claude-platform-playbook

Claude Platform — Common Misconceptions

As of 2026-05. A skeptic-disarmer companion to decision-memes.html. Every entry below is a misread that drives a measurable mis-budget, mis-architecture, or mis-staffing call. If a myth doesn’t change a decision, it’s not in this list.

How to read this. Format: MythRealityWhat you’d mis-decideCite. Pinned to the current model surface (Opus 4.7 / Sonnet 4.6 / Haiku 4.5). Refreshed monthly against docs/feature-inventory.md — the single source of truth.


1. Model + capability misreads

“Claude’s context window is still 200K tokens.”

“Bigger context = better answers, so always use 1M.”


2. Claude Code (CLI) surface

“Claude Code is just a chat wrapper for the terminal.”

“Custom slash commands and Skills are different systems.”

“Hooks just run shell scripts.”

sandbox.denyRead blocks the Read tool from seeing my .env.”

“Claude Code has a free tier.”


3. API economics

“Pro / Max subscriptions include API credits.”

“Prompt caching is too complex / not worth it.”

“Long context costs 2× once you cross 200K tokens.”

“Batch API is just a queue — same price.”

“Haiku is too weak for production.”


4. Safety + refusal

“Claude refuses everything sensitive.”


5. Rate limits + Computer Use

“Rate limits are just the dashboard percentage.”

“Computer Use is a paid product feature on its own SKU.”


How this artifact connects to the rest


© gmanch94 · CC-BY-4.0 · As of 2026-05. Verify pricing/models at anthropic.com.