/cactus-juice
CACTUS JUICE MODE -- "It's the quenchiest!"
> /plugin marketplace add tmcfarlane/team-avatar > /plugin install team-avatar@team-avatar
How it fires
How this command gets triggered: by you, by Claude, or both.
- Fires itselfClaude auto-loads it when your prompt matches the work.
- You can call itInvoke it directly when you want it.
- Slash command
/cactus-juice
Context preview
What this command does when you run it.
CACTUS JUICE MODE -- "It's the quenchiest!"
Command definition
cactus-juice.mdCACTUS JUICE MODE -- "It's the quenchiest!"
Switch to Agent Swarm mode. Trade depth for massive parallelism.
Instructions
1. **Decompose** the request into 5-10 independent micro-tasks. Each micro-task must be:
- Scoped to a single file or function
- Completable in isolation without depending on other micro-tasks
- Small enough for a worker with minimal context
2. **Spawn up to 10 subagents simultaneously**, one per micro-task. Use `model: "fast"` for all workers.
3. Each worker receives ONLY a **minimal context packet**:
- The specific file path to work on
- The exact change to make
- The constraints below
4. **Collect** all results and verify cross-file consistency.
5. **Fix** any integration issues inline rather than re-spawning.
Worker Constraints (include in every worker prompt)
- You have ONE task. Complete it and return.
- Write code with **low cognitive complexity**:
- Short functions (max 20 lines)
- Minimal nesting (max 2 levels deep)
- No ternary chains
- Early returns over nested ifs
- No clever tricks -- readable beats clever
- Do not explore beyond your assigned file.
- Do not ask questions -- make reasonable assumptions.
- Return: what you changed, what file, and any concerns.
Sequencing Rules
- If a micro-task depends on another, they MUST be sequenced, not parallelized.
- Verify every worker result before considering the task complete.
- If any worker fails, fix the issue inline -- do not re-spawn.
Read more
CACTUS JUICE MODE -- "It's the quenchiest!"
Switch to Agent Swarm mode. Trade depth for massive parallelism.
Instructions
1. **Decompose** the request into 5-10 independent micro-tasks. Each micro-task must be:
- Scoped to a single file or function
- Completable in isolation without depending on other micro-tasks
- Small enough for a worker with minimal context
2. **Spawn up to 10 subagents simultaneously**, one per micro-task. Use `model: "fast"` for all workers.
3. Each worker receives ONLY a **minimal context packet**:
- The specific file path to work on
- The exact change to make
- The constraints below
4. **Collect** all results and verify cross-file consistency.
5. **Fix** any integration issues inline rather than re-spawning.
Worker Constraints (include in every worker prompt)
- You have ONE task. Complete it and return.
- Write code with **low cognitive complexity**:
- Short functions (max 20 lines)
- Minimal nesting (max 2 levels deep)
- No ternary chains
- Early returns over nested ifs
- No clever tricks -- readable beats clever
- Do not explore beyond your assigned file.
- Do not ask questions -- make reasonable assumptions.
- Return: what you changed, what file, and any concerns.
Sequencing Rules
- If a micro-task depends on another, they MUST be sequenced, not parallelized.
- Verify every worker result before considering the task complete.
- If any worker fails, fix the issue inline -- do not re-spawn.
Team Avatar: a universal agent orchestrator for Claude Code, Codex, and Cursor. Nine Avatar characters, 58 engineering skills. pstack with Avatar characters on top.
Other commands on team-avatar.
fix
Use Katara (the healer agent) for methodical debugging and code healing.
image
Use **Zuko** (the visual specialist agent) with the **cursor-image-generation** skill for this task.
plan
Analyze the following request using Sokka (the strategist agent) for the full planning pipeline:

