| name | assumption-excavator |
| description | Find unstated assumptions in any plan, design, or argument. Test what breaks if each assumption is false. Use when reviewing a design, planning a project, or reading a paper. |
| tags | ["novelty","thinking","analysis"] |
| difficulty | intermediate |
| argument-hint | The plan, design, or argument to excavate |
Assumption Excavator — Find What Everyone Assumed
Every failed project had a critical assumption that nobody checked. You are the tool that finds them before they fail.
Protocol
Step 1: Surface Every Assumption
Read the input and list every assumption made — both stated and unstated. Categorize:
| Category | Examples |
|---|
| Resource assumptions | "We have enough GPUs." "The API stays free." |
| Behavioral assumptions | "Users will read the instructions." "Engineers will write tests." |
| Environmental assumptions | "The network is reliable." "The database is always available." |
| Temporal assumptions | "Nothing changes while we build." "The market stays the same." |
| Causal assumptions | "X causes Y." "If we build it, they will come." |
| Scalability assumptions | "Works at 10 users = works at 10M." "Linear cost growth." |
| Compositional assumptions | "Components that work independently work together." |
Step 2: Test Each Assumption
For each assumption, ask:
- What would break if this assumption is false?
- How would we know it's false before it's too late?
- What's the cheapest way to validate this assumption?
Step 3: Identify the Critical Few
Most assumptions are safe. 1-3 are dangerous. Flag the ones where:
- Falsehood would be catastrophic
- Falsehood is plausible
- We have no way to detect falsehood early
Step 4: Design a Test
For each critical assumption, specify the minimum experiment that could falsify it.
Example Output
Input: "We're building a code generation agent. We'll sell it to enterprise engineering teams for $50/seat/month. They hook it into their GitHub and it generates PRs automatically."
Assumptions surfaced:
- Engineering teams have authority to buy tools without security review — Critical
- Developers trust AI-generated PRs — Critical
- GitHub API rate limits allow automated PR generation at scale
- Enterprise codebases are clean enough that AI-generated code integrates cleanly
- Security teams will approve a tool that writes code
- $50/seat/month is below the "expenseable without VP approval" threshold
- The tool works equally well across codebases in different languages
- Generated code doesn't introduce security vulnerabilities
- PR review costs are lower than the code-writing costs being saved
- Teams want more PRs, not fewer
Critical assumptions:
#1 (buying authority): If VP approval is required, the sales cycle goes from 1 week to 9 months. Most security teams block AI code generation tools.
Test: Survey 20 engineering leaders at target companies. Ask: "Could you buy this with a company card or does it need VP+ approval?"
#2 (trust): If developers spend as much time reviewing AI-generated PRs as writing code themselves, the value proposition collapses.
Test: Run a pilot with 5 engineers. Measure time-to-merge for AI-generated vs human-written PRs. If review time > 0.5x writing time, the value prop is weak.
#5 (security approval): If security teams block the tool, enterprise sales are zero regardless of engineering interest.
Test: Ask 3 enterprise security engineers: "What would you need to approve an AI that writes production code?" If the answer includes things you can't provide (e.g., SOC2, on-prem deployment, human-in-the-loop for every PR), adjust the product.
Anti-Patterns
| Mistake | Why it fails | Fix |
|---|
| Only surfacing obvious assumptions | Stated assumptions are usually fine | Look for the ones nobody says aloud |
| Not differentiating critical from trivial | Everything is an assumption | Flag the 1-3 that would kill the project |
| Not suggesting a test | Assumptions are only dangerous if untested | Always propose the minimum falsification experiment |
PRISM Integration
In PRISM mode, output findings as structured YAML:
pattern: assumption-excavator
input: "<original plan>"
findings:
- claim: "<hidden assumption>"
type: assumption
category: <resource | behavioral | environmental | temporal | causal | scalability | compositional>
catastrophic_if_false: <true | false>
cheapest_test: "<minimum falsification experiment>"
confidence: <HIGH | MEDIUM | LOW | EXPLORATION>
Consumed by: contrarian (invert critical assumptions), pragmatist (cost of false assumptions)
Consumes from: counterfactual (surface assumptions in suppressed alternatives)
Trigger Conditions
Use this skill when:
- Reviewing a project plan
- Evaluating a proposal
- The user says "it should work because..."
- Before committing to a significant decision
- A project is behind schedule (something was assumed)
