# wu-cherny-use-claude-code-builders-every-2025-10-29

## Veille

Cat Wu and Boris Cherny (Anthropic) explain how to use Claude Code like its creators: antfooding, plan mode, subagents, hooks, and extensibility — Every's AI & I podcast

## Titre Article

How to Use Claude Code Like the People Who Built It

## Date

2025-10-29

## URL

https://every.to/podcast/how-to-use-claude-code-like-the-people-who-built-it

## Keywords

Claude Code, Cat Wu, Boris Cherny, Anthropic, antfooding, plan mode, subagents, slash commands, stop hooks, settings.json, dangerous mode, code review, code migrations, tacit code sharing, MCP, Puppeteer, Playwright, Sentry, Asana, Bash as universal interface, CLI, IDE extension, GUI, non-technical users, extensibility

## Authors

Rhea Purohit (interviewer: Dan Shipper), Cat Wu, Boris Cherny

## Ton

**Profile:** Podcast-Interview-Technical | Third-person journalistic | Educational-Insights | Intermediate-Expert

Every (Rhea Purohit) adopts a podcast-synthesis voice capturing the conversation between Dan Shipper and the founding engineers of Claude Code. The structure by tips and themes (practical recommendations → philosophy → future vision) facilitates actionable takeaways. Product insider language (antfooding, dangerous mode, stop hooks, settings.json) reveals internal practices rarely visible to external users. A privileged-access tone typical of podcast interviews with builders sharing hard-won lessons. Direct quotes from Wu/Cherny ("In the end, the result is awesome", "this stuff's just moving so fast") preserve the authentic voice of practitioners. The emphasis on extensibility and "hackability" reflects the product philosophy. A format typical of AI & I podcasts (Every, Latent Space, Acquired) capturing builders' wisdom for an audience of practitioners seeking a competitive edge through insider knowledge.

## Pense-betes

- **Antfooding**: Anthropic employees = "ants", internal dogfooding of Claude Code, feedback every 5 minutes
- **Plan mode is crucial**: doubles/triples the success rate on complex tasks vs. one-shot attempts
- **Shared settings.json**: pre-approve common commands, block risky files, defaults for the whole team
- **Stop hooks**: automated actions when Claude finishes (e.g., run the test suite, fix failures, keep going)
- **Sparring subagents**: code review spawns multiple subagents that challenge each other
- First pass: style, history, obvious bugs
- Second pass: 5 subagents look for holes in the initial findings
- Result: finds real issues, eliminates false positives
- **Code migrations**: engineers spend $1,000+/month, the main agent creates the to-do list, subagents process items in parallel
- **Diary entries**: Claude writes logs after each task (tried, succeeded, failed), separate agents synthesize the lessons
- **Tacit code sharing**: subagents look at how other apps handle a feature, without an API
- **Favorite slash commands**:
- `/commit`: automates commits/push without permission prompts
- `/feature-dev`: structured development (spec → plan → to-do → step-by-step)
- `/code-review`: first pass on pull requests, the human approves the final merge
- **Top MCPs**: Puppeteer, Playwright, Sentry, Asana
- **Bash as a universal interface**: replaces dozens of specialized tools, reduces context, "Swiss Army knife"
- **Dangerous mode**: auto-accepts all changes, current models run 30 hours straight, next ones likely days
- **Form factors being explored**: CLI, IDE extension, GUI, GitHub integration, web, mobile
- **Upcoming challenge**: Claudes monitoring Claudes, optimizing Claude-to-Claude communication
- **Non-technical users**: data scientists, researchers, analysts, product managers adopt the tool
- **VS Code extension**: point-and-click interface vs. terminal commands
- **Philosophy**: build for everyone, let experts push the limits, observe "abuse" → build for that demand
- **Pruning discipline**: remove features when a simpler approach is found, don't add bloat
- **Every's transformation**: each feature makes the next one easier, the CEO ships to unfamiliar codebases, non-technical people in the terminal
- **Meta parallel**: Facebook Dating born from observing 60% of profile views being opposite-gender non-friends

## RésuméDe400mots

Cat Wu and Boris Cherny, founding engineers of Claude Code at Anthropic, share on Every's AI & I podcast how they use their own product and what they've learned watching hundreds of engineers adopt it through "antfooding" (Anthropic employees = "ants", their version of dogfooding).

**Practical usage insights**

**Plan mode is critical**: don't try to one-shot everything. Enabling plan mode (Claude maps out steps before coding) doubles or triples the odds of success on complex tasks by aligning the approach before execution.

**Shared settings.json**: create a settings file in the codebase with pre-approved commands (routine tasks without permission prompts) and blocked files (never touch). The whole team inherits sensible defaults instead of individual configuration.

**Creative stop hooks**: automated actions when Claude finishes. Example: run the test suite, and on failures, tell Claude to fix the problems and keep testing instead of stopping. "You can just make the model keep going until the thing is done."

**Sparring subagents**: Cherny spawns multiple subagents for code review that challenge each other. First pass: style guidelines, project history, obvious bugs. Second pass: 5 subagents poking holes in the original findings. Result: captures real issues without false alarms.

**Code migrations**: engineers spend $1,000+/month. The main agent creates a to-do list, and subagents tackle items in parallel. Particularly effective for switching testing frameworks, where output is easy to verify.

**Diary entries and institutional knowledge**: Claude writes logs after each task. Separate agents review and distill them into reusable insights. Challenge: distinguishing one-off instructions ("make button pink") from genuine best practices applicable universally.

**Tacit code sharing at Every**: when building a new feature, subagents look at how other apps have already done it. "You don't need API or ask anyone, just [ask AI] 'How do we do this already?'"

**Favorite slash commands**: `/commit` (automate commits/push), `/feature-dev` (spec → plan → to-do → structured execution), `/code-review` (first pass on PRs, human final approval).

**Product philosophy**

**Build for everyone, experts push the edges**: a product simple enough for anyone, powerful enough for advanced users discovering novel use cases the creators never imagined. Observing how users "abuse" the product reveals latent demand for new features. Meta parallel: Facebook Dating was born from 60% of profile views being opposite-gender non-friends.

**Extensibility as a core belief**: every engineering environment is unique. Slash commands, hooks, and plugins let users shape their workflow. "Insert determinism at pretty much any step." The new user experience is intuitive: "drop in and it works."

**Simplify as you scale**: discipline of pruning as much as building. Features get unshipped when a simpler approach is found. Bash as a universal interface replaces dozens of specialized tools, reducing context: "Swiss Army knife vs. a drawer full of single-purpose gadgets."

**Future vision**

**Exploring new form factors**: CLI, IDE extension, GUI, GitHub integration, web, mobile. "No one knows what [new] form factors are, this stuff's moving so fast."

**Dangerous mode evolution**: current models run 30 hours straight. The next generation will likely run for days. Practical problem: you can't leave your laptop open for days. Emerging challenge: Claudes monitoring other Claudes, optimizing Claude-to-Claude communication over human inspection.

**Non-technical users**: data scientists, researchers, analysts, and PMs are adopting it. The VS Code extension offers a point-and-click interface instead of the terminal. Every is launching a Claude Code for Beginners camp (Nov 19): a hands-on workshop covering install, assigning tasks, and building an app end-to-end, with no programming experience required.

The article illustrates how Every was transformed by Claude Code: each feature makes the next one easier, the CEO ships to unfamiliar codebases, and non-technical people suddenly find themselves in terminals.

## GrapheDeConnaissance

- Boris Cherny —a_créé→ Claude Code (TECHNOLOGIE, 0.98)
- Cat Wu —a_créé→ Claude Code (TECHNOLOGIE, 0.98)
- Anthropic —emploie→ Boris Cherny (PERSONNE, 0.97)
- Anthropic —emploie→ Cat Wu (PERSONNE, 0.97)
- Claude Code —utilise→ antfooding (METHODOLOGIE, 0.95)
- Boris Cherny —recommande→ plan mode (METHODOLOGIE, 0.96)
- Boris Cherny —recommande→ settings.json (TECHNOLOGIE, 0.92)
- Boris Cherny —utilise→ subagents (CONCEPT, 0.95)
- Claude Code —utilise→ Bash (TECHNOLOGIE, 0.93)
- Cat Wu —recommande→ Puppeteer (TECHNOLOGIE, 0.88)
- Cat Wu —recommande→ Playwright (TECHNOLOGIE, 0.88)
- Boris Cherny —prédit→ les prochains modèles tourneront plusieurs jours en autonomie (AFFIRMATION, 0.85)
- Claude Code —s_applique_à→ utilisateurs non-techniques (CONCEPT, 0.9)
- Boris Cherny —affirme_que→ l'observation des usages imprévus révèle la demande latente (AFFIRMATION, 0.88)
- Claude Code —améliore→ Every (ORGANISATION, 0.92)

---
Canonical: https://www.thekb.eu/en/fiches/wu-cherny-use-claude-code-builders-every-2025-10-29/