Claude's Creator Just Told Us How to Actually Use Claude in 2026

Overview

The video argues that best practices for using Claude in 2026 are fundamentally different from those from ~6 months prior. This shift is attributed to:

• Newer Anthropic models responding better with less instruction
• Improved Claude “agent” capabilities, including verification and safer autonomy

---

1) “Subtraction” / Ablation (remove old instructions)

The speaker cites Claude Code lead Boris Cherny, who recommends users remove:

• The Claude MD (long custom instruction/system-style file)
• Custom skills
• Custom instructions

Why this helps

The rationale is based on an internal ablation process: Anthropic supposedly removes most of the system prompt and then re-adds only what’s needed. Newer models are claimed to perform better when users don’t over-direct them, because many prior instructions existed to compensate for older model weaknesses.

Practical guidance

• Delete old instruction layers, experiment to see what breaks, then gradually re-add only necessary lines.
• If you remove anything, remove instructions that micromanage behavior (e.g., “think/respond/act like this”).
• Keep only essentials:
  • context
  • goals
  • a clear task success definition
• Warning: too much context/instruction can reduce performance.

Mentioned resource

The creator provides a free “ablation prompt” that reads your current settings and outputs a concise, tailored instruction set.

---

2) Verification via better prompt composition (avoid over-specifying)

The second lesson claims the most common user mistake is overly detailed prompts, especially ones that explain how to think step-by-step.

Instead, users should:

• Specify the desired outcome
• Define what success means
• Avoid micromanaging the model’s internal reasoning steps

Claimed benefit

Newer models (especially “Mythos class models” like Fable) reportedly do worse when users over-direct them.

Example: Bun rewrite

• Bun is a JavaScript runtime originally written in Zig
• The creator rewrote it in Rust to reduce memory errors/crashes/leaks
• Claude Code reportedly rewrote 530k+ lines in ~11 days using:
  • only one engineer prompting
  • many “agent coders” mentioned (~64)

Key takeaway

The value isn’t just coding—it’s that Claude can verify work at each stage, making checking more automatic rather than manual step-by-step supervision.

Additional claim: CEO Dario Amodei said Claude-written code is increasingly generated by Claude itself under engineering supervision, shifting humans toward monitoring/testing.

---

3) Test hidden capabilities + weekly “impossible” challenges

The third lesson focuses on discovering that models can do more than expected.

Anecdote

Claude was reportedly found to be able to draw portraits/landscapes using OpenCV without prior teaching.

Guidance

• If something “crazy” comes to mind, try it—creators may not know all capabilities.
• Set a weekly goal: give Claude a task you believe is impossible.

Claude Code automatic mode

In automatic mode, Claude can act on your behalf without constant approvals and can run scheduled/looped tasks (with later “cycles/loops” discussed).

---

4) Agents managing other agents (agentic workflows while you sleep)

The fourth lesson is “agents managing agents.”

Claim

Attributed to Boris: instead of communicating directly with agents, he has agents communicate with other agents—described as thousands working while he sleeps.

Example of internal implementation

• Claude independently maintains its own codebase
• Performs 20–30 daily routines, each described as one sentence

Practical advice

• Don’t launch an “army” of many agents at once.
• Start by automating one recurring weekly/daily task.
• Add more over time (aim: automate several workflows across months).

Claude feature mentioned

Users can schedule a task so Claude automatically creates/launches an agent.

---

5) “Cycles” / Loops for autonomy + safer execution (prompt injection mitigation)

The fifth lesson argues that “prompting is dead” (or insufficient alone). The future is loops/cycles, where the model tests and corrects errors itself.

Safety change highlighted

Earlier autonomous agents were risky due to prompt injection (malicious instructions). The claim is that Anthropic models now include classifiers and improved infrastructure so prompt injection “no longer works.”

How cycles are structured (conceptual syntax)

• A /goal defines the end result
• A /loop defines the schedule/interval
• Claude runs mini-checks, attempts the task, and returns only when readiness/success criteria are met

Orchestration across models

• Use Fable as an orchestrator to select which model to run per step (route tasks to Opus/Haiku/Sonnet depending on optimization).
• Fable is described as more expensive, so model choice depends on complexity and budget.

Example loop

• Scan a competitor channel for new videos
• Record metrics into a “tracker MD file”
• Flag anomalous videos and stop after a number of starts

---

6) Graph engineering (turn workflows into structured graphs for AI)

Additional best practices emphasize graph engineering as a productivity unlock.

Core idea

Turn business/process work (often “in the head”) into:

• variables
• a defined sequence of steps
• a visual diagram (graph) represented in HTML/Markdown

What the graph clarifies

• what must happen at each stage
• why each step matters
• who/what executes it:
  • AI vs human approval
  • which AI + skills

Claimed benefit: “knowledge compression”

Transferring how you think into AI in a format humans can understand, helping identify which SOPs can be AI-executed versus where humans must verify.

Practical workflow

• Create one graph per task/process
• Maintain a library (e.g., Notion / Google Sheets + HTML/MD files)
• Update graphs as the business changes
• Start with one process—don’t get overwhelmed

---

Reviews / guides / tutorials explicitly promised or referenced

• A free “ablation prompt” for safer instruction removal/restoration
• A link to a “Loops masterclass” (loop/cycle design)
• A link to a “graph engineering” masterclass, plus more graphs/resources in a School community
• Testing encouragement:
  • one-day ablation experiments
  • weekly “impossible” challenges
  • scheduled/automatic agent runs

---

Main speakers / sources

• Boris Cherny (Anthropic; creator of Claude Code) — central source for “subtraction/ablation,” verification, and internal practices
• Dario Amodei (Anthropic CEO) — cited about Claude-written code under engineering supervision
• Video creator/host — main narrator; discusses their business implementation, tests, and provides prompts/resources (name not stated in subtitles)