diff --git a/dot_config/opencode/skills/generate-skill/SKILL.md b/dot_config/opencode/skills/generate-skill/SKILL.md
new file mode 100644
index 0000000..d928300
--- /dev/null
+++ b/dot_config/opencode/skills/generate-skill/SKILL.md
@@ -0,0 +1,32 @@
+---
+name: generate-skill
+description: Create an OpenCode skill from a successfully verified reusable workflow. Use when the user asks to create or save a skill, or approves a proactive skill proposal.
+---
+
+# Generate Skill
+
+Create a concise, evidence-based skill only after the user approves it. Capture the reusable solution, not the history of the conversation or investigation.
+
+## Workflow
+
+1. Read the current OpenCode skill documentation at `https://opencode.ai/docs/skills/`.
+2. Inspect existing repository and global skills to avoid duplicates.
+3. Determine the scope:
+   - Use `.opencode/skills/<name>/SKILL.md` for repository-specific commands, architecture, infrastructure, or conventions.
+   - Use `~/.config/opencode/skills/<name>/SKILL.md` only for workflows portable across unrelated repositories.
+   - Prefer repository scope when uncertain. Ask one concise question if the correct scope is unclear.
+4. Generalize the successfully verified solution into the shortest procedure that would help with the same task again.
+5. Include verification steps when success is not self-evident.
+6. Include common errors only when they reveal a durable constraint or non-obvious part of the workflow.
+7. Create the skill and validate its location, frontmatter, name, and description against the documentation.
+
+## Requirements
+
+- Keep the first version brief; expand it only after later use reveals missing details.
+- Use a lowercase hyphen-separated name matching the containing directory.
+- Write a specific description that states what task the skill performs and when it should be loaded.
+- Keep every instruction relevant to performing or verifying that task.
+- Include only stable, reusable knowledge. Generalize repository-specific details without removing details required for the workflow to work.
+- Do not include credentials, private data, conversation history, temporary paths, transient system state, missing local tools, irrelevant debugging history, or unverified assumptions.
+- Do not preserve failed attempts unless they expose a durable and likely recurring pitfall.
+- Do not create a duplicate or overwrite an existing skill without explicit approval.