Before a task, say in at most 5 lines which model and effort to use and how to phrase the prompt so it needs fewer back-and-forths. Use for "/coach", "coach", "qué modelo uso para esto", "cómo le pido esto", "which model should I use", "how should I phrase this", or before starting a task when unsure how to ask.
Installs into .claude/skills of the current project.
Sponsors
SSkills DirectorySkills Directory
Know which skills are safe — weekly.
SSkills DirectorySkills Directory
Is that skill safe to install?
Are you the author of Coach?
Add the live security badge to your README. It updates with every re-scan.
[](https://www.skillsdirectory.com/skills/sebabottero-coach)
---
name: coach
description: Before a task, say in at most 5 lines which model and effort to use and how to phrase the prompt so it needs fewer back-and-forths. Use for "/coach", "coach", "qué modelo uso para esto", "cómo le pido esto", "which model should I use", "how should I phrase this", or before starting a task when unsure how to ask.
---
# Coach: before the task
Goal: fewer back-and-forths. Short and scannable, in the user's language: no paragraphs, each bullet one line. No theory, no survey.
1. **Classify** the task from what the user wrote (and the repo, if it is obvious): question, small edit, new feature, bug, big ticket, setup change.
2. **Model and effort** (cheapest that does the job dependably):
- question, small edit, search, summary → fast model, low effort;
- feature or refactor with a known plan → default model, medium effort;
- hard bug diagnosis, architecture, ambiguous scope → strongest model, high effort, only for that step.
3. **Prompt**: rewrite the user's request with the four things that avoid rework: goal · success criteria · what is out of scope · files or commands to start from. Mark with `?` what is missing instead of inventing it.
4. **Process hint**, one line only if it applies: vague → `grillme`; big ticket → `ticket-flow`; bug → `systematic-debugging`; long unrelated context → `/clear` first; large PDF/docx/pptx/xlsx referenced or attached → `convert-to-markdown` first.
Output format (receipt style: bold labels, one short bullet per item, blank line between blocks, prompt in a code block to copy):
```
**Tarea**: <type>
**Modelo**: <model> · esfuerzo <level>
- <why, 5 words>
**Prompt** (copiar):
> <improved request, split into Objetivo / Hecho cuando / Fuera de alcance / Partir de, one per line>
**Falta** (responde y listo):
- <missing item as a question with the recommended default>
- <max 3 items, or "nada">
**Proceso**: <one hint, only if it applies>
```
When the user answers **Falta**, regenerate on your own, without being asked: reply with only the updated **Prompt** block (answers folded in, no `?` left) plus **Modelo** if the answers change the size of the task, and **Falta** only if something is still open. Do not repeat Tarea or Proceso.
Do not start the task. Do not read files unless one is needed to classify. Wait for the user to send the prompt.
