People use agents for specific UI designs or complex data analysis pipelines.

43:31 - 45:32

Multiple agents can be used to battle-test ideas and find the most robust solution.

22:34 - 24:34

Non-engineers are increasingly using agentic tools to automate their weekly tasks.

00:00 - 02:00

Claude is used to rename files and automate receipts, helping non-technical teams.

08:40 - 09:40

Technical skills are no longer required for people in finance and HR to build agents.

16:07 - 18:07

Product managers use agentic tools to build functional prototypes incredibly fast.

06:30 - 08:30

People without technical backgrounds are learning to use Claude for complex data analysis.

27:58 - 29:59

Agents make technical tasks accessible to everyone, regardless of their background or schedule.

17:22 - 18:22

Anyone with an idea for a product can now build it themselves using AI.

48:11 - 50:11

The shift to non-technical users adopting these tools is considered a major milestone.

10:28 - 12:28

Managing usage limits is now as important as engineering the actual context.

03:32 - 05:34

Code is a really interesting product and.

00:29 - 01:33

Best practices we've figured out both internally and from our users uh for getting the most out of.

33:56 - 35:58

Smarter models use fewer tokens in the long run because they require less steering.

05:53 - 07:54

Better initial information allows the model to work more effectively and use fewer tokens.

19:53 - 21:53

There is a trade-off between model power and the cost of token constraints.

00:45 - 01:16

Usage engineering ensures users get the most value without wasting tokens.

13:56 - 14:58

A whole feature can be built for as little as three dollars in API costs.

09:04 - 10:06

The team uses a radical internal dogfooding cycle to improve their coding agents.

10:38 - 11:38

Claude Code is essentially a product of the team's own daily coding work.

02:18 - 03:21

A leaderboard was built to track how much Anthropic employees use the internal tools.

34:18 - 36:18

Engineers at Anthropic form their own opinions through intensive daily dogfooding.

00:22 - 01:24

- Yeah, Claude Code, it's a way to do agentic coding.

16:41 - 18:42

Internal teams at Anthropic are the primary testers and developers of the tool.

00:33 - 01:33

Less code is being written by humans.

12:05 - 14:05

Internal teams use their own models and tools to fix bugs in a few hours.

07:52 - 08:52

The product lacks a flashy UI to stay out of the model's way.

04:00 - 05:00

Interacting with models requires a harness, similar to a horse requiring a saddle.

03:58 - 04:58

Anthropic aims for the simple thing that works, prioritizing pure agent behavior.

57:44 - 59:44

The goal is to discover a new design language that maximizes agent capabilities.

14:42 - 16:43

User experience priorities focus on tools being extremely snappy to use.

22:43 - 24:46

Engineers prefer extension points and settings that let them customize their own tools.

01:10 - 02:10

Terminal-based tools are effective because they are the simplest and most universal.

08:06 - 09:06

The agent independently figures out how to chain tools together for complex work.

05:04 - 06:06

Claude explores code bases through agentic search rather than static indexing.

01:51 - 03:52

The agent searches and refines terms to understand a code base like a human would.

06:18 - 07:18

The model can autonomously look through Git history to answer deep architectural questions.

07:01 - 08:02

Agentic tools explore code bases and read files to provide a deep understanding.

00:09 - 02:09

Utilizing agents for research takes the engineering process a step further.

48:18 - 50:19

Agentic search tools have different internal 'vibes' for accuracy and speed.

04:05 - 05:05

Flash initial research is conducted by the agent combing through documents.

03:21 - 04:23

Developers are advised to use agents to ask questions and research code first.

00:52 - 01:52

The programming profession is being dramatically refactored into sparse human contributions.

00:18 - 00:49

Senior engineers are now shipping production code they have never even read.

01:56 - 02:27

AI tools amplify what the expert developer brings to the problem domain.

14:02 - 15:02

Future engineering roles will focus on higher-level system requirements and design.

12:39 - 13:41

A mental shift must occur for those used to handwriting code for decades.

01:51 - 02:52

Developers now orchestrate Claude to proactively make and review changes.

17:24 - 18:24

Running multiple clouds requires a shift to orchestrating rather than direct coding.

01:20:48 - 01:22:50

Modern engineering is about orchestrating tools rather than traditional manual programming.