What is Verbalized Sampling (VS)?

VS is a protocol to generate diverse intent hypotheses ranked by likelihood, with an explicit actor-critic review documenting weaknesses, contradictions, and oversights for each sample.

How many samples should I generate?

Baseline is N >= 5, but you should scale N based on ambiguity and risk (higher for complex tasks; architectural contexts may require N >= 10; there is no hard cap).

When do I start asking clarifying questions?

Only after completing both VS and the exploratory edge-case steps; then ask the maximum possible number of precise, challenging planning questions that address uncertainty and gaps.

askme

npx machina-cli add skill OutlineDriven/odin-claude-plugin/askme --openclaw

Files (1)

SKILL.md

1.9 KB

Ask Me Command

Before proceeding to ask planning questions, you must proactively and critically execute both Verbalized Sampling (VS) and exploration:

For Verbalized Sampling, generate and sample at least N distinct, diverse candidates that represent different possible user intents or directions, ranked by likelihood, where N is dynamic by ambiguity/risk/scope (baseline N>=5; trivial N>=3; high ambiguity/risk N>=7; architectural N>=10; no hard cap). Run actor-critic on each VS sample: explicitly record one weakness, contradiction, and oversight before selecting a direction. VS prevents over-engineering by surfacing simpler alternatives; expand only while new samples materially change planning decisions, and prefer the smallest sufficient N.

Required VS Output Format:

1. [Most likely] hypothesis here
   - Weakness: [potential flaw]
   - Contradiction: [logical conflict if any]
   - Oversight: [what this misses]

2. [Alternative] hypothesis here
   ...

For exploration, deliberately seek out unconventional, underexplored, and edge-case possibilities relating to the user's objective, drawing on both the provided context and plausible but non-obvious requirements. Include at least 3 edge cases (at least 5 if architectural), and stop expanding once additional cases no longer change decisions.

Only after completing both critical VS and exploration steps, proceed to use the question tool to ask the maximum possible number of precise, clarifying, and challenging planning questions that holistically address the problem space, taking into account uncertainty, gaps, and ambiguous requirements.

Source

git clone https://github.com/OutlineDriven/odin-claude-plugin/blob/main/skills/askme/SKILL.mdView on GitHub

Overview

AskMe employs Verbalized Sampling (VS) to explore diverse user intents before planning. It generates multiple candidate directions, evaluates them with an actor-critic lens (weakness, contradiction, oversight), and then conducts edge-case exploration to surface non-obvious requirements before asking the maximum number of clarifying questions.

How This Skill Works

First, generate at least N diverse hypotheses representing possible user intents, and rank them by likelihood. Next, run an actor-critic evaluation on each hypothesis to document weakness, contradiction, and oversight. Then perform deliberate exploration to surface unconventional edge cases, stopping when new cases no longer change decisions. Only after both steps are complete do you deploy the question tool to ask the maximum possible number of precise clarifying questions.

When to Use It

When a task starts with ambiguity or unclear goals
When there are multiple valid interpretations of user intent
When planning runs the risk of over-engineering and you need simpler alternatives
When high-stakes or architectural decisions require exhaustive clarification
When you need to surface edge-case intents and non-obvious requirements before committing

Quick Start

Step 1: Identify ambiguity and set a baseline N for VS (adjust by risk/complexity)
Step 2: Generate N diverse hypotheses, rank by likelihood, and run actor-critic for each
Step 3: Do edge-case exploration, then use the question tool to ask the maximum number of clarifying questions

Best Practices

Always start with VS: generate at least N diverse hypotheses and rank by likelihood
For each VS sample, perform an actor-critic evaluation and record Weakness, Contradiction, and Oversight
Keep N dynamic: baseline N>=5; adjust for ambiguity; avoid expanding if decisions don’t change with more samples
Consciously perform exploration to surface unconventional edge cases (min 3; 5+ if architectural) and stop when decisions stabilize
Proceed to asking clarifying questions only after completing both VS and exploration steps

Example Use Cases

Clarifying a vague feature request by exploring intents such as UX improvement, accessibility constraints, and performance targets
Disambiguating a multi-interpretation automation task by considering different workflow directions and failure modes
Planning a research task with unclear success criteria across stakeholders and data requirements
Prioritizing a project backlog where conflicting stakeholder goals must be reconciled through explicit intents
Developing a data pipeline fix where possible failure modes are ambiguous and edge cases must be anticipated

Frequently Asked Questions

Add this skill to your agents