Echobot Skill Authoring

EchoBot Skill Authoring

Create or revise repository-local skills that EchoBot can discover and activate.

Authoring goals

• Put project-specific skills under skills/<skill-name>/.
• Use lowercase kebab-case for the folder name and the name frontmatter field.
• Keep frontmatter minimal. EchoBot routing reads name and description from SKILL.md.
• Keep name as a single-line value. description may be a single line or a YAML block scalar with > or |.
• Write description as the trigger surface: what the skill does, when to use it, and what kinds of user requests or contexts should activate it. Include likely request phrasing when that improves triggering.
• Keep the main body procedural and concise. Move deeper material into small files under references/ or scripts/.
• Design for lazy loading. After activation, EchoBot exposes the body plus a resource summary. It does not auto-load bundled files.
• Only put text resources that an agent may need to read into references/, scripts/, or agents/. read_skill_resource only reads UTF-8 text files.
• Use assets/ for templates or binary output artifacts, not for essential instructions.
• Do not treat agents/openai.yaml as required. Current EchoBot discovery ignores it.
• Avoid extra docs like README.md or changelog files inside a skill unless the runtime truly needs them.

Frontmatter checklist

• Keep name stable and specific. Renaming a skill changes explicit /skill-name and $skill-name activation.
• Make description concrete. Good descriptions say both the task and the trigger context.
• Do not put "when to use" guidance only in the body. The body is loaded after activation.
• EchoBot currently ignores extra frontmatter keys. The validator is more permissive for compatibility, but project-local skills should usually keep only name and description.

Resource layout

skills/<skill-name>/
|-- SKILL.md
|-- references/   # short, focused UTF-8 docs loaded on demand
|-- scripts/      # executable helpers or deterministic workflows
|-- assets/       # templates or binary resources not meant for context loading
`-- agents/       # optional helper prompts or agent-specific text resources

You do not need every folder. Create only what the skill actually uses.

Practical workflow

1. Choose a stable kebab-case skill name and create or update skills/<skill-name>/SKILL.md.
2. Draft or refine the trigger description in SKILL.md before writing body details.
3. Keep the main body short and procedural. Tell the agent which resource to inspect first when the skill has multiple files.
4. Split long or variant-specific details into focused resource files.
5. Validate the skill with python -X utf8 echobot/skills/skill-creator/scripts/quick_validate.py skills/<skill-name>.
6. If the skill changes runtime expectations, update tests.

EchoBot-specific notes

• Project skills override built-in skills with the same name.
• Explicit /skill-name or $skill-name activates the skill immediately.
• Otherwise the model may call activate_skill.
• Activation loads the skill body and a resource summary, not the bundled file contents.
• After activation, the agent must call list_skill_resources and read_skill_resource to inspect bundled files.
• If a skill depends on bundled files, make the body tell the agent which file to open first and why.
• If you change discovery, parsing, or activation behavior in echobot/skill_support/, run python -m unittest tests.test_skill_support tests.test_chat_agent tests.test_agent -v.

Read references/runtime.md before changing the skill runtime or designing a skill with many bundled files.