assumption-surface

Force assumptions into the open. Do not start work while high-blast gaps remain.

Trigger on

• "before you start"
• "what would you assume"
• "what am I missing"
• "think this through first"
• "is this clear enough to build"
• Any request where scope, format, data source, or target is not specified

Blast radius

Common assumption categories

Process

1. Read the request. List every decision the implementer must make that was not stated.
2. Map each to a category above.
3. For each, note the default most LLMs would assume silently.
4. Rank by blast radius.
5. Extract questions for high + medium. Low proceeds with a noted default.
6. If no answer comes, generate 2–3 candidate interpretations ranked by likelihood.

Output

request: {one-line summary}

assumptions (high blast):
- {category}: {assumption} → default: {typical guess} → ask: {question}
- ...

assumptions (medium):
- {category}: {assumption} → default: {guess} → ask: {question}

assumptions (low — proceeding with defaults):
- {category}: {default}

minimum questions to proceed:
1. {question}
2. {question}

interpretations if no answer comes:
A. {reading 1} → implies {concrete choice chain}
B. {reading 2} → implies {concrete choice chain}
C. (optional) {reading 3}

recommended path: {A | B | C | hold for answers}
reasoning: {one line on why this path}

Worked example

Input: "add a way to export users"

request: export users functionality

assumptions (high blast):
- scope: all users vs filtered subset → default: all → ask: scope?
- delivery: file download vs API response → default: API → ask: how delivered?
- fields: which columns → default: id+name+email → ask: which fields (privacy)?

assumptions (medium):
- format: JSON vs CSV → default: JSON → ask: format?
- volume: thousands vs millions → default: thousands → ask: scale?

assumptions (low — proceeding with defaults):
- persistence: no caching needed
- auth: admin-only endpoint

minimum questions to proceed:
1. Scope + field set (privacy-critical)
2. Delivery mechanism
3. Expected volume (affects pagination / streaming choice)

interpretations if no answer:
A. Admin API endpoint → paginated JSON, id+email+name, admin-auth
B. One-off data dump → CSV file, all fields, run via script

recommended path: A
reasoning: API is safer default — paginated, auth-bound, does not need ops support

Hard rules

• List assumptions before writing code, prose, or structure.
• If zero high-blast assumptions exist, say "no hidden assumptions" and proceed.
• Never exceed 5 minimum questions. More = request is too broad, recommend decomposition.
• Never ask about style preferences unless the task is style-bound (writing, naming, UI).
• Defaults come from project context first, general conventions second.
• If the request is one sentence and hits 3+ high-blast assumptions, flag "request underspecified" at the top.
• Never proceed on low-blast defaults if user explicitly asked you to enumerate everything.