Assumption Check

skill: assumption-check

purpose

List the assumptions baked into a plan before executing it. Surfaces hidden guesses that could make the whole thing wrong. Blocks high-risk ones before they cascade.

whentouse

• Before executing a multi-step plan where one wrong assumption cascades
• Acting on behalf of a user without full context (delegated tasks, automated jobs)
• Task involves a domain you haven't verified facts about

whennotto_use

• Task fully specified with verified facts — no inference needed
• Single-step tasks with no dependencies
• You've already confirmed assumptions with the user this session

input

plan: string      # The plan or action you're about to execute
context: string   # What was explicitly stated vs what you inferred

output

{
  "assumptions": [
    {
      "statement": "string",
      "risk": "low" | "medium" | "high",
      "source": "inferred" | "stated",
      "verify_by": "string | null"
    }
  ],
  "safe_to_proceed": true | false,
  "blocker": "string | null"
}

instructions

Read the plan. Find every place you are acting on something assumed rather than stated.

State each assumption as a plain fact sentence: "The user wants X", "The file format is Y."

Risk levels:
- high: if wrong, output is useless or harmful
- medium: if wrong, requires significant rework  
- low: if wrong, minor correction needed

For high-risk assumptions: state exactly how to verify before proceeding.

safe_to_proceed is false if any high-risk unverified assumption exists.
blocker is the single most dangerous unverified assumption.

Return JSON only. Max 6 assumptions — if you find more than 6, include only the 6 highest-risk ones and set safe_to_proceed: false with blocker: "Task has too many unverified assumptions — scope it before proceeding."

constraints

• Max output tokens: 300
• Max 6 assumptions — more means the task itself needs clarification
• Only material assumptions — not obvious facts

example

Input:

plan: "Refactor auth module to use JWT, update middleware, deploy to production."
context: "User said: 'We should modernize our auth.' No other details."

Output:

{
  "assumptions": [
    {
      "statement": "JWT is the desired replacement (not OAuth, API keys, or another approach)",
      "risk": "high",
      "source": "inferred",
      "verify_by": "Ask: which auth method do you want to migrate to?"
    },
    {
      "statement": "Production deployment is in scope for this task",
      "risk": "high",
      "source": "inferred",
      "verify_by": "Confirm: refactor only, or also deploy?"
    },
    {
      "statement": "Session-based auth is the only current auth system",
      "risk": "medium",
      "source": "inferred",
      "verify_by": "Check codebase for other auth paths before modifying middleware"
    }
  ],
  "safe_to_proceed": false,
  "blocker": "JWT is the desired replacement — wrong choice invalidates the entire plan"
}

feedback_log

2026-05-29 | plan with 8+ interdependent assumptions | no output behavior defined for >6 — skill was silent | added: output top 6 by risk, set safetoproceed:false with explicit blocker

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: 0-uddeshya-0
• Source: 0-uddeshya-0/agent-skills
• License: MIT

Install and usage instructions live in the source repository linked above.