Generate a validated ReWOO plan DAG from a user request and tool catalog. Use when you need help with rewoo planner.
Scanned 9/8/2026
Install to Claude Code
npx -y skills add anubhavg-icpl/vibe --skill rewoo-planner --agent claude-codeInstalls into .claude/skills of the current project.
Are you the author of Rewoo Planner?
Add the live security badge to your README — it updates automatically with every re-scan.
[](https://www.skillsdirectory.com/skills/anubhavg-icpl-rewoo-planner)More formats (shields.io, HTML) on the badges page.
---
name: rewoo-planner
description: Generate a validated ReWOO plan DAG from a user request and tool catalog. Use when you need help with rewoo planner.
license: CC-BY-NC-SA-4.0
phase: 14
lesson: 02
metadata:
version: 1.0.0
tags: [rewoo, plan-and-execute, planning, dag, distillation]
---
Given a user request and a tool catalog (name, input schema, description), produce a ReWOO plan: a DAG of steps with tool calls and evidence references (`#E1`, `#E2`, ...). Validate the plan before handing it to an executor.
Produce:
1. A plan DAG. Each node has id (`E1`, `E2`, ...), tool name, argument dict (strings may contain `#E<k>` references), and optional `parallel_group` label.
2. Validation output. Acyclicity check via topological sort; reference resolution check (every `#E<k>` has a preceding producer); tool existence check (every tool name is in the catalog); arg schema check (each argument matches the tool's input schema).
3. Parallelism hint. For every topological level, list the nodes that can execute concurrently.
4. Planner/solver split recommendation. If the plan has fewer than 3 steps, recommend ReAct instead. If the plan has an unbounded loop requirement (replanning on every step), recommend Plan-and-Execute with replanner. If the plan exceeds 30 steps or targets web/mobile, recommend Plan-and-Act with synthetic plan data.
Hard rejects:
- Plans with cycles. ReWOO assumes a DAG; cycles are a ReAct or LATS concern.
- Plans that reference `#E<k>` where `k` does not exist yet in the topological order. Emit the specific edge that fails.
- Plans that call tools not in the catalog. Do not invent tools to make a plan work.
- Plans where the argument type for a reference does not match the tool's schema (e.g., `#E1` substitutes a string but the tool expects an int).
Refusal rules:
- If the task is open-ended exploration (unknown tools needed, unknown steps), refuse and recommend ReAct or LATS (Lesson 04).
- If the tool catalog contains destructive tools without a gating approval tool, refuse and point to Lesson 09 (permissions, sandboxing).
Output: a structured plan (JSON or YAML), a validation report, a parallelism map, and a follow-up action pointing to the executor (ReWOO Worker), a replanner (Plan-and-Execute), or a larger trajectory-sampling loop (Plan-and-Act).
End with a "what to read next" note pointing to Lesson 03 (Reflexion) if the task class has been attempted before, or Lesson 04 (LATS) if the plan would benefit from search.
Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.
No comments yet. Be the first to comment!