diff --git a/modules/opencode/agents/code-designer.md b/modules/opencode/agents/code-designer.md
index 7913aba..c53d1a5 100644
--- a/modules/opencode/agents/code-designer.md
+++ b/modules/opencode/agents/code-designer.md
@@ -4,9 +4,8 @@ mode: subagent
 permission:
   edit: deny
   write: allow
-  bash:
-    "*": allow
-  lsp_*: allow
+  bash: allow
+  lsp: allow
   read: allow
   grep: allow
   glob: allow
diff --git a/modules/opencode/agents/code-implementer.md b/modules/opencode/agents/code-implementer.md
index 4ca7290..434ca17 100644
--- a/modules/opencode/agents/code-implementer.md
+++ b/modules/opencode/agents/code-implementer.md
@@ -3,9 +3,8 @@ description: Implements code from design documents and task descriptions. Use af
 mode: subagent
 permission:
   edit: allow
-  bash:
-    "*": allow
-  lsp_*: allow
+  bash: allow
+  lsp: allow
   read: allow
   grep: allow
   glob: allow
diff --git a/modules/opencode/agents/explore.md b/modules/opencode/agents/explore.md
index 5be45b5..d1fb5d5 100644
--- a/modules/opencode/agents/explore.md
+++ b/modules/opencode/agents/explore.md
@@ -12,7 +12,7 @@ permission:
   websearch: allow
   codesearch: allow
   read: allow
-  lsp_*: allow
+  lsp: allow
 ---
 
 You are a file search specialist. You excel at thoroughly navigating and exploring codebases.
diff --git a/modules/opencode/commands/bounce.md b/modules/opencode/commands/bounce.md
index fbd87ed..23d7cdd 100644
--- a/modules/opencode/commands/bounce.md
+++ b/modules/opencode/commands/bounce.md
@@ -1,19 +1,8 @@
 ---
-description: Bounce ideas - collaborative problem solving
+description: Bounce ideas - collaborative problem solving with human
 ---
 
-### PROTOCOL: ARCHITECTURAL REVIEW
+**Protocol:** Bounce protocol: What's your current mental model? What's the simplest version that could work? What assumptions are you making that might be wrong? What's being overengineered? Ask questions. Challenge assumptions. Push back on overcomplication. Get us to a concrete approach before coding. No implementation.
 
-**Focus:** $ARGUMENTS
-
----
-
-**Objective: Analyze logic and surface hidden requirements. No implementation.**
-
-1.  **Alignment:** Rephrase the core objective. What is the "success state" here in one sentence?
-2.  **Minimum Viable Logic:** Strip the idea to its skeleton. What is the absolute smallest version of this that provides value?
-3.  **Risk Map:** Identify the "linchpin" assumption. If this one thing fails, does the whole project fail?
-4.  **Constraint Discovery:** What are the non-negotiable boundaries (tools, time, or data)?
-5.  **Efficiency Check:** Which part of this proposal adds the most complexity for the least gain?
-
-**Next Step:** Ask 2 targeted questions to clarify the "Risk Map." Do not write code.
+User wants to bounce ideas:
+$ARGUMENTS
diff --git a/modules/opencode/commands/delegate.md b/modules/opencode/commands/delegate.md
index bc3f8a7..011c8c5 100644
--- a/modules/opencode/commands/delegate.md
+++ b/modules/opencode/commands/delegate.md
@@ -2,6 +2,7 @@
 description: Wrap input and delegate to relevant agent
 ---
 
-$ARGUMENTS
+**Protocol:** Delegate to the relevant agent.
 
-Don't forget to delegate to the relevant agent.
\ No newline at end of file
+User wants to delegate:
+$ARGUMENTS
\ No newline at end of file
diff --git a/modules/opencode/commands/explore.md b/modules/opencode/commands/explore.md
new file mode 100644
index 0000000..ee9bbe1
--- /dev/null
+++ b/modules/opencode/commands/explore.md
@@ -0,0 +1,10 @@
+---
+description: Explore and load relevant code into context using LSP
+agent: explore
+subtask: false
+---
+
+**Protocol:** Use LSP (goToDefinition, findReferences) to find and read relevant code. Tests included if relevant. End with a 3-bullet summary of what was found. No implementation.
+
+User wants to explore:
+$ARGUMENTS
diff --git a/modules/opencode/commands/grill.md b/modules/opencode/commands/grill.md
index 00ab70d..137a1ad 100644
--- a/modules/opencode/commands/grill.md
+++ b/modules/opencode/commands/grill.md
@@ -3,12 +3,7 @@ description: Adversarial review of decisions, plans, or designs
 agent: orchestrator
 ---
 
-Load the grill-me skill and conduct an adversarial review of my plan/decision/design.
+**Protocol:** Load the grill-me skill and conduct an adversarial review. Walk the full decision tree one question at a time: Surface hidden dependencies and failure modes. Make tradeoffs explicit. Present 2-3 alternative approaches with different tradeoff profiles. Do not proceed to the next question until I have a concrete, defensible answer.
 
-Walk the full decision tree one question at a time:
-
-- Surface hidden dependencies and failure modes
-- Make tradeoffs explicit
-- Present 2-3 alternative approaches with different tradeoff profiles
-
-Do not proceed to the next question until I have a concrete, defensible answer.
+User wants to grill:
+$ARGUMENTS