What Claude Code Can Do That You Haven't Tried - Gui Ferreira - NDC Oslo 2026

Summary

Gui Ferreira presents a rapid-fire tour of Claude Code features and workflows that go beyond code generation. His central message is to treat Claude Code as a configurable, extensible assistant—not just an autocomplete tool—and to choose a few capabilities that improve your own workflow.

Setup and configuration

• Import settings from other coding tools: Use /import to bring over items such as MCP servers, configuration files, and subagents.
• Keep CLAUDE.md concise: Ferreira says long instruction files can create duplication and confusion. He cites an ETH Zurich study reporting that lengthy files reduced performance by about 3% and increased costs by about 20%. He recommends keeping only broadly applicable instructions, moving occasional procedures into skills, and avoiding details the agent can discover from the project itself.
• Organize instructions by scope:
  • Use @path imports to include shared or specialized reference files without duplicating their contents.
  • Put language- or file-specific guidance in .claude/rules.
  • Use CLAUDE.local.md for personal, Git-ignored preferences, and the user-level .claude directory for instructions that should apply across projects.
• Check configuration: Use /doctor to review configuration and project instructions.
• Protect sensitive files and commands: Configure deny permissions in .claude/settings.json to prevent Claude Code from reading specified files or running specified commands.
• Customize the interface: Choose a theme and create a status line that displays useful session details, such as the project, Git branch, model, and context usage.

Day-to-day workflow

• Plan before implementing: Use plan mode to research an approach. Press Ctrl+G to open the plan as a Markdown file in an editor, revise it directly, and let Claude continue from the edited version.
• Clarify requirements: Ask Claude to keep interviewing you until key requirements and assumptions are clear.
• Rewind instead of correcting a bad direction in place: Use double Escape or /rewind to return to an earlier point and avoid unnecessary implementation history in the context.
• Choose models and reasoning effort deliberately: Switch models with /model, adjust effort levels, and use “ultrathink” for a single prompt that needs maximum effort.
• Manage prompts while Claude works: Ctrl+S stashes a prompt for later; /btw asks a side question while the main task continues.
• Track spending and context: /status shows session and model usage; /context reveals what is taking up context, including automatically loaded material. Ferreira advises checking an empty session to spot unnecessary overhead. He describes 40% context usage as a commonly cited danger-zone heuristic, not a research-backed threshold.
• Compact context proactively: Use /compact at a natural break in a task, especially when the conversation is getting repetitive or stuck in trial and error.
• Audit memories: Use /memory to inspect saved information, remove incorrect or irrelevant entries, and promote useful preferences into project instructions.
• Run simple commands directly: Prefix a command with ! to execute it as Bash rather than asking Claude to reason through it. Command output still enters the context, so avoid exposing sensitive information.
• Review changes: Use /diff to view modified files and line-level changes.

Parallel work and automation

• Remote control lets you monitor and interact with a running Claude Code session from the Claude app, including while away from your desk.
• Voice mode (/voice) can help you explain a problem aloud and turn speech into a prompt. The speaker notes that it may not suit open offices.
• Use images as input: Paste or drag screenshots into the terminal to provide visual references, such as a design to reproduce.
• Use Git worktrees to work on separate tasks in parallel within the same repository. Ferreira suggests using /color and /rename to make sessions easier to distinguish.
• Set a verifiable goal with /goal: Define both the desired outcome and how Claude should test or verify it, so it can iterate without constant supervision.
• Reduce status noise with /focus: Show the original request and final result rather than every intermediate update.
• Manage multiple agents: Use the desktop app or the terminal’s agents view to see which agents are running, finished, or need attention.

Extending Claude Code

• Hooks can automatically run tests, linters, or other checks and feed their results back to Claude. Ferreira recommends them for fast feedback rather than relying only on written instructions.
• Skills can package more than prompt text: a skill folder can include scripts, documentation, and supporting files, making repeatable team procedures easy to invoke.
• Built-in skills include:
  • /patch for dividing large changes across agents.
  • /simplify for addressing duplication and maintainability issues.
  • /loop for repeating a check at an interval, such as monitoring whether a deployment has completed.
• Share MCP configuration at the project level: Put project-used servers in .mcp.json so the team can review and share the setup.
• Use headless mode: claude -p runs a prompt without the interactive session flow. This can support shell scripts, pipelines, and other command-line automation.

Uses beyond software development

Ferreira also describes using Claude Code with ordinary files and personal workflows:

• Review project work to help prepare quarterly performance summaries.
• Sort a cluttered downloads folder into categories for easier cleanup.
• Work with Markdown notes in an Obsidian vault to summarize meetings, surface future tasks, and connect related notes.
• Create skills that import Strava run history into a vault and help plan a sustainable weekly running schedule.

The talk is a practical feature guide rather than a step-by-step product demo. Ferreira’s closing suggestion is to try /insights to examine how you use Claude Code.

Speaker and sources

• Speaker: Gui Ferreira, who describes himself as an independent Claude Code user and workshop instructor, not affiliated with Anthropic.
• Research cited: An ETH Zurich study on the effects of agent instruction files such as CLAUDE.md and AGENTS.md.
• Event/source: NDC Oslo 2026.