I Deleted All My Claude Skills... And Claude Got Smarter

Technological Concepts & Claims in the Video

• Claude Code’s system prompts and components change over time. The video argues that Anthropic updates Claude Code frequently—by adding/removing system prompt text, changing tool sets, and adjusting prompts for tools as new models release.

• Different models require different “personalities.” The creator claims that behaviors corrected by system prompts for one model may be unnecessary or ineffective for a newer, more capable model.

• “Unhobbling” / reducing constraints can improve results. The core idea: when using Claude Code, deleting or simplifying skills and system instructions can let the model perform better—especially with newer models (example mentioned: Opus 5).

---

Product / Review / Test Findings Highlighted

Practical guidance for non-agentic users (Claude Code users)

• The video recommends that every ~6 months, users should delete/retest their:
  • Claude.md
  • skills
  • hooks

The goal is to see what the model can do without those constraints.

• For Opus 5, the claim is that this can help because older instructions may have been tuned for degraded behavior in earlier models.

Counterintuitive takeaway

• Removing customization does not necessarily reduce quality. Instead, it may free the model to use its improved reasoning rather than being forced into outdated workflows.

Creator’s experiment (repo / skill removal)

• The creator removed Claude.md and skills and asked Claude to produce a YouTube resource guide.
• With skills/context:
  • Output looked more visually polished and aligned with branding (e.g., header image, colors, links).
• Without skills/context:
  • Output was less formatted, but the content structure was more idea-driven, with timestamps, and arguably better substance.

Proposed compromise: Keep skills for brand/system context (e.g., inserting header images, applying color schemes, adding channel links), but let the model handle the main task and structure more freely.

---

“Guide / Tutor” Style Process Advice Extracted

Don’t blanket-apply advice from others

The video warns that recommendations from model/harness builders (e.g., Anthropic-heavy engineering perspectives) may not fit everyone’s workflow—such as research/knowledge work vs. large-scale software engineering. Adapt the guidance to your use case.

Use higher-level instructions instead of step-by-step micromanagement

The creator emphasizes:

• Give the model harder tasks, but with more general directions.
• Avoid rigid “do step 1, step 2…” instructions.
• Instead, specify:
  • the task
  • guardrails
  • exit criteria
  • then let the model “cook”

Slash-goals / exit criteria + verification loops

• Success depends on defining what “done” means and requiring the agent to verify it meets the standard.
• The author describes wanting outputs that are:
  • tested
  • iterated
  • QA’d
  • ready “to ship tomorrow,” not just a prototype.
• The creator argues that verification is the most commonly mishandled part of agent workflows.

Agent management framing (product-oriented)

• Treat skills/tooling as agent orchestration, not just prompt engineering.
• Mindset: don’t micromanage—outsource thinking to agents, but retain understanding.
• Use agents to:
  • check their own work
  • optionally act as devil’s advocate reviewers

---

Key “Three Things Everyone Should Be Doing”

The subtitles mention “three super important things” but the provided text does not explicitly list all three in full. What is clearly covered includes:

1. Regularly delete/simplify Claude.md + skills + hooks to avoid outdated constraints (about every ~6 months).
2. Use higher-level task framing + guardrails + exit criteria rather than overly specific step lists.
3. Force/require verification—agents must check their work against a defined standard.

---

Main Speakers / Sources (As Referenced)

• Boris Cherny — Anthropic creator of Claude Code Primary technical claims include system prompt behavior, unhobbling/hobbling, harder tasks, and verification.

• Video narrator / author — the person describing experiments and applying the advice Includes the repo/Claude.md/skills removal test and resource guide examples.

• Y Combinator YouTube channel — referenced source A ~35-minute interview with Boris Cherny is mentioned.