Video summary

Claude Code tried to improve /init... Is it any better?

Main summary

Key takeaways

Technology

Background / Motivation

  • The creator previously argued not to run Claude init because it produced large, bloated Claude.md files that often didn’t help much.
  • They later learned that the Claude Code team (via Tarik) wanted to make init more useful for setting up Claude Code in a repo.
  • The new approach incorporates ideas like:
    • Progressive disclosure
    • Keeping Claude.md extremely minimal (mainly environment info)

What the New init Does (Test Run on a React/TS Repo)

  1. Asks setup questions similar to a typical init CLI, including:
    • Whether to create project and/or personal Claude.md files
    • Whether to set up skills and hooks
  2. Explores the codebase (took ~1 min 54 sec) and then recommends:
    • Skills
    • Hooks
    • Based on what it finds (e.g., existing skill/hook opportunities)

Main Critique: Unnecessary Questions + “Sycophancy”

The speaker’s main objections include:

  • Dislike of init asking questions instead of proactively presenting conclusions.
    • Example: it asks something like “should you also set up skills and hooks?” and only afterward explores.
  • Strong dislike of Claude Code’s “ask user question” UI when used for skills:
    • They find it hard to view/edit
    • They consider it poor UX
  • Concern about the tool being overly agreeable (“naturally sycophantic”):
    • The speaker repeatedly tries to “grill” the tool to justify whether items should be included.

Evaluation of Generated Claude.md / Skills / Hooks

Accepted

  • Format on edit hook
    • init suggests a “format on edit” hook that runs Prettier after edits
    • The speaker accepts it.

Removed as Redundant

  • Type check command duplication
    • init suggests preferring npm run type check over npx tsc
    • The speaker already enforces type checking via hooks, so they remove it as redundant.

Kept (with justification)

  • useEffect reducer vs React.useReducer
    • init proposes a Claude.md instruction to use useEffect reducer (a specific library) instead of React’s useReducer.
    • The speaker argues it’s discoverable in code, and also worries about LLM instruction budget (they reference a rough limit like ~500 instructions).
    • They ultimately accept the line as “cheap insurance” in case the model makes a review mistake.
    • They conclude the one-line cost is worth testing.

Deemed Redundant

  • Testing patterns (Vitest / DB setup)
    • Proposed instructions such as:
      • create test DB
      • truncate all tables
      • specific Vitest + effect testing patterns
    • The speaker considers these redundant because the agent should discover such patterns during the normal exploration phase before writing tests.

Moved to a Dedicated Skill

  • Effect package installation recommendation
    • init proposes “npm install --force for effect packages.”
    • The speaker considers this rare.
    • They prefer putting it into a dedicated skill only invoked when relevant, rather than burning general Claude.md instruction budget.
    • The tool agrees (the speaker paraphrases the response as “Smart. That’s a better home for it.”).
    • The speaker still dislikes that the agreement feels too easy/sycophantic.

Final Outcome: What the Speaker Keeps vs Removes

  • The speaker trims Claude.md down to a very minimal file.
  • They keep only the useEffect reducer instruction.
  • They remove most other general guidance lines (like testing/build/type-check commands) as too redundant or unnecessary for this repo.
  • They accept a generated skill, but only for a narrow situation:
    • Skill: when installing effect packages, use npm install --force
    • Marked not user-invocable (i.e., not meant to show up as an invoke-able skill list).
  • They reject other general-purpose Claude.md guidance lines as too narrow and likely unnecessary for this repository.

Overall Verdict

  • The speaker concludes the new init is a clear improvement over the earlier approach that produced huge Claude.md automatically.
  • Remaining request to the Claude Code team:
    • Make init less sycophantic
    • Make it more proactive
    • Be better at making/defending decisions (e.g., “steelman” arguments for including items in Claude.md)

Main Speakers / Sources

  • Primary speaker: The YouTube creator evaluating the workflow (unnamed in subtitles; referred to as “I”)
  • Claude Code team member mentioned: Tarik
  • Other referenced individuals/projects:
    • Dex (contributor with progressive disclosure ideas)
    • David (author of the useEffect reducer library)
    • “Cohort course on Claude code for real engineers” (the speaker’s own course)

Original video