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

Nate Herk | AI Automation11 min2026-08-12 ▶ Watch on YouTube

What this video is

⚡ a 12-minute video, readable in 60 seconds

This video summarizes comments from Boris Cherny, the creator of Claude Code, on why Claude's performance can feel like it's degraded and how to fix it. It explains that Anthropic rewrites Claude Code's system prompt with every new model release, having deleted over 80% of it for Opus 5 because the model no longer needs corrections older models required. Cherny's core advice is to avoid overspecifying tasks, instead describing the goal, guardrails, and exit criteria, then letting Claude work and verify its own output, and to periodically strip CLAUDE.md, skills, and hooks to see what the model does unaided. The video closes by framing management of AI agents like management of people: outsource the thinking but never the understanding.

Anthropic deleted over 80% of Claude Code's system prompt for Opus 5 [00:29]

Key takeaways

Every new model triggers a rewrite of the system prompt, tools, and tool prompts because models behave differently [00:46]
Opus 5 is intelligent enough that old system-prompt corrections are no longer needed [01:12]
Recommendation: every six months, delete your CLAUDE.md, skills, and hooks to see what the model does on its own [02:04]
Boris Cherny, Claude Code's creator, said in a Y Combinator talk recorded one day after Opus 5 shipped that the harness is never finished [04:21]

+ 7 more takeaways

Evals are more durable than prompts but survive only one to three model generations before saturating [04:39]
An undocumented env var, "Claude Code simple=1," strips all system and tool prompts; Anthropic uses it internally as an ablation because the model is slightly more intelligent without them [05:09]
"Product overhang": today's models can already do things people haven't realized or discovered [07:41]
A common mistake is giving Claude overly specific step-by-step instructions instead of higher-level direction [08:10]
Boris recommends describing the task, guardrails, and exit criteria, then letting the model work and checking back later [08:28]
Boris says verification is probably the single most important thing people don't get right [10:15]
Key line: outsource some of the thinking, but never outsource the understanding [11:21]