26 Hidden Commands and the Case for Opinionated AI Skills
Hidden slash commands in Claude Code reveal what power users actually need: sharp, opinionated workflows. Here's what skill builders should learn from them.
The Claude Code community has been digging through leaked source material again, and this time the findings are instructive for anyone building AI skills. Twenty-six hidden slash commands, over 120 environment variables, and 32 feature flags paint a picture of where developer tooling is heading. More importantly, they reveal a design philosophy -- consistent with the 43 built-in tools architecture -- that every skill builder should internalize.
A Note on Sources
The findings discussed in this article come from CCLeaks, a community project that analyzes publicly available Claude Code artifacts. This content is AI-generated, may contain errors or misinterpretations, and is not affiliated with or endorsed by Anthropic. Nothing here should be treated as confirmed product roadmap. Think of this as signal reading -- interpreting patterns in available data to make educated guesses about where the ecosystem is heading. Build your skills based on user needs, not leaked commands.
Key Takeaways
- 26 hidden slash commands, 120+ environment variables, and 32 feature flags were found in Claude Code's source artifacts
- Every hidden command is opinionated -- it assumes a specific context, workflow, and outcome rather than trying to handle everything
- Generic skills get installed and forgotten; narrow, workflow-specific skills get used daily
- The best skill design pattern: identify a repeated workflow, make strong assumptions, define clear input/output contracts, and name it like a command
- Build-time code elimination of feature flags means platform capabilities are evolving fast -- build skills that sit above platform features, not replicate them
The 26 Commands Nobody Told You About
The discovered commands break into several clusters, and each cluster tells a story about what sophisticated users actually need from their AI coding assistant.
Workflow Automation Commands
Commands like /autofix-pr, /subscribe-pr, and /bughunter point toward a future where Claude Code does not just respond to prompts -- it actively participates in engineering workflows. /autofix-pr suggests automatic remediation of pull request issues. /subscribe-pr implies persistent monitoring of PR state changes. /bughunter hints at proactive bug detection rather than reactive fixing.
These are not generic "help me code" features found in the Claude Code documentation. They are sharp, opinionated tools that assume a specific workflow: you open a PR, the system watches it, finds problems, and fixes them without you asking.
Planning and Cognition Commands
/ultraplan is perhaps the most interesting discovery. It appears to relate to remote cloud-based planning -- the idea that Claude could spin up a separate planning process that runs independently before returning a structured execution plan. /dream connects to memory consolidation, suggesting the system might process and reorganize learned context during idle periods.
These commands indicate that the future of AI development tools is not faster responses to prompts. It is deeper, more deliberate cognition that happens in the background.
Debugging and Introspection Commands
/debug-tool-call, /perf-issue, /ant-trace, and /ctx-viz form a cluster around observability. /ctx-viz likely visualizes the context window -- what the model is actually seeing when it processes your request. /ant-trace suggests distributed tracing for multi-step agent operations.
The existence of these commands tells us that power users need to understand what is happening inside the system, not just consume its outputs.
System Management Commands
/force-snip, /mock-limits, /break-cache, /reset-limits, /bridge-kick, and /backfill-sessions reveal the operational surface area of running Claude Code at scale. These are the commands engineers use when things go wrong or when they need to test edge cases.
What Hidden Commands Teach Us About Skill Design
Here is the lesson that matters for skill builders: every one of these 26 commands is opinionated. None of them are "help me with anything." Each one assumes a specific context, a specific workflow, and a specific outcome.
This is the single most important principle for building skills that people actually use.
Generic Skills Die on the Shelf
The temptation when building an AI skill is to make it broad. "A skill that helps with code review." "A skill that assists with documentation." "A skill for project management."
These skills get installed, tried once, and forgotten. They do not outperform a well-crafted prompt because they try to handle every situation and end up handling none of them well.
Opinionated Skills Get Used Daily
The hidden commands show a different approach. /autofix-pr does not "help with pull requests." It fixes specific issues on a specific PR automatically. /bughunter does not "assist with debugging." It proactively hunts for bugs in a defined scope.
The best skills in the aiskill.market catalog follow this pattern, as previous CCLeaks analysis has consistently shown. They make assumptions about your workflow, your tools, and your desired output format. Those assumptions are what make them fast and reliable.
The Narrow-But-Powerful Pattern
Consider what a skill version of each hidden command cluster would look like:
Workflow skills that assume specific engineering practices:
- "Turn product manager notes from a Notion doc into a structured GitHub issue with acceptance criteria, technical notes, and effort estimate"
- "Convert a merged PR into release notes formatted for our changelog, Slack announcement, and customer email"
- "Audit a landing page for GEO readiness and output a prioritized fix list with code snippets"
Planning skills that structure thinking before execution:
- "Analyze a feature request against the current codebase and produce an implementation plan with file-level changes, risk assessment, and estimated token cost"
- "Review a quarter's worth of support tickets and produce a prioritized technical debt backlog"
Debugging skills that go beyond "find the bug":
- "Trace a user-reported error from the frontend through API calls to the database query and produce a root cause analysis with fix recommendations"
- "Profile the critical rendering path of a Next.js page and produce specific optimization recommendations with expected impact"
Each of these is narrow. Each makes strong assumptions. Each delivers a specific, valuable output. That is the pattern.
Environment Variables and Feature Flags: The Configuration Layer
Beyond the slash commands, the discovery of 120+ environment variables and 32 feature flags reveals something about how Anthropic thinks about feature delivery.
Build-Time Code Elimination
The 32 feature flags use build-time code elimination, meaning features are not just toggled off -- they are physically removed from the shipped binary when disabled. This is a sophisticated approach that suggests these features are in various stages of readiness, from internal testing to staged rollout.
For skill builders, this matters because it tells you the platform is moving fast. Features you build skills around today might become native capabilities tomorrow. The smart play is to build skills that sit above platform features rather than replicate them.
Secret CLI Flags
The discovered flags -- --bare, --dump-system-prompt, --daemon-worker, --bg, --spawn -- reveal the architectural direction. --daemon-worker and --bg suggest background processing. --spawn implies the ability to launch sub-processes. These are the building blocks of agentic systems that can run multiple operations concurrently.
Skills that can orchestrate multi-step workflows, kicking off background tasks and coordinating results, will become increasingly natural as the platform evolves to support these patterns natively. For more on how Claude Code's tool architecture supports this direction, see our deep dive on the 43 built-in tools.
Building Your Own Opinionated Skills
If these findings have convinced you that narrow, opinionated skills are the path forward, here is how to start building them.
Step 1: Identify a Specific Workflow You Repeat
Do not start with "what would be useful." Start with "what did I do three times this week that followed the same pattern." The best skills automate workflows you already execute manually.
Step 2: Make Strong Assumptions
Your skill should assume a specific tech stack, a specific project structure, a specific output format. Every assumption you make is a decision the user does not have to make, which means faster execution and more consistent results.
Step 3: Define the Input and Output Contract
The hidden commands each have an implicit contract. /autofix-pr takes a PR reference and produces fixed code. /bughunter takes a codebase scope and produces a bug report. Your skill should have equally clear boundaries.
Step 4: Handle the Happy Path Extremely Well
Do not try to handle every edge case in version one. Handle the most common scenario perfectly. Users will tell you what edge cases matter, and you can iterate from there.
Step 5: Name It Like a Command, Not a Feature
/autofix-pr is better than "Pull Request Enhancement Assistant." Name your skill after what it does, not what it is. Users scan skill names looking for verbs and outcomes.
The Library of Narrow Capabilities
The 26 hidden commands point toward a future where your AI development environment is not one monolithic assistant. It is a library of narrow, powerful capabilities that you invoke as needed. Some of those capabilities ship with the platform. Many more come from the community.
The skill marketplace exists to be that library. The skills that win are not the ones that try to do everything. They are the ones that do one thing so well that using them becomes automatic -- the same way you type git commit without thinking about it.
The hidden commands show us what Anthropic's own engineers need: specific tools for specific jobs. Build your skills the same way, and you will build skills that people actually use.
Frequently Asked Questions
What are opinionated AI skills?
Opinionated AI skills are tools that make strong assumptions about context, workflow, and output format rather than trying to handle every possible scenario. They assume a specific tech stack, project structure, and desired outcome, which makes them faster and more reliable than generic alternatives.
What hidden commands exist in Claude Code?
Analysis of Claude Code artifacts has surfaced 26 hidden slash commands spanning workflow automation (/autofix-pr, /subscribe-pr, /bughunter), planning and cognition (/ultraplan, /dream), debugging and introspection (/debug-tool-call, /ctx-viz, /ant-trace), and system management (/force-snip, /mock-limits, /break-cache).
What is the difference between a generic skill and a workflow skill?
A generic skill tries to handle broad categories like "help with code review" and ends up handling none of them well. A workflow skill automates a specific, repeatable process -- like converting a merged PR into formatted release notes for a changelog, Slack announcement, and customer email.
How should I name my AI skill?
Name your skill after what it does, not what it is. /autofix-pr is better than "Pull Request Enhancement Assistant." Users scan skill names looking for verbs and outcomes, so lead with the action.
Are the hidden commands available to use today?
These commands were discovered through community analysis of publicly available artifacts and are not officially documented by Anthropic. They may be in various stages of development, testing, or internal use. Build your skills based on user needs rather than unreleased features.
What to Build Next
If you are looking for skill ideas, work backward from the command clusters:
- Workflow automation: What repetitive multi-step process in your team could become a single command?
- Planning and analysis: What decision do you make regularly that would benefit from structured, deliberate analysis?
- Debugging and observability: What diagnostic process do you wish you could delegate entirely?
- System management: What operational task requires tribal knowledge that should be codified?
The answers to these questions are your next skills. Make them narrow. Make them opinionated. Make them useful.
Explore production-ready AI skills at aiskill.market/browse or submit your own skill to the marketplace.