stream-coding（流编码）

Stream Coding v3.5: Documentation-First Development

⚠️ CRITICAL REFRAME: THIS IS A DOCUMENTATION METHODOLOGY, NOT A CODING METHODOLOGY

The Goal: AI-ready documentation. When documentation is clear enough, code generation becomes automatic.

The Insight:

"If your docs are good enough, AI writes the code. The hard work IS the documentation. Code is just the printout."

v3.5 Core Additions:

1. Phase 2.5 Adversarial Review — stress-test specs with a hostile critic before execution. New time allocation: 40/40/5/10/5.
2. Renamed internal 13-item checklist to Spec Gate (structural completeness for code generation)

Two-gate pipeline — know the difference: - Spec Gate (13 items, this skill) → *"Can AI execute this without asking questions?"* - Clarity Gate (9 points, separate tool) → *"Will AI mistake assumptions for facts?"* Spec Gate → Clarity Gate → Adversarial Review → Verified Specification → Execution

CHANGELOG

THE STREAM CODING TRUTH

Messy Docs → Vague Specs → AI Guesses → Rework Cycles → 2-3x Velocity
Clear Docs → Clear Specs → AI Executes → Minimal Rework → 10-20x Velocity

Why Most "AI-Assisted Development" Fails:

• People feed AI messy docs
• AI generates code based on assumptions
• Code doesn't match intent
• Endless revision cycles
• Result: Marginally faster than manual coding

Why Stream Coding Achieves 10-20x:

• Documentation is clarified FIRST
• AI has zero ambiguity
• Code matches intent on first pass
• Minimal revision
• Result: Documentation time + automatic code generation

DOCUMENT TYPE ARCHITECTURE

The Rule: Not all documents need all sections. Putting implementation details in strategic documents violates single-source-of-truth.

"If AI has to decide where to find information, you've already lost velocity."

Document Types

Section Placement Matrix

Why This Matters

Wrong (violates single-source-of-truth):

Master Blueprint
├── Strategy content
├── Anti-patterns ← WRONG: duplicates Technical Spec
├── Test Cases ← WRONG: duplicates Testing doc
└── Error Matrix ← WRONG: duplicates Error Handling doc

Right (single-source-of-truth):

Master Blueprint (Strategic)
├── Strategy content
└── References
    └── Pointer: "Anti-patterns → Technical Spec, Section 7"

Technical Spec (Implementation)
├── Implementation details
├── Anti-patterns ← CORRECT: lives here
├── Test Cases ← CORRECT: lives here
└── Error Matrix ← CORRECT: lives here

THE 4-PHASE METHODOLOGY

Time Allocation

The logic: Spending 5% upfront to break specs saves more than 5% in Phase 3 rework. If you bulletproof docs adversarially, execution gets faster — not slower.

PHASE 2.5: ADVERSARIAL REVIEW (5% of time)

When to Run

After Spec Gate passes (9+/10). Before any code generation.

The principle: The same AI that wrote your specs has the same blind spots you do. A different model — or a human with instructions to attack — finds what you can't see.

"When code fails, fix the spec — not the code. Phase 2.5 finds spec failures before there's any code to fail."

The Process

1. Spec Gate passes (9+/10) → proceed to adversarial step
2. Submit specs to DIFFERENT AI model (Gemini, GPT-4, Perplexity)
   OR trusted human reviewer
3. Use the adversarial prompt below
4. Categorize findings: CRITICAL / HIGH / MEDIUM / LOW
5. Fix ALL CRITICAL issues → return to Spec Gate → re-score
6. Document HIGH issues with explicit accept/defer decision
7. Gate: zero CRITICAL remaining → proceed to Phase 3

Why a different model: The AI that generated or reviewed your docs learned your assumptions. A different model has no context, no charitable interpretation, no benefit of the doubt. It finds gaps your primary AI normalizes.

The Adversarial Prompt Template

You are a skeptical senior developer and hostile critic reviewing
this specification before it goes to an AI agent for execution.

## Your Mission
Find every flaw. Assume problems exist — your job is to find them.
Do not be helpful. Do not suggest minor improvements. Attack the spec.

## What to Look For

### 1. LOGICAL CONTRADICTIONS
- Claims that conflict with each other within the spec
- Numbers that don't add up
- Requirements that are mutually exclusive

### 2. CREDIBILITY RISKS
- Overclaims ("zero bugs", "always", "never", "guaranteed")
- Unverifiable statements with no measurement method
- Claims a hostile reader would immediately challenge

### 3. IMPLICIT DEGREES OF FREEDOM
- Points where the AI agent must CHOOSE between valid interpretations
- Anything where two different developers would implement differently
- Edge cases that are mentioned but not fully specified

### 4. MISSING CONSIDERATIONS
- Error states that have no specified handling
- Concurrency or race conditions not addressed
- External dependencies with no fallback specified
- Security assumptions not made explicit

### 5. DEFENSIBILITY GAPS
- "What would a hostile HN commenter use to debunk this?"
- "What would a junior developer get wrong from this spec?"
- "What happens when the happy path fails?"

## Output Format

For each issue found:
**[SEVERITY]** — Issue title
Location: Where in the spec
Problem: What exactly is wrong
Fix: Specific rewrite needed

Severity:
- **CRITICAL:** Execution will fail or produce wrong output without this fix
- **HIGH:** Significant risk of incorrect implementation
- **MEDIUM:** Minor ambiguity, lower risk
- **LOW:** Polish, not blocking

## Success Criteria

A good adversarial review finds:
- At least 2 CRITICAL issues (if zero, you haven't looked hard enough)
- At least 4-5 HIGH issues
- 10+ total issues across all severities

If you find fewer, state explicitly why the spec is unusually strong.

Gate Criteria

• Zero CRITICAL issues remaining
• All HIGH issues documented with explicit decision: fix now / accept risk / defer
• Spec Gate re-run if any CRITICAL was fixed (score may have changed)

What Counts as "Different AI"

Rule: always use a different provider than your primary workflow. Same provider = confirmation bias.

This step is intentionally manual — paste specs into a different chat and apply the adversarial prompt. No API key or integration required.

Never use the same Claude session that helped write the docs. Start a fresh session at minimum — different model preferred.

PHASE 1: STRATEGIC THINKING (40% of time)

Decision Tree: Where Do You Start?

Phase 1: Strategic Product Thinking
│
├─ Have existing documentation?
│   └─ YES → Start with Documentation Audit → then 7 Questions
│
└─ Starting fresh?
    └─ Skip to 7 Questions

Documentation Audit (Conditional)

Skip this step if starting from scratch. The Documentation Audit only applies when you have existing documentation—previous specs, inherited docs, or accumulated notes.

Why clean existing docs? Because most documentation accumulates cruft:

• Aspirational statements ("We will revolutionize...")
• Speculative futures ("In 2030, we might...")
• Outdated decisions (v1 architecture in v3 docs)
• Duplicate information across files
• Motivational fluff with no implementation value

The Audit Process:

Apply the Clarity Test to all existing documentation:

Audit Checklist:

• Remove all "vision" and "future state" language
• Delete motivational conclusions and preambles
• Consolidate duplicate information to single source
• Update all outdated architectural decisions
• Remove speculative features not in current scope

Target: 40-50% reduction in volume without losing actionable information.

Once clean, proceed to the 7 Questions.

The 7 Questions Framework

Before ANY new documentation, answer these with specificity. Vague answers = vague code.

Phase 1 Exit Criteria

• All 7 questions answered at "Require" level
• Strategic Blueprint document created
• Architecture Decision Records (ADRs) for major choices
• Zero ambiguity about WHAT you're building

PHASE 2: AI-READY DOCUMENTATION (40% of time)

The 4 Mandatory Sections (Implementation Docs)

Every implementation document MUST include these four sections. Without them, AI guesses—and guessing creates the velocity mirage.

1. Anti-Patterns Section

Why: AI needs to know what NOT to do.

## Anti-Patterns (DO NOT)

| ❌ Don't | ✅ Do Instead | Why |
|----------|---------------|-----|
| Store timestamps as Date objects | Use ISO 8601 strings | Serialization issues |
| Hardcode configuration values | Use environment variables | Deployment flexibility |
| Use generic error messages | Specific error codes per failure | Debugging impossible otherwise |
| Skip validation on internal calls | Validate everything | Internal calls can have bugs too |
| Expose internal IDs in APIs | Use UUIDs or slugs | Security and flexibility |

Rules: Minimum 5 anti-patterns per implementation document.

2. Test Case Specifications

Why: AI needs concrete verification criteria.

## Test Case Specifications

### Unit Tests Required
| Test ID | Component | Input | Expected Output | Edge Cases |
|---------|-----------|-------|-----------------|------------|
| TC-001 | Tier classifier | 100 contacts | 20-30 in Critical tier | Empty list, all same score |
| TC-002 | Score calculator | Activity array | Score 0-100 | No events, >1000 events |

### Integration Tests Required
| Test ID | Flow | Setup | Verification | Teardown |
|---------|------|-------|--------------|----------|
| IT-001 | Auth flow | Create test user | Token refresh works | Delete test user |

Rules: Minimum 5 unit tests, 3 integration tests per component.

3. Error Handling Matrix

Why: AI needs to know how to handle every failure mode.

## Error Handling Matrix

### External Service Errors
| Error Type | Detection | Response | Fallback | Logging | Alert |
|------------|-----------|----------|----------|---------|-------|
| API timeout | >5s response | Retry 3x exponential | Return cached | ERROR | If 3 in 5 min |
| Rate limit | 429 response | Pause 15 min | Queue for retry | WARN | If >5/hour |

### User-Facing Errors
| Error Type | User Message | Code | Recovery Action |
|------------|--------------|------|-----------------|
| Quota exceeded | "You've used all checks this month." | 403 | Show upgrade CTA |
| Session expired | "Please sign in again." | 401 | Redirect to login |

Rules: Every external service and user-facing error must be specified.

4. Deep Links (All Document Types)

Why: AI needs to navigate to exact locations. "See Technical Annexes" is useless.

## References

### Schema References
| Topic | Location | Anchor |
|-------|----------|--------|
| User profiles | [Schema Reference](../schemas/schema.md#user_profiles) | `user_profiles` |
| Events table | [Schema Reference](../schemas/schema.md#events) | `events` |

### Implementation References
| Topic | Document | Section |
|-------|----------|---------|
| Auth flow | [API Spec](../specs/api.md#authentication) | Section 3.2 |
| Rate limiting | [API Spec](../specs/api.md#rate-limiting) | Section 5 |

Rules: NEVER use vague references. ALWAYS include document path + section anchor.

⚠️ THE SPEC GATE (v3.5 - COMPLETE)

⛔ NEVER SKIP THIS GATE.

This is the difference between stream coding and vibe coding. A 7/10 spec generates 7/10 code that needs 30% rework.

The 13-Item Spec Gate Checklist

Before ANY code generation, verify ALL items pass:

Foundation Checks (7 items)

Document Architecture Checks (6 items - v3.3 Critical)

Gate Enforcement

- [ ] All 7 Foundation Checks pass
- [ ] All 6 Document Architecture Checks pass
- [ ] AI Coder Understandability Score ≥ 9/10

If ANY item fails → Fix before proceeding to Phase 3

AI CODER UNDERSTANDABILITY SCORING

Use this rubric to score documentation. Target: 9+/10 before Phase 3.

The 6-Criterion Rubric

Score Interpretation

Self-Assessment Questions

Before Phase 3, ask yourself:

1. Actionability: "Does every section tell AI exactly what to do?"
2. Specificity: "Are there any numbers I left vague?"
3. Consistency: "Is any information stated in more than one place?"
4. Structure: "Could I convert any prose paragraphs to tables?"
5. Disambiguation: "Have I listed at least 5 anti-patterns per implementation doc?"
6. Reference Clarity: "Do any references say 'see elsewhere' without exact location?"

If you answer "no" or "yes" to any question that should be opposite → Fix before proceeding.

AI-ASSISTED SPEC GATE (Meta-Prompt)

Use this prompt to have Claude score your documentation:

**ROLE:** You are the Spec Gatekeeper. Your job is to ruthlessly
evaluate software specifications for ambiguity, incompleteness, and
"vibe coding" tendencies.

**INPUT:** I will provide a technical specification document.

**TASK:** Grade this document on a scale of 1-10 using this rubric:

**RUBRIC:**
1. **Actionability (25%):** Does every section dictate a specific
   implementation detail? (Reject aspirational like "fast" or
   "scalable" without metrics)
2. **Specificity (20%):** Are data types, error codes, thresholds,
   and edge cases explicitly defined? (Reject "handle errors appropriately")
3. **Consistency (15%):** Single source of truth? No duplicates?
4. **Structure (15%):** Tables over prose? Clear hierarchy?
5. **Disambiguation (15%):** Anti-patterns present? Edge cases explicit?
6. **Reference Clarity (10%):** Deep links only? No vague references?

**OUTPUT FORMAT:**
1. **Score:** [X]/10
2. **Criterion Breakdown:** Score each of the 6 criteria
3. **Hallucination Risks:** List specific lines where an AI developer
   would have to guess or make an assumption
4. **The Fix:** Rewrite the 3 most ambiguous sections into AI-ready specs

**THRESHOLD:**
- 9-10: Ready for code generation
- 7-8: Needs revision before proceeding
- <7: Return to Phase 2

PHASE 3: EXECUTION (10% of time)

The Generate-Verify-Integrate Loop

1. GENERATE: Feed spec to AI → Receive code
2. VERIFY: Run tests → Check against spec
   - Does output match spec exactly?
   - Yes → Continue
   - No → Fix SPEC first, then regenerate
3. INTEGRATE: Commit → Update documentation if needed

The Golden Rule of Phase 3

"When code fails, fix the spec—not the code."

If generated code doesn't work:

1. ❌ Don't patch the code manually
2. ✅ Ask: "What was unclear in my spec?"
3. ✅ Fix the spec
4. ✅ Regenerate

Why: Manual code patches create divergence between spec and reality. Divergence compounds. Eventually your spec is fiction and you're back to manual development.

PHASE 4: QUALITY & ITERATION (5% of time)

The Rule of Divergence

Every time you manually edit AI-generated code without updating the spec, you create Divergence. Divergence is technical debt.

Why Divergence is Dangerous:

• If you fix a bug in code but not spec, you can never regenerate that module
• Future AI iterations will reintroduce the bug
• You've broken the stream

Preventing Divergence

The "Day 2" Workflow

1. Isolate the Module: Target the specific module, not the whole app
2. Update the Spec: Add the new edge case, requirement, or fix
3. Regenerate the Module: Feed updated spec to AI
4. Verify Integration: Run test suite for regressions

This takes 5 minutes longer than a quick hotfix. But it ensures your documentation never drifts from reality.

TRIGGER BEHAVIOR

This methodology activates when the user says:

• "Build [feature]" → Full methodology (Phases 1-4)
• "Create [component]" → Full methodology
• "Implement [system]" → Check: Do clear docs exist?
• "Document [project]" → Phases 1-2 only
• "Spec out [feature]" → Phases 1-2 only
• "Clean up docs for [X]" → Documentation Audit only

Response Protocol

1. Check for existing docs: "Do you have existing documentation for this project?"
2. If existing docs: "Let's start with a Documentation Audit to clean them before building."
3. If Phase 1 incomplete: "Before building, let's clarify strategy. [Ask 7 Questions]"
4. If Phase 2 incomplete: "Before coding, let's ensure documentation is AI-ready. [Run Spec Gate]"
5. If Spec Gate not passed: "Documentation scores [X]/10. Let's fix [specific issues] before proceeding."
6. If Phase 3 ready: "Documentation passes Spec Gate (9+/10). Generating implementation..."
7. If maintaining (Phase 4): "Is this change spec-conformant? Let's update docs first."

THE STREAM CODING CONTRACT

YOU MUST:

Documentation Audit (if existing docs):

• Run Clarity Test on all existing documentation
• Remove aspirational/future state language
• Consolidate duplicates to single source
• Target 40-50% reduction without losing actionable content

Phase 1:

• Answer all 7 questions at "Require" level
• Create Strategic Blueprint with Implementation Implications
• Write ADRs for major architectural decisions

Phase 2:

• Identify document type (Strategic vs Implementation vs Reference)
• Add 4 mandatory sections to each implementation doc
• Add deep links to ALL documents
• Use pointers (not duplicates) in strategic docs

Phase 2.5 — Adversarial Review:

• Submit specs to different AI model or human reviewer
• Use adversarial prompt template
• Fix ALL CRITICAL issues before Phase 3
• Document HIGH issues with accept/defer decision
• Re-run Spec Gate if CRITICAL issues were fixed

Spec Gate:

• Pass all 13 checklist items
• Score 9+/10 on AI Coder Understandability
• Answer all 6 self-assessment questions correctly

Phase 3-4:

• Show code before creating files
• Run quality gates (lint, type, test)
• When code fails: fix spec, regenerate
• Never create divergence (update spec with every code change)

YOU CANNOT:

• ❌ Build on existing docs without running Documentation Audit first
• ❌ Skip to coding without clear docs
• ❌ Accept vague specs ("handle errors appropriately")
• ❌ Skip Spec Gate (even if you wrote the docs yourself)
• ❌ Put Anti-patterns/Test Cases/Error Handling in strategic docs
• ❌ Use vague references ("see Technical Annexes")
• ❌ Duplicate content across document types
• ❌ Iterate on code when problem is in spec
• ❌ Edit code without updating spec (creates Divergence)

DOCUMENT TEMPLATES

Strategic Document Template

# [Document Title] (Strategic)

## 1. [Strategic Section]
[Strategic content]

**Implementation Implication:** [Concrete effect on code/architecture]

## 2. [Another Section]
[Strategic content]

**Implementation Implication:** [Concrete effect on code/architecture]

## N. REFERENCES

### Implementation Details Location
| Content Type | Location |
|--------------|----------|
| Anti-patterns | [Technical Spec, Section 7](path#anchor) |
| Test Cases | [Testing Doc, Section 3](path#anchor) |
| Error Handling | [Error Handling Doc](path#anchor) |

### Schema References
| Topic | Location | Anchor |
|-------|----------|--------|
| [Topic] | [Path](path#anchor) | `anchor` |

*This document provides strategic overview. Technical documents provide implementation specifications.*

Implementation Document Template

# [Document Title] (Implementation)

## 1. [Implementation Section]
[Technical details]

## N-3. ANTI-PATTERNS (DO NOT)

| ❌ Don't | ✅ Do Instead | Why |
|----------|---------------|-----|
| [Anti-pattern] | [Correct approach] | [Reason] |

## N-2. TEST CASE SPECIFICATIONS

### Unit Tests
| Test ID | Component | Input | Expected Output | Edge Cases |
|---------|-----------|-------|-----------------|------------|
| TC-XXX | [Component] | [Input] | [Output] | [Edge cases] |

### Integration Tests
| Test ID | Flow | Setup | Verification | Teardown |
|---------|------|-------|--------------|----------|
| IT-XXX | [Flow] | [Setup] | [Verify] | [Cleanup] |

## N-1. ERROR HANDLING MATRIX

| Error Type | Detection | Response | Fallback | Logging |
|------------|-----------|----------|----------|---------|
| [Error] | [How detected] | [Response] | [Fallback] | [Level] |

## N. REFERENCES

| Topic | Location | Anchor |
|-------|----------|--------|
| [Topic] | [Path](path#anchor) | `anchor` |

QUICK REFERENCE

The 13-Item Spec Gate

Foundation (7):

1. Actionable? 2. Current? 3. Single source? 4. Decision not wish?
2. Prompt-ready? 6. No future state? 7. No fluff?

Architecture (6): 8. Type identified? 9. Anti-patterns placed correctly? 10. Test cases placed correctly? 11. Error handling placed correctly? 12. Deep links present? 13. No duplicates?

The Scoring Rubric

Time Allocation

┌─────────────────────────────────────────────────────────────┐
│ Have existing docs? → Documentation Audit (conditional)     │
├─────────────────────────────────────────────────────────────┤
│                                                             │
│  Phase 1 (Strategy): 40% ──┐                                │
│  Phase 2 (Specs): 40% ─────┼── 80% Documentation            │
│                            │                                │
│  ⚠️ SPEC GATE ─────────────┘                                │
│                            │                                │
│  🔴 Phase 2.5 (Adversarial): 5% ← different AI attacks spec │
│                            │                                │
│  Phase 3 (Code): 10% ──────┼── 15% Code + Quality           │
│  Phase 4 (Quality): 5% ────┘                                │
│                                                             │
└─────────────────────────────────────────────────────────────┘

Core Mantras

1. "Documentation IS the work. Code is just the printout."
2. "When code fails, fix the spec—not the code."
3. "A 7/10 spec generates 7/10 code that needs 30% rework."
4. "If AI has to decide where to find information, you've already lost velocity."

Version: 3.5 Changes from 3.4:

• Phase 2.5 Adversarial Review added (full section with process, prompt template, gate criteria)
• Time allocation updated: 40/40/5/10/5 (was 40/40/15/5)
• "Clarity Gate" renamed to "Spec Gate" throughout (structural completeness, 13 items)
• Two-gate pipeline documented: Spec Gate (this skill) + Clarity Gate (separate standalone tool for epistemic quality)
• Contract section updated with Phase 2.5 checklist
• Quick reference diagram updated
• Core insight: Spec Gate catches completeness. Adversarial Review catches correctness. Clarity Gate (standalone) catches epistemic drift.

*Stream Coding by Francesco Marinoni Moretto — CC BY 4.0* *github.com/frmoretto/stream-coding*

END OF STREAM CODING v3.5