prp

Planificar una feature compleja antes de implementarla. Genera un PRP (Product Requirements Proposal) con objetivo, comportamiento esperado, modelo de datos, y fases de implementación. Activar SIEMPRE

SKILL.md

GitHub

Skill: Generate a PRP (Product Requirements Proposal)

Generate a PRP for: $ARGUMENTS


What a PRP is

A PRP is the blueprint of a piece of software. It defines WHAT to build before a single line of code.

It is the human-AI contract. The human defines the goal and the why. The agent researches the context, proposes the architecture and generates the phase plan. Together they validate before executing.

Without a PRP: vibe coding, spaghetti code, features that don't fit. With a PRP: clear architecture, defined phases, learnings that persist (self-hardening).


Process

Step 1: Read the base template

~/.claude/PRPs/prp-base.md   (or PRPs/prp-base.md in the Jarvis repo)

Step 2: Interview if context is missing

If $ARGUMENTS lacks detail, ask directly:

  1. Goal: What do you want to build? (end state in 1-2 sentences)
  2. Why: What problem does it solve? What's the value?
  3. Success criteria: How do you know it's done? (3-5 measurable checkboxes)
  4. Constraints: Anything it must NOT do, or a critical dependency?

Don't ask all 4 if the user already gave enough context.

Step 3: Research the codebase

Before writing the PRP:

  • Grep/Glob: find existing code related to the feature
  • Read: read relevant files to understand current patterns
  • Look for similar features already implemented to reuse their patterns

Step 4: Generate the PRP

Create the PRP following prp-base.md:

File name: .claude/PRPs/PRP-XXX-{feature-name}.md — XXX sequential, {feature-name} in kebab-case.

Mandatory content:

  • Goal (1-2 sentences)
  • Why (problem/solution table + value)
  • Success criteria (measurable checkboxes)
  • Expected behaviour (happy path)
  • Context (references, proposed architecture, data model if applicable)
  • Blueprint (ONLY phases, no subtasks — bucle-agentico generates those)
  • Empty sections for Learnings, Gotchas, Anti-patterns

Step 5: Present it to the user

Summarise: goal, number of phases, key architecture decisions — and ask whether to adjust before approving.

Implement NOTHING yet. The PRP must be approved before executing.


After the PRP

Once approved, implementation runs with /bucle-agentico, which uses the PRP as its guide, phase by phase with just-in-time context mapping. For work spanning several sessions, turn the success criteria into a long-run-harness feature list.

Learnings discovered during implementation are written back into the PRP (Learnings section) — active self-hardening.


Rules

  • ALWAYS read prp-base.md before generating
  • NEVER generate subtasks inside the phases (bucle-agentico does that)
  • NEVER implement code in this skill (only the document)
  • ALWAYS research the codebase before proposing architecture
  • The PRP is created in PENDING state until the user approves
  • PRPs live in .claude/PRPs/ inside the current project