Experimental Mode

Normally Claude Code loads a skill's (or agent's) stable release. Experimental mode lets you serve its live, in-progress draft instead — so you can iterate and see the changes in your very next session. It's gated by two switches that both have to be on. This page describes skills; the same gate applies to agents, workflows, and commands.

The two-key gate

A skill serves its live draft only when both of these are true:

1. The CLI session is in Experimental mode — a global toggle in superclawd config (named Experimental mode), off by default.
2. The skill is flagged experimental — a per-skill flag you set from the skill list.

If either switch is off, the skill behaves normally: it serves its stable release, or is skipped if it has no release. This means non-experimental skills behave identically whether or not the CLI is in experimental mode — only the skills you've deliberately flagged are affected.

When the draft loads

Because "draft" means the live draft, every in-progress edit is immediately visible to experimental sessions — including half-finished states. That's the intended behavior for an opt-in "bleeding edge" mode.

Turning on Experimental mode in the CLI

Run the config menu and toggle Experimental mode on:

superclawd config

The setting applies to future launches and is off by default; turn it off again from the same menu. While it's on, the CLI tells the backend that this session accepts experimental content.

Flagging a skill experimental

In the skill list, open a skill's 3-dots menu and choose Mark experimental (or Unmark experimental to turn it off). A flagged skill shows an amber Experimental badge in the list, alongside its Stable or Draft release badge. Admins and maintainers manage the flag; members have read-only access.

Agents work the same way

Agents use the very same two-key gate. The CLI's Experimental mode toggle is global — there's one switch for the whole session, shared by skills and agents alike — and each agent carries its own per-agent experimental flag. When both are on, that agent serves its live draft; otherwise the session gets the agent's stable release (or the agent is skipped if it has no release).

The only practical difference is that an agent is served whole — the full subagent definition at catalog time — so an experimental session picks up the agent's live draft definition in its entirety. Flag an agent experimental from its 3-dots menu exactly as you would a skill; a flagged agent shows the same amber Experimental badge.

Workflows work the same way

Workflows use the very same two-key gate. The CLI's Experimental mode toggle is global — one switch for the whole session, shared by skills, agents, workflows, and commands alike — and each workflow carries its own per-workflow experimental flag. When both are on, that workflow serves its live draft graph — the current nodes and edges, including half-finished states; otherwise the session gets the workflow's stable release (or the workflow is skipped if it has no release). Flag a workflow experimental from its 3-dots menu exactly as you would a skill; a flagged workflow shows the same amber Experimental badge.

Commands work the same way

Commands use the same two-key gate too. With the global Experimental mode toggle on and a command flagged experimental, that command serves its live draft — both its body and its arguments. Because the slash command's argument hint (catalog) and its executed body (on-demand) both honor the gate, an experimental session sees the draft hint and runs the draft body — they never disagree. Otherwise the session gets the command's stable release (or the command is skipped if it has no release). Flag a command experimental from its 3-dots menu; a flagged command shows the same amber Experimental badge.

Memories, MCP servers, and resources (no draft)

Memories, workspace MCP servers, and resources honor the same Experimental mode toggle, but they have no draft/release split — there's just the one record. So the gate is simpler: when one of these is flagged experimental, it's only injected into experimental-mode sessions; everyone else's session skips it entirely. Use it to trial a memory, a new integration, or a reference doc with yourself before rolling it out to the whole workspace. Non-experimental items are always injected. Each flagged item shows the same amber Experimental badge.

A typical workflow

1. Flag the skill you're iterating on as experimental.
2. Turn on Experimental mode with superclawd config.
3. Edit the draft and relaunch — your session now runs the live draft for that skill.
4. When the draft is solid, create a New Release and Promote it to stable (see Versions & Releases).
5. Unmark experimental and turn the CLI toggle back off so everyone runs the stable release.