Nate Herk | AI Automation - Anthropic Engineers Just 10x'd Everyone's Claude Code

Short Summary

Main recommendation: Audit and trim your Claude Code setup (CLAUDE.md, skills, and memory files) and remove instructions that newer models no longer need, keeping only context such as your voice, goals, audience, and the intent behind tasks. The video draws on Anthropic's engineering guidance to use "the smallest amount of context that gets the result you want," and shows the presenter using the built-in /doctor command to find redundant and broken configuration. The presenter frames this as an ongoing process, not a one-time cleanup.

Key Points

• An Anthropic engineer referred to in the video as "Thoric" reportedly removed over 80% of the Claude Code system prompt for newer models and saw no measurable loss on their evaluations.
• Rules written for older models can become too restrictive; an example is a hard limit on code comments, replaced with an instruction to follow the style of the surrounding code.
• The video's core idea is minimal, outcome-focused context ("best context is the smallest amount of context that gets you the highest quality result at the cheapest cost").
• Extra input can hurt performance: a 2025 paper cited by the presenter reported drops of 13.9% to 85% across five models on math, QA, and coding as input got longer, though these are controlled experiments on older models.
• Loading MCP tool definitions only when needed reportedly cut token usage 85% in Anthropic's example, with MCP evaluation scores rising from 49% to 74% (Opus 4) and 79.5% to 88.1% (Opus 4.5).
• Boris Cherny's advice, as relayed by the presenter, is to delete all skills about every six months.

Details Worth Noting

• Concrete instructions:
  • Run /doctor inside Claude Code in a new session. It checks skills, memory, plugins, settings, and installation, and shows each skill's location, recorded uses, and estimated token cost.
  • Review flagged skills before deleting them. The audit flagged 18 older skills with no recorded use but kept recent ones; check whether a skill is used through another tool first.
  • Look for duplicated content across memory sources (project CLAUDE.md, user-wide instructions, auto-memory index, and local CLAUDE.md). Remove repeated sections, such as overlapping project map and routing map content.
  • Check for stray instruction files in other project folders, since they can load into sessions and add tokens you don't expect.
  • Fix broken skills: correct wrong file names and fix front matter formatting so Claude can read each skill's description.
  • For tasks, state the outcome, the reason, and the constraints, then leave room for the model to do the work. The presenter's audit prompt was: "Audit this repo and show me the smallest change that would make the path from a raw YouTube idea to a filming ready script faster. Use the systems already here. Cite the exact files you would reuse and do not edit anything yet."
  • Keep context that Claude cannot infer: your voice, goals, audience, and business details, plus the reason behind each request.
  • Move detailed procedures (such as a video editing workflow) out of always-loaded files and into the moment they are needed.
  • Standardize how research is handed off, for example by requiring a sources.md file that the script step checks, instead of relying on verification repeated at each step.
  • Re-run this audit monthly or quarterly, and after switching to any new model.
• The presenter's workflow audit found 25 video folders with a script but only three with a sources.md file, showing inconsistent evidence handoff.
• The presenter's trimmed setup audit found about 2,250 characters of cuts with an estimated saving of 563 tokens, plus an extra roughly 2,000 tokens loaded from an unrelated CLAUDE.md file.
• The presenter notes that the 85% and percentage figures come from controlled experiments and do not directly predict results for your own CLAUDE.md.
• The presenter suggests scheduling a local task to run the audit and then reviewing recommendations to approve or deny them.

Bottom Line

The main lesson is to treat your AI instructions as a living config that needs regular pruning, because each new model may need less guidance than the last. Start with a /doctor audit and keep only the context the model cannot infer for itself.