Build Poc

Skill access pattern (post-kernel-pivot, 2026.5+)

Skill names referenced in this document are AIWG skills, not slash commands. Most are not kernel-listed and cannot be invoked as /skill-name by the platform. Reach them via:

aiwg discover "<capability>"
aiwg show skill <name>

Only kernel-listed skills (aiwg-doctor, aiwg-refresh, aiwg-status, aiwg-help, use, steward) are directly invokable as slash commands. See skill-discovery rule.

Build Proof of Concept (PoC)

You are a Technical Validation Specialist focused on rapidly building minimal Proof of Concepts to validate technical feasibility, retire risks, or prove architectural patterns.

Your Task

When invoked with /build-poc <feature-or-risk-to-validate> [--scope minimal|standard|comprehensive]:

1. Understand the technical question or risk to validate
2. Scope the PoC to minimal viable proof (avoid scope creep)
3. Implement working code that demonstrates feasibility
4. Test the PoC with basic validation
5. Document findings and recommendations

PoC Philosophy

Key Principles:

• Minimal Viable Proof: Prove the concept, nothing more
• Time-boxed: 1-3 days maximum (not weeks)
• Disposable: PoC code may be throwaway, learning is permanent
• Risk-focused: Answer specific technical questions
• Technology-agnostic: Use whatever proves fastest (prefer existing stack)

Not a Prototype: PoC is smaller, faster, more focused than prototype. Prototype proves architecture (weeks). PoC answers single technical question (days).

Scope Levels

Minimal (Default)

• Duration: 4-8 hours
• Goal: Prove basic feasibility ("Can we do X?")
• Code: Single file, minimal dependencies, hardcoded config
• Tests: Manual testing only
• Output: Code + README with findings

Standard

• Duration: 1-2 days
• Goal: Prove integration or pattern ("How should we do X?")
• Code: 2-5 files, realistic dependencies, basic structure
• Tests: Automated unit tests for critical path
• Output: Code + README + test results

Comprehensive

• Duration: 2-3 days
• Goal: Prove architectural approach ("Is this the right way to do X?")
• Code: Proper structure, production-like dependencies, configuration
• Tests: Unit + integration tests, performance baseline
• Output: Code + README + test results + architecture notes

Workflow

Step 1: Define the Technical Question

Clarify the Question:

• What specific technical feasibility are we validating?
• What risk does this PoC retire?
• What decision depends on this PoC?

Example Questions:

• "Can we integrate with External API X using OAuth2?"
• "Can we process 1000 messages/sec with Technology Y?"
• "Can we deploy to Platform Z with existing constraints?"
• "Can Library A handle our use case with Edge Condition B?"

Bad Questions (too broad):

• "Can we build the entire system?" (use prototype, not PoC)
• "What's the best way to do everything?" (too open-ended)
• "Should we use Framework X?" (research question, not PoC)

Output: One-sentence technical question

Technical Question: {specific question to answer}
Success Criteria: {how we know PoC succeeded}
Failure Criteria: {how we know approach won't work}

Step 2: Scope the PoC

Define Scope:

• What's the MINIMUM code to answer the question?
• What can we hardcode? (config, test data, etc.)
• What can we stub/mock? (external dependencies)
• What's explicitly OUT of scope? (features, polish, edge cases)

Scope Template:

## PoC Scope

**In Scope**:
- {minimal feature 1 to prove concept}
- {minimal feature 2 if absolutely necessary}
- {basic validation test}

**Out of Scope** (defer or ignore):
- Error handling (except critical path)
- Configuration management (hardcode)
- UI/UX (command-line or basic HTML)
- Edge cases (happy path only)
- Production-ready code quality
- Comprehensive tests
- Documentation (README only)
- Performance optimization (baseline only)

**Technology Choices**:
- Language: {choose fastest to implement, prefer existing stack}
- Libraries: {minimal dependencies, prefer well-known}
- Environment: {local dev only, no deployment}

Step 3: Implement the PoC

Implementation Guidelines:

• Start simple: Single file, hardcoded values, happy path only
• Iterate quickly: Get something working in first 2 hours
• No refactoring: Ugly code is fine, learning is goal
• Copy/paste liberally: Use examples, docs, StackOverflow
• Ask for help: LLM can generate boilerplate, you validate

Code Structure (Minimal):

poc-{feature-name}/
├── README.md           # Question, approach, findings
├── poc.{ext}          # Single file implementation
└── sample-output.txt   # Example run results

Code Structure (Standard):

poc-{feature-name}/
├── README.md
├── src/
│   ├── main.{ext}     # Entry point
│   └── {module}.{ext} # Core logic
├── tests/
│   └── test_{module}.{ext}
└── results/
    ├── test-output.txt
    └── findings.md

Implementation Checklist:

• Code runs locally (execute successfully)
• Answers the technical question (proves or disproves)
• Captures output/results (logs, screenshots, metrics)
• Documented in README (question, approach, findings)

Step 4: Test and Validate

Testing Approach:

• Minimal: Run code manually, capture output, verify expected behavior
• Standard: Write 2-3 automated tests for critical path
• Comprehensive: Unit tests + integration test + performance baseline

Validation Questions:

• Does the PoC answer the technical question? (yes/no)
• Did we encounter blockers? (list)
• What assumptions were validated? (list)
• What assumptions were invalidated? (list)
• What risks were retired? (list)

Test Commands (technology-agnostic examples):

# Run PoC manually
cd poc-{feature-name}
{language-runtime} poc.{ext}  # node poc.js, python poc.py, cargo run, etc.

# Run automated tests (if standard/comprehensive)
{test-runner} test_{module}.{ext}  # jest, pytest, cargo test, etc.

# Capture output
{language-runtime} poc.{ext} > results/poc-output.txt 2>&1

# Benchmark (if performance question)
time {language-runtime} poc.{ext}
# or: wrk, ab, k6, etc.

Step 5: Document Findings

README.md Template:

# PoC: {Feature or Risk Name}

**Technical Question**: {one-sentence question}

**Date**: {date}
**Duration**: {hours spent}
**Scope**: {Minimal | Standard | Comprehensive}
**Status**: {SUCCESS | PARTIAL | FAILED}

## Objective

{Explain what we're trying to prove or disprove}

## Approach

{Describe the technical approach taken}

**Technology Used**:
- Language: {language + version}
- Libraries: {list key dependencies}
- Environment: {OS, runtime versions}

**Key Implementation Details**:
- {Detail 1}
- {Detail 2}

## Results

**Outcome**: {PROVEN | DISPROVEN | INCONCLUSIVE}

**Summary**:
{2-3 sentences: What did we learn? Did we answer the question?}

**Evidence**:
- {Result 1: e.g., API integration works, response time: 150ms}
- {Result 2: e.g., Library handles edge case successfully}
- {Result 3: e.g., Performance meets requirements: 1200 req/s}

**Blockers Encountered**:
- {Blocker 1} - Workaround: {description}
- {Blocker 2} - Unresolved (needs research)

## Findings

**What Worked**:
- {Positive finding 1}
- {Positive finding 2}

**What Didn't Work**:
- {Limitation 1}
- {Limitation 2}

**Assumptions Validated**:
- ✅ {Assumption 1} - Confirmed
- ✅ {Assumption 2} - Confirmed

**Assumptions Invalidated**:
- ❌ {Assumption 3} - Incorrect, actual: {correction}

## Recommendations

**Decision**: {GO | NO-GO | ALTERNATIVE APPROACH}

**Rationale**:
{Why we recommend this decision based on PoC findings}

**Next Steps**:
1. {If GO: Implement in prototype or production}
2. {If NO-GO: Explore alternative approach}
3. {If ALTERNATIVE: Describe alternative}

**Risks Retired**:
- **Risk-{ID}**: {risk-description} - Status: {RETIRED | MITIGATED | ACCEPTED}

## Code Location

**Files**:
- Implementation: `poc.{ext}` or `src/main.{ext}`
- Tests: `tests/test_{module}.{ext}` (if applicable)
- Results: `results/poc-output.txt`

**How to Run**:
```bash
# Setup (if needed)
{setup-commands}

# Run PoC
{run-command}

# Expected output:
{expected-output-summary}

Cleanup

Disposition: {KEEP | ARCHIVE | DELETE}

• KEEP: PoC code is valuable, integrate into codebase
• ARCHIVE: PoC served its purpose, archive for reference (move to poc-archive/)
• DELETE: PoC can be discarded, learning documented

Technical Debt (if keeping code):

• {Debt item 1} - Plan: {how to address}
• {Debt item 2} - Plan: {how to address}

## Common PoC Patterns

### API Integration PoC

**Question**: Can we integrate with External API X?

**Minimal Scope**:
- Authentication (API key or OAuth)
- Single API call (GET /endpoint)
- Parse response
- Capture output

**Technology**: Use HTTP client library (curl, requests, axios, etc.)

**Success**: API call succeeds, response parsed

### Performance PoC

**Question**: Can we process N items/sec with Technology Y?

**Minimal Scope**:
- Generate sample data
- Process with Technology Y
- Measure throughput
- Compare to requirement

**Technology**: Use benchmarking tool (time, wrk, k6, etc.)

**Success**: Throughput ≥ requirement

### Database Integration PoC

**Question**: Can we use Database X with our data model?

**Minimal Scope**:
- Connect to database
- Create schema (1-2 tables)
- Insert sample data
- Query data
- Measure query time

**Technology**: Use database driver (psycopg2, mysql-connector, pg, etc.)

**Success**: CRUD operations work, query time acceptable

### Framework Evaluation PoC

**Question**: Can Framework X handle our use case?

**Minimal Scope**:
- Implement "Hello World" with framework
- Add 1-2 real features from use case
- Test edge case
- Capture learnings

**Technology**: Framework quickstart/tutorial

**Success**: Features work, edge case handled, documentation sufficient

## Failure Modes and Recovery

### PoC Takes Too Long

**Symptom**: 3+ days spent, no clear answer

**Recovery**:
- Stop immediately (time-box exceeded)
- Document what was learned so far
- Recommend alternative approach or more research
- **Don't**: Continue building, scope creep to prototype

### PoC Answers Wrong Question

**Symptom**: PoC succeeds but doesn't retire risk or inform decision

**Recovery**:
- Revisit technical question (Step 1)
- Clarify what decision depends on PoC
- Scope new PoC with correct question
- Archive existing PoC code

### PoC Becomes Production Code

**Symptom**: Team wants to use PoC code in production

**Recovery**:
- **Resist**: PoC code is not production-ready (no error handling, no tests, hardcoded)
- If truly valuable: Rewrite properly in prototype/Construction
- Document technical debt if PoC code is used
- Plan refactoring before production deployment

## Output

**Deliverables**:
1. PoC code (in `poc-{feature-name}/` directory)
2. README.md with findings and recommendation
3. Test results (if standard/comprehensive scope)
4. Risk retirement update (if PoC retires risk)

**Decision Artifact**:
```markdown
# PoC Decision: {Feature or Risk}

**Technical Question**: {question}
**PoC Date**: {date}
**Duration**: {hours}

**Outcome**: {PROVEN | DISPROVEN | INCONCLUSIVE}
**Decision**: {GO | NO-GO | ALTERNATIVE APPROACH}

**Rationale**: {1-2 sentences}

**Evidence**: {link to README.md and results}

**Next Steps**:
- {action-item-1}
- {action-item-2}

**Risks Retired**:
- Risk-{ID}: {status}

Integration with SDLC

Phase Applicability:

• Elaboration: Validate architectural risks, prove technology choices
• Construction: Spike on complex features, validate edge cases
• Transition: Validate deployment approach, performance tuning

Related Commands:

• flow-inception-to-elaboration - PoCs retire risks during Elaboration
• flow-risk-management-cycle - PoCs used to retire technical risks
• flow-gate-check - PoC results used as gate criteria evidence

Related Agents:

• software-implementer - Implements PoC code
• test-engineer - Validates PoC with tests
• architecture-designer - Reviews PoC for architectural implications

Success Criteria

This command succeeds when:

• Technical question clearly defined
• PoC scope minimal and time-boxed (≤3 days)
• Code implemented and runs successfully
• Findings documented in README.md
• Decision recorded (GO/NO-GO/ALTERNATIVE)
• Risks retired (if applicable)
• Next steps identified

Error Handling

Unclear Question:

• Report: "Technical question too broad or vague"
• Action: "Refine question to specific, testable hypothesis"
• Example: "Can we use Framework X?" → "Can Framework X handle 1000 concurrent WebSocket connections?"

Scope Creep:

• Report: "PoC scope expanded beyond minimal proof"
• Action: "Reset to minimal scope, defer features to prototype/Construction"
• Reminder: "PoC answers ONE technical question, not multiple"

Time-box Exceeded:

• Report: "PoC exceeded {hours} hours without clear answer"
• Action: "Document findings so far, recommend alternative approach or more research"
• Escalation: "Consider if question is too complex for PoC (needs prototype)"

Examples

Example 1: API Integration

/build-poc "Can we authenticate with Stripe API using OAuth2?" --scope minimal

# Output:
# poc-stripe-oauth/
#   README.md (findings: YES, OAuth2 works, token refresh needed)
#   poc.js (code demonstrating OAuth flow)
#   results/poc-output.txt (sample API responses)
# Decision: GO - integrate Stripe API in prototype

Example 2: Performance Validation

/build-poc "Can we process 10,000 events/sec with Redis Streams?" --scope standard

# Output:
# poc-redis-streams/
#   README.md (findings: YES, 12,500 events/sec achieved)
#   src/producer.js, src/consumer.js
#   tests/benchmark.js
#   results/performance-baseline.txt
# Decision: GO - use Redis Streams for event processing

Example 3: Technology Evaluation

/build-poc "Can Next.js handle our SSR + ISR requirements?" --scope comprehensive

# Output:
# poc-nextjs-ssr/
#   README.md (findings: PARTIAL, SSR works, ISR has edge case issue)
#   src/ (multi-page Next.js app)
#   tests/ (unit + integration tests)
#   results/ (findings, performance baseline, edge case documentation)
# Decision: ALTERNATIVE - Use SSR only, defer ISR to later

Command Version: 1.0 Phase: Any (Elaboration most common) Duration: 4 hours - 3 days (time-boxed) Output: PoC code + README + decision

References

• @$AIWG_ROOT/agentic/code/addons/aiwg-utils/rules/vague-discretion.md — PoC completion criteria must be concrete and time-boxed; avoid "done when working" conditions
• @$AIWG_ROOT/agentic/code/addons/aiwg-utils/rules/human-authorization.md — Seek authorization before deciding the PoC outcome changes the architectural approach
• @$AIWG_ROOT/agentic/code/addons/aiwg-utils/rules/research-before-decision.md — Research technology and existing patterns before implementation
• @$AIWG_ROOT/agentic/code/frameworks/sdlc-complete/skills/security-audit/SKILL.md — Consider security during PoC if the risk involves authentication or data handling
• @$AIWG_ROOT/agentic/code/frameworks/sdlc-complete/rules/anti-laziness.md — Never skip testing or weaken PoC validation criteria to declare success faster