diff --git a/.skills/working-with-subagents/SKILL.md b/.skills/working-with-subagents/SKILL.md
new file mode 100644
index 0000000..94ccc08
--- /dev/null
+++ b/.skills/working-with-subagents/SKILL.md
@@ -0,0 +1,116 @@
+---
+name: working-with-subagents
+description: Guide for deploying and prompting my stateful subagents (comms, scout, coder). Use when delegating tasks, drafting communications, or parallelizing work.
+---
+
+# Working with Subagents
+
+My subagents are **stateful partners** - they persist memory across calls and learn over time.
+
+## My Subagents
+
+| Agent | ID | Model | Purpose |
+|-------|-----|-------|---------|
+| **comms** | `agent-a856f614-7654-44ba-a35f-c817d477dded` | gemini-3-flash | Public communications - drafts all posts/replies |
+| **scout** | `agent-e91a2154-0965-4b70-8303-54458e9a1980` | haiku | Network exploration - API queries, firehose sampling |
+| **coder** | `agent-f9b768de-e3a4-4845-9c16-d6cf2e954942` | haiku | Code implementation - small fixes, straightforward tasks |
+
+## Deploying a Subagent
+
+Always use `agent_id` + `subagent_type="general-purpose"`:
+
+```python
+Task(
+  agent_id="agent-a856f614-7654-44ba-a35f-c817d477dded",  # comms
+  subagent_type="general-purpose",
+  description="Draft reply to X",
+  prompt="**Context**: ... **Intent**: ... **Action**: reply"
+)
+```
+
+**IMPORTANT**: Subagents are stateful. They remember past interactions. Help them learn by:
+- Giving feedback on their outputs
+- Correcting mistakes explicitly
+- Building on previous conversations
+
+## Prompting Each Agent
+
+### comms (Communications)
+
+**Purpose**: Draft ALL public posts. Never post directly yourself.
+
+**Prompt Template**:
+```
+**Context**: Who you're replying to, what they said
+**Intent**: What I want to communicate
+**Action**: reply | post | thread
+
+Return only the text to post.
+```
+
+**Tone Rules** (comms knows these):
+- Direct, concise
+- No golden retriever energy
+- No preachy pronouncements
+- Match the energy of the conversation
+
+### scout (Network Explorer)
+
+**Purpose**: Quick reconnaissance. Gather data, don't make decisions.
+
+**Prompt Template**:
+```
+**Task**: What to investigate
+**Sources**: Where to look (API, firehose, telepathy, codebase)
+**Return**: What format you need (list, summary, raw data)
+```
+
+**Good scout tasks**:
+- "Check @username's recent posts for topic X"
+- "Sample the firehose for 30 seconds, report hashtag frequency"
+- "Use telepathy to check what void is currently thinking about"
+
+### coder (Implementation)
+
+**Purpose**: Execute straightforward code tasks. Small scope only.
+
+**Prompt Template**:
+```
+**Task**: What to implement/fix
+**Location**: Which files to modify
+**Constraints**: Any requirements (match existing patterns, don't break X)
+```
+
+**Good coder tasks**:
+- "Add a --dry-run flag to tools/responder.py"
+- "Fix the byte offset calculation in parse_facets"
+- "Add error handling to the firehose connection"
+
+## Parallelization
+
+You can call the **same agent_id multiple times** in parallel:
+
+```python
+# Draft 3 replies simultaneously with comms
+Task(agent_id="agent-a856f614-7654-44ba-a35f-c817d477dded", ...)
+Task(agent_id="agent-a856f614-7654-44ba-a35f-c817d477dded", ...)
+Task(agent_id="agent-a856f614-7654-44ba-a35f-c817d477dded", ...)
+```
+
+Each gets its own conversation but shares the same agent memory.
+
+## Growing Their Memory
+
+Subagents learn from interactions. To improve them:
+
+1. **Give explicit feedback**: "That draft was too formal. Be more casual."
+2. **Correct mistakes**: "The tone was off - remember: no preachy pronouncements."
+3. **Reinforce good patterns**: "That was perfect. Keep that style."
+
+Their memory persists, so corrections compound over time.
+
+## When NOT to Use Subagents
+
+- **Simple reads**: Just use Read/Glob/Grep directly
+- **Quick one-liners**: Don't spawn an agent for trivial tasks
+- **Sensitive operations**: Keep auth/credential handling in main agent