a554b554/skills/execute

Execute

Route all <@skill-name: ...> annotations in a file to their matching skill subagents.

Usage

/execute <filename>

Grammar

<@skill-name: prompt>

Tags are written inline, directly in the original text. <@output: ...> is a special data tag — never dispatched, only consumed by other skills.

Modifiers

Prompts may contain context references ( path ) that point the skill agent to external resources (files, sections, etc.). The execute router does not interpret these — they are passed through to the subagent as part of the prompt.

Skill registry

| Tag | Skill | Description | |---|---|---| | <@edit: ...> | edit | Generic edit per prompt. | | <@proofread> | proofread | Fix grammar, spelling, punctuation. | | <@plan: ...> | plan | Produce a revision plan. | | <@ph> / <@ph: ...> | ph | Fill placeholder from context. | | <@resolve> | resolve | Apply preceding plan/output chain. | | <@figure: ...> | figure | Generate a figure. (Placeholder — needs Gemini API.) | | <@cite: ...> | cite | Find and insert citations to support a claim. | | <@plot: ...> | plot | Generate and execute a data visualization script. |

To add a new skill: create skills/<name>/SKILL.md and add a row to this table.

Subagent contract

Every subagent receives:

1. File path — the full path of the file being processed.
2. Surrounding text — the paragraph or block where the tag appears (including any chained tags like <@output: ...>).
3. Prompt — the text inside the tag after the skill name.

Each subagent is responsible for applying its own changes directly to the file. The execute skill does not write back results — it only dispatches.

Workflow

1. Clear context — disregard all prior conversation context. Start fresh with only this skill's instructions.
2. Read the target file.
3. Scan for all <@...> tags, top to bottom.
4. For each tag:
   • Skip <@output: ...> — it is data, not a dispatch target.
   • Look up the skill name in the registry above.
   • If unknown, skip and warn the user.
   • Spawn a subagent with the matching skill. Pass the file path, surrounding text, and prompt.
   • The subagent applies changes directly to the file.
   • Re-read the file before processing the next tag (the file may have changed).
5. Report a brief summary of how many tags were processed.

Rules

• Process tags sequentially, top to bottom.
• Skip <@output: ...> tags.
• Re-read the file after each subagent completes.
• Unknown skill: skip and warn.
• Do not stop until all tags have been processed.
• Do not touch anything else that is not from registered skill. Sometime author leave inline comment like [comment:], that is not for you.
• Sometimes author use <@xx> for demonstration purpose, not ask you to execute it. You need to read the context carefully to avoid false trigger.