[Bug] Claude Code Agent Lacks Knowledge of Built-in Features and Commands

Bug Description
Bug Report: Claude Code Agents Cannot Advise on Claude Code's Own Features

From: Jordan Dea-Mattson
GitHub: @jordan-of (OrdinaryFolk), @jordandm (personal)
Email: jordan-of@users.noreply.github.com
Related: #33059 (Improve product knowledge accuracy and coverage for Claude Code)

Problem

When a user asks a Claude Code agent about Claude Code's own features, the agent frequently cannot help — or worse, fabricates details that don't exist. This is especially frustrating when the
user learned about the feature from Anthropic's own marketing — a blog post, a video demo, or Anthropic staff promoting capabilities on Twitter, Threads, or LinkedIn. The user sees the feature
showcased, gets excited, asks the agent "how do I use this?" — and the agent either draws a blank or invents an answer.

Example 1: Complete ignorance. The user asked about /remote-control — a standard Claude Code slash command listed in the / menu. The agent responded "I don't know what that is" and "I haven't seen
it documented anywhere." This is a built-in command in the product the agent is running inside of.

Example 2: Confident fabrication. When a research subagent was tasked with documenting /remote-control, it produced detailed information that was largely accurate but included fabricated UI
details — describing a "green status dot" that indicates online status. The actual UI shows sessions in static lists grouped by "Idle" and "Archived" headers, with a small blue dot per session.
The dot does not change based on session state (working, waiting for input, or idle). The agent invented a dynamic status indicator that doesn't exist, got the color wrong, and missed the actual
list structure — presenting all of this with the same confidence as real details, making it impossible for the user to know what's accurate without independently verifying every claim.

The cost of this: The typical flow is: multiple Anthropic staff working on Claude Code make yet another post on social media excitedly telling us about a new feature with a 15-second video clip
showing yet another amazing feature (YAAF) that can make my life easier → user asks their agent about it → agent can't help → user spends an hour or more Himalayan Wandering (the learning
equivalent of Yak Shaving) — testing the feature manually, taking screenshots, comparing reality to the agent's fabrications, filing bug reports about the discrepancies. Work that should have been
a 30-second conversation turns into an exploratory expedition. This is the opposite of what an AI assistant should deliver.

The Friction

Claude Code is moving fast — new features, commands, and capabilities are shipping regularly. Anthropic actively promotes these features through social media, blog posts, and video demos. But the
agents running inside Claude Code have stale or absent knowledge of the very features being promoted. This creates a frustrating cycle:

• Anthropic promotes a new capability (video, tweet, blog post)
• User gets excited and asks their agent "how do I use this?"
• Agent says "I don't know" or fabricates details that sound right but aren't
• User has to research it themselves — watching the video frame by frame, reading docs, trial and error
• Or the user trusts the agent's fabricated answer and wastes time chasing things that don't exist
• The feature that was supposed to save time ends up costing time

This undermines trust in the agent's capabilities broadly. If it can't accurately answer "what does /remote-control do?" — a question about its own product — why would the user trust it on harder
questions?

What's Needed

1. Claude Code's own documentation should be part of the agent's knowledge. Every built-in command, flag, hook event, plugin API, and configuration option should be in the model's training data or

available as a tool.

1. The agent should be able to introspect. It should be able to list its own commands, describe what they do, and explain how to use them — without guessing.
2. Changelog awareness. When features ship, the agent should know about them. The gap between "feature shipped" and "agent knows about it" should be days, not months. If Anthropic is promoting a

feature on social media, the agent should already know about it.

1. When uncertain, say so clearly. "I'm not sure about that feature — check /help or the docs" is better than "that doesn't exist" (when it does) or fabricating details (the green dot that isn't

there).

1. Never fabricate UI details. If the agent hasn't verified a UI element exists, it should not describe it. "The session appears in your session list" is honest. "The session appears with a green

status dot" is a lie that wastes the user's time looking for something that doesn't exist.

Impact

This is not a nice-to-have. Claude Cod…
Note: Content was truncated.