RebundleRebundle Stack
Repositorio

Skill · Proceso · /ama

AMA: Ask Me Anything

Modo entrevista: antes de ponerse a trabajar, Claude te hace rondas de preguntas hasta cubrirlo todo y guarda un brief con lo decidido.

Instálala

Guarda el fichero de abajo en tu carpeta de skills y Claude Code lo carga en la siguiente sesión.

# guárdalo como
~/.claude/skills/ama/SKILL.md

# y llámalo con
/ama

El fichero

SKILL.mdFrontmatter
---
name: ama
description: Deep-dive interview mode for any task. Conducts adaptive rounds of focused questions using AskUserQuestion to eliminate assumptions and gaps before writing code, plans, text, or any deliverable. Use when the user invokes /ama alongside a task they want explored in depth before execution.
---
Instrucciones

AMA: Ask Me Anything Interview Mode

When invoked via /ama, conduct a thorough multi-round interview before doing any actual work. The user has already described a task in the conversation. Your job is to deeply understand their intent, constraints, preferences, and edge cases so the final output is exactly right on the first attempt.

How It Works

  1. Map what's already known. Read the conversation history, any referenced files or materials, and the project's CLAUDE.md, so you don't ask questions whose answers you already have.
  2. Run a blindspot pass. Now invert step 1: inspect the artifact the task actually touches (the codebase, the existing doc, the dataset, the design) and surface the unknown unknowns -- decisions the user never specified, hidden constraints, integration points, relevant prior art, anything that would reshape the approach if they knew about it. Why: the user can only raise questions about what they already know to worry about, so an interview built on their prompt alone inherits their blind spots. This pass is what lets the questions reach the decisions neither of you had flagged, which is the depth they invoked /ama for. Feed what you surface into the interview.
  3. Interview using AskUserQuestion. Group questions in one call only when they're genuinely independent (one answer wouldn't change how you'd ask another); when answers are coupled, ask one at a time so each reply can reshape the next question. Batching is an efficiency, not the goal -- over-batching coupled questions spends the adaptivity /ama exists to provide. Build on previous answers, following up on interesting or ambiguous ones rather than jumping to a new topic.
  4. Cover these dimensions (skip any that are clearly irrelevant to the task):
    • Goal clarity: What does "done" look like? What would make this a 10/10?
    • Technical choices: Architecture, stack, dependencies, integration points
    • UX/UI: Layout, interaction patterns, responsive behavior, accessibility
    • Audience: Who sees this? What do they expect?
    • Tone & style: Formal/casual, brand voice, visual identity
    • Scope boundaries: What's explicitly out of scope? What's a "nice to have" vs "must have"?
    • Tradeoffs: Speed vs quality, simplicity vs flexibility, cost vs features
    • Edge cases: What happens when things go wrong? Empty states, error handling
    • Safe defaults & one-way doors: When execution later hits a decision the interview didn't settle, what's the conservative default you'd want picked -- and which actions are irreversible (deploys, migrations, deletes, sending things externally) where the agent should stop and ask instead of defaulting? Capturing this here means the executing agent starts with your risk tolerance instead of guessing it mid-run.
    • Context: Why now? What prompted this? What was tried before?
    • Output format: File type, structure, length, delivery method
  5. Avoid obvious questions. Don't ask things you can infer from context, CLAUDE.md preferences, or the codebase. The user chose /ama because they want depth, not a checkbox survey. Surprise them with questions they hadn't considered -- most come straight from the blindspot pass in step 2.
  6. Adapt depth to complexity. Simple tasks might need 2 rounds. Complex, ambiguous, or high-stakes tasks might need 6+. Read the room -- if the user gives short, confident answers, the task is probably clear to them and you can wrap up sooner. If answers reveal new complexity, keep digging.

After the Interview

When you've gathered enough context:

  1. Present a brief summary of the key decisions and your understanding of the task. Wait for the user to confirm before moving on. Do NOT skip this step or combine it with saving the brief.
  2. Once the user confirms, save a brief file before starting work. Save to briefs/ in the project root (create the folder if needed). Filename: YYYY-MM-DD-[topic-slug].md. Use this template:
# Brief: [topic]
Fecha: YYYY-MM-DD
Plan: [path to plan file if one was created, or "N/A"]

## Contexto
[What needs to be done and why]

## Resumen del AMA
[Key questions and answers/decisions from the interview, organized by theme
 in whatever structure best represents what was discussed]

## Notas adicionales
[Anything else relevant that came up: references, discarded ideas,
 identified risks, inspiration, links, extra context that doesn't
 fit above but is worth preserving]

## Deviations
[Empty at handoff. During execution, when you hit a decision this brief
 didn't settle, pick the conservative option, append it here (what came up,
 what you chose, why it was the safe call), and keep going.]

## Siguiente paso
[What will be done with this information]

Adapt the language of the brief to match the language used during the AMA (Spanish if the conversation was in Spanish, English if in English, etc.). Keep the ## Deviations heading as-is regardless of language, so the executing agent can always find the log.

  1. Tell the user where the brief was saved (full path). This is the last checkpoint -- once confirmed, transition directly into the work. As you work, honor the safe defaults captured above: if you hit a decision the brief didn't settle, pick the conservative option, log it under the brief's ## Deviations section (what came up, what you chose, why), and keep going. Why: a long run shouldn't stall on every unspecified fork, but silent divergence from what was agreed is worse -- the log keeps momentum while leaving the user a reviewable trail of every judgment call made on their behalf.

Key Principles

  • Use AskUserQuestion for every question (never dump questions as plain text)
  • Batch only genuinely independent questions in one call; ask coupled ones and follow-ups one at a time, so each answer can reshape what you ask next
  • Prioritize questions where the user's answer would change the architecture or approach -- if two possible answers lead to the same plan, the question isn't worth asking
  • If the user gives a short answer, that dimension probably isn't important to them -- move on
  • The interview should feel like a conversation with a thoughtful collaborator, not a form

Rebundle · Stack de IA curado