Nw Spike Methodology Agent Skill

Spike Methodology

What a Spike IS

Throwaway code that validates exactly one assumption. Max 1 hour. Binary outcome: works or doesn't work. Code is deleted after validation; only findings persist.

What a Spike IS NOT

| Not This | Why |

|----------|-----|

| Prototype | Prototypes evolve into production code. Spikes are deleted. |

| MVP | MVPs ship to users. Spikes never leave the developer's machine. |

| Walking Skeleton | Skeletons wire end-to-end architecture. Spikes test one mechanism. |

| POC | POCs demonstrate feasibility to stakeholders. Spikes answer a developer question. |

The 4 Questions

Every spike answers exactly these:

• Does the core mechanism work? Can we parse the output? Call the API? Does the algorithm produce correct results?
• Does it meet the performance requirement? Is it under the time/memory/throughput budget?
• What edge cases exist? Empty input, malformed data, missing dependencies, concurrent access.
• What did we assume wrong? The design said X, but reality is Y. Document the delta.

When to Spike

• New mechanism never tried before in this codebase
• Performance requirement that cannot be validated by reasoning alone
• External integration (API, library, OS feature) with unknown behavior
• Algorithm where correctness is non-obvious

When to Skip

• Pure refactoring (mechanism already proven)
• Bug fix (behavior already understood)
• Feature estimated at less than 1 day (spike overhead not justified)
• Mechanism already validated by a prior spike or existing code

Rules

• Max 1 hour wall clock. If it takes longer, the problem is bigger than expected. Stop and escalate.
• No tests. No types. No error handling. No ports. No abstractions.
• Code lives in /tmp/spike_{feature_id}/ or a scratch directory. Never in src/.
• One file preferred. Two files maximum.
• Use time.perf_counter() for timing, not time.time().
• Print results to stdout. No logging frameworks.
• After validation: delete the code, keep the findings.

Spike Code Structure

#!/usr/bin/env python3
"""SPIKE: {feature} -- {one-line question being validated}

Throwaway code. Not production. Will be deleted after validation.
"""
import time

# 1. Setup (minimal, no frameworks)
# ...

# 2. Core mechanism attempt
start = time.perf_counter()
# ... the thing you're testing ...
elapsed = time.perf_counter() - start

# 3. Print findings
print(f"Mechanism: {'WORKS' if success else 'FAILS'}")
print(f"Timing: {elapsed*1000:.1f}ms (budget: {budget}ms)")
print(f"Edge cases: {edge_cases}")

Findings Output Format

Output goes to docs/feature/{feature-id}/spike/findings.md:

# Spike Findings -- {feature-id}

## Verdict: WORKS / DOESN'T WORK

## Question tested
{the one assumption being validated}

## Core mechanism
- Tested: {what was tested}
- Result: {what happened}

## Timing
- {operation}: {time}ms
- Total: {time}ms (budget: {budget}ms) -- PASS / FAIL

## Edge cases discovered
1. {edge case}: {what happened}
2. {edge case}: {what happened}

## Design implications
- {assumption that was wrong}: {correct reality}
- {approach the spike validated or invalidated}

## Spike code
Deleted. Was at /tmp/spike_{feature_id}/

Escalation

If the spike reveals the problem is fundamentally different from what was assumed:

• Stop the spike
• Write findings with verdict "DOESN'T WORK" or "BIGGER THAN EXPECTED"
• Return to DISCUSS to re-scope the feature
• The spike findings become input to the revised DISCUSS wave