Skill for project-specific skills creation, editing or validation

.gitignore

@@ -13,7 +13,13 @@ npm-debug.log*
 
 # OpenCode local assets
 .opencode/
+.codex
+.codex/
 ai-docs/
+eval/runtime
 
 # Pack artifacts
 *.tgz
+
+# dev folder
+dev/

doc/en/README.md

@@ -11,18 +11,18 @@ This section contains complete user documentation for installing, launching, and
 
 1. [Installation](./installation.md)
 2. [Quick Start](./quick-start.md)
-3. [Agents and When to Use Them](./agents.md)
-4. [Integrations and API Keys](./integrations-and-keys.md)
-5. [Update and Backups](./update-and-backup.md)
-6. [Troubleshooting](./troubleshooting.md)
-7. [FAQ](./faq.md)
+3. [Skill-Authoring Guide](./skill-authoring.md)
+4. [Agents and When to Use Them](./agents.md)
+5. [Integrations and API Keys](./integrations-and-keys.md)
+6. [Update and Backups](./update-and-backup.md)
+7. [Troubleshooting](./troubleshooting.md)
+8. [FAQ](./faq.md)
 
 ## Package in Brief
 
 `assistagents` is an interactive installer that prepares the `~/.opencode` directory:
 
 - installs/updates `agents`, `skills`, and `commands` from package templates;
-- optionally installs experimental hash-based tools into `tools`;
 - helps enable MCP integrations;
 - stores keys in `~/.opencode/keys`;
 - generates `~/.opencode/opencode.jsonc` with the required MCP config and permission restrictions.
@@ -35,7 +35,7 @@ Run this command via the `build/dev` agent.
 
 - Run without arguments to auto-detect languages from repository code and configs.
 - Run with arguments to pass a comma-separated language list, for example `typescript,csharp`.
-- Output: creates or updates project-local skills in `.opencode/skills/coder/` inside the current repository.
+- Output: creates or updates project-local skills in `.opencode/skills/` inside the current repository.
 
 For step-by-step usage and examples, see [Quick Start](./quick-start.md).
 

doc/en/installation.md

@@ -26,9 +26,8 @@ The installer asks step by step:
 1. Whether to create a zip backup of the current `~/.opencode`.
 2. Preferred response language for agents.
 3. Whether to explicitly set models for selected agents (optional).
-4. Whether to enable experimental hash-based file tools.
-5. Which MCP integrations to enable.
-6. Key values only for integrations that require them.
+4. Which MCP integrations to enable.
+5. Key values only for integrations that require them.
 
 ## Which Paths Are Updated
 
@@ -37,7 +36,6 @@ During setup, the installer overwrites/creates:
 - `~/.opencode/agents/`
 - `~/.opencode/skills/`
 - `~/.opencode/commands/`
-- `~/.opencode/tools/` (if hash tools are enabled)
 - `~/.opencode/keys/`
 - `~/.opencode/opencode.jsonc`
 

doc/en/quick-start.md

@@ -29,7 +29,7 @@ Run it via the `build/dev` agent.
 
 What this command does:
 
-- creates or updates skills only under `.opencode/skills/coder/` in the current repository;
+- creates or updates skills only under `.opencode/skills/` in the current repository;
 - auto-detects languages from code and configs when no arguments are provided;
 - supports an explicit comma-separated language list via `$ARGUMENTS`, for example:
 
@@ -43,6 +43,14 @@ When to run it:
 - after significant stack changes (new language, framework, test runner, linter);
 - when you want to re-sync local coder skills with current repository state.
 
+## Skill-authoring
+
+If you want to create, check, or improve skills themselves, use `/skill-authoring` directly.
+
+You do not need to pick a runtime agent manually. The command routes to a dedicated authoring agent automatically.
+
+For the full workflow, see the [Skill-Authoring Guide](./skill-authoring.md).
+
 ## 4) Use the Recommended Delivery Loop
 
 Short iterations are more reliable than one huge prompt:

doc/en/skill-authoring-prompts.md

@@ -0,0 +1,44 @@
+# Skill-Authoring Prompt Templates
+
+Use these prompts directly in an OpenCode chat with `/skill-authoring`.
+
+For the full flow, artifact layout, static validation, trigger tests, and final `apply / reject / revise: ...` contract, see the [Skill-Authoring Guide](./skill-authoring.md).
+
+Rules:
+
+- Keep the user-facing chat in the user's language.
+- Keep the resulting `SKILL.md` content in English.
+- Show the diff in visible chat before changing the target skill.
+- In full mode, pre-apply artifacts under `ai-docs/skill-authoring/**` are expected.
+- Before `apply`, the real target skill must remain unchanged.
+- The final decision should be one of: `apply`, `reject`, or `revise: ...`.
+
+## 1. Create a New Project Skill
+
+```text
+/skill-authoring create .opencode/skills/<skill-name>. This is a goal-driven change. I want a skill that does the following: <desired behavior>. Use the relevant project files as references, ask any missing questions, propose the first version, show the diff in chat, and end with a final decision prompt. Keep the target skill unchanged until I explicitly type apply. After apply, write the skill and run static validation. Keep the final skill text in English.
+```
+
+## 2. Improve an Existing Project Skill
+
+```text
+/skill-authoring improve .opencode/skills/<skill-name>. This is a <instruction-driven or goal-driven> change. I want the skill to change in this way: <request>. Use relevant project files as context. Show the proposed diff with rationale first, keep the real target skill unchanged until I type apply, and finish the preview with apply / reject / revise: ... . Then apply the approved change and run static validation. Keep the final skill text in English.
+```
+
+## 3. Check an Existing Project Skill
+
+```text
+/skill-authoring check .opencode/skills/<skill-name>/SKILL.md. Validate the current skill, infer and confirm its current meaning only if I did not already describe the intended meaning, point out weak spots, prepare a non-applied proposal when justified, and end the preview with apply / reject / revise: ... .
+```
+
+## 4. Check Against an Explicit Intended Meaning
+
+```text
+/skill-authoring check .opencode/skills/<skill-name>/SKILL.md. The skill should mean: <intended behavior>. Evaluate the current text against that meaning, skip inferred-meaning confirmation, prepare a non-applied proposal if needed, and end with apply / reject / revise: ... .
+```
+
+## 5. Quick Edit
+
+```text
+/skill-authoring improve .opencode/skills/<skill-name> quick. Make only this change: <exact requested change>. Show the diff first, keep the target skill unchanged until I type apply, then apply it. Skip validation and tests.
+```