Claude Code slash commands: files, names, and what was found

Choose session control or your own command

/help, /clear, and /compact are session controls. A custom command is your own reusable prompt. It is not a shell alias: type it inside Claude Code, rather than asking PowerShell to execute a slash-prefixed filename. The official command reference distinguishes built-in commands from skills.

Keep a harmless discovery check separate from a command that deploys, deletes, or changes a repository. The fixture below only asks for a marker and explicitly forbids tool calls. Its body was not executed by a model in our run.

# .claude/commands/sm-probe.md
---
description: Harmless SpecMatter command discovery fixture
---
Return SPECMATTER_DISCOVERY_ONLY.
Do not call tools or edit files.

The official skills documentation describes the newer SKILL.md format and backward compatibility. This page tests the command surface; our separate skills field note covers triggering and loading failures.

Eight fixtures, one unchanged CLI

Each case had its own disposable project and user configuration directory. We checked auth status --json first and continued only when loggedIn was false. The same print-mode command then emitted discovery metadata and a result. Credentials were removed from the child process environment; no subscription or API key was added.

All eight final result records reported zero input tokens, zero output tokens, and zero dollars. Seven stopped with Not logged in · Please run /login; the hidden-skill case returned an empty local result instead. Neither result establishes that a model followed the fixture instructions.

Observation CSV Discovery field extracts Original result lines

Three observations that prevent a false diagnosis

A nested file was found under a qualified name

The nested fixture was not missing. Its initialization record advertised team:sm-nested. Our attempted input was the shorter /sm-nested, and authentication then stopped the run. This measures the advertised name, not whether both spellings can invoke the command. The latter remains untested. Copy the reported identifier when investigating a discovery problem; do not assume a basename is the complete name.

Two loaded files did not mean two visible command names

The collision fixture loaded one project skill and one legacy command, while the slash-command metadata contained a single sm-shared name. The skill list also contained sm-shared. Loader counts therefore cannot tell you which body was selected. The documentation says a same-name skill takes precedence, but our unauthenticated capture did not independently verify the winning body.

2026-10-02T01:09:38.557Z [DEBUG] Loaded 2 unique skills (2 unconditional, 0 conditional, managed: 0, user: 0, project: 1, additional: 0, legacy commands: 1)

Exit zero was not a generated answer

The hidden fixture was counted by the loader but absent from both advertised lists. Its result had local_command: "custom", no turns, and an empty string. A pipeline that checks only the process exit code would call this success while receiving no answer. In this experiment, success needs separate evidence for discovery and for the expected marker; only discovery was tested.

"is_error": false,
"num_turns": 0,
"result": "",
"local_command": "custom",
"total_cost_usd": 0

These are selected original fields, not a complete result object. The downloadable original-result file preserves each full result line, in fixture order 01 through 08. Discovery extracts omit working-directory paths, session identifiers, and unrelated fields; they are labeled projections rather than full stdout. The installed CLI also advertised the built-in agents-md plugin, so this was isolated user/project configuration, not a claim that every built-in component was disabled.

Reproduce without borrowing our conclusion

Download the fixture JSON and replay script into the same folder. Supply the absolute path to an already installed official Claude Code executable. The script creates a new temporary directory, checks authentication, closes stdin, saves stdout, and leaves its files available for inspection. It refuses an authenticated configuration. It neither installs software nor opens a login flow.

Eight fixtures Replay script

node claude-code-slash-commands-replay.mjs C:\path\to\claude.exe 02-project-md
node claude-code-slash-commands-replay.mjs C:\path\to\claude.exe 07-wrong-extension
node claude-code-slash-commands-replay.mjs C:\path\to\claude.exe 08-repaired-extension

The measured CLI invocation used these arguments after the isolated environment was prepared:

claude --print --output-format stream-json --verbose \
  --setting-sources user,project --debug-file debug.log \
  /sm-probe

Record your version before comparing results. Our extension repair used a fresh directory with identical command text and a Markdown filename; it was not a claim about hot reload in an existing session. We did not test interactive autocomplete, successful authentication, argument substitution, permission approval, or a generated marker. If you test those later, retain the new output separately rather than relabeling this capture.

Sources and evidence boundaries

• Official command reference — session controls and custom command distinction.
• Official skills reference — legacy compatibility, visibility, and documented precedence.
• Official CLI reference — print mode, structured output, source filtering, and debug options.

Sources checked 2026-10-02. The eight local observations and fixture files are original SpecMatter evidence. This is not a benchmark of model quality or a promise that other versions behave identically. SpecMatter is not affiliated with Anthropic.

Skill loading and triggeringSettings source resolutionPackaged extensionsHook events