Prototype

prototype: spike to learn, then throw it away

Some uncertainty is cheaper to build through than to talk through. A prototype exists to answer ONE named question; the deliverable is the answer, never the code.

Contract (agree before writing code)

1. The question: one sentence, falsifiable. "Can X stream 10k rows

without buffering?" not "explore X".

1. Timebox: default 30 to 60 min of agent work. Hit the box with no answer

→ that IS a finding (question too big; split it or grill further).

1. Kill criteria: what result answers the question either way.
2. Where it lives: an authorised, manifest-owned scratch dir outside

production source. Never import it from real code or touch external systems, credentials or user data without separate authority.

Rules

• Build only the checks needed to answer the question. Production quality gates

do not apply to disposable scratch code; evidence, authority and containment do.

• Cheat aggressively: hardcode data, stub auth, fake the network. Only

the question's variable needs to be real.

• One question per spike. A second question mid-spike → note it, finish,

spike it separately if it still matters.

• Harvest, then retire. Put the finding and evidence in the owning spec,

decision record or scoping thread when artifact authority exists. Remove only paths proven run-owned and explicitly authorised; otherwise quarantine or hand them back with a named cleanup action.

• Prototype code never "graduates" to production. Productionisation routes

through scope when decisions changed, then implement; rebuild the smallest production solution with the appropriate test/refactor method.

Harvest format (into the owning doc/thread)

**Spike:**   (YYYY-MM-DD, timebox 45m)
**Answer:** yes/no/partial — 
**Evidence:** 
**Implication:** 

Red flags

• "The spike works, let's ship it" → scope any changed decision, then use

implement; do not wire the spike into production.

• Spike touching production files or secrets → wrong dir, stop.
• No timebox agreed → not a spike, just unscoped wandering.
• Answer known from docs/a search in <10 min → didn't need a spike.

Source & license

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

• Author: mblauberg
• Source: mblauberg/provenant
• License: MIT

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