Result Synthesizer

Result Synthesizer

Combine internal investigation findings + external AI responses into a single, actionable report with cross-validated conclusions.

Announce at start: "I'm using the /result skill to synthesize all research findings."

Pipeline position:

/question → prompts + internal findings
     ↓
User pastes prompt → external AI responses
     ↓
/result ① ← synthesize question findings → approach decided
     ↓
/research → prompts + internal findings
     ↓
User pastes prompt → external AI responses
     ↓
/result ② ← synthesize research findings → decisions ready
     ↓
/spec → lock decisions → /guide → 구현 → /validation

/result is reusable — call it after ANY investigation round.
It reads whatever internal findings + external responses exist and synthesizes them.

STEP 0: Gather All Sources

Collect everything available:

Internal Sources (auto-read)

Scan for files matching these patterns:
  docs/reports/*-question-internal-findings.md
  docs/reports/*-research-internal-findings.md
  docs/reports/*-validation-internal-findings.md
  docs/reports/*-problem-internal-findings.md
  docs/reports/*-research-report.md  (legacy — pre-pipeline manual reports)

External Sources (user provides)

The user pastes or references external AI responses. Accept them as:

• Direct paste in chat
• File paths (if user saved responses to files)
• URLs (if responses are in a shared doc — may fail for auth-protected pages; if URL cannot be fetched, paste content directly)

Label each source for traceability:

• Internal-Claude (our own agent investigation)
• External-ChatGPT / External-Gemini / External-Perplexity / External-Claude

STEP 1: Source Inventory

Before synthesizing, list what we have:

## Source Inventory

| # | Source | Type | Topic | Word Count |
|---|--------|------|-------|------------|
| 1 | Internal investigation | question | Tech landscape | ~2000 |
| 2 | ChatGPT response | question | Tech landscape | ~1500 |
| 3 | Gemini response | question | Tech landscape | ~1800 |
| 4 | Internal investigation | research | AudioWorklet deep-dive | ~3000 |
| 5 | Gemini response | research | Implementation plan | ~2500 |

Total sources: 5

If < 2 sources: Warn user that cross-validation is limited.

Context Mode Detection

Check how this skill was invoked:

Full mode (via orchestrator pipeline):

• IF args contain enriched_prompt: path → Read the enriched prompt file and use as primary context
• This enriched prompt already contains claude_guide knowledge, project context, and complexity analysis

Degraded mode (direct invocation):

• IF no enriched prompt → Read claude_guide/INDEX.md and select 1-2 relevant documents
• Load only those selected documents for lightweight context
• Note: "Running in standalone mode. For best results, use /orchestrator."

STEP 2: Cross-Validation Matrix

Synthesis Mode Detection

Detect inbound context to adjust synthesis behavior:

How to detect: Check source file patterns (priority order):

• *-problem-internal-findings.md → Root Cause Analysis mode
• *-validation-internal-findings.md → Quality Assessment mode
• *-research-internal-findings.md → Decision mode
• *-question-internal-findings.md → Discovery mode
• Multiple types present → use highest-priority mode from the list above

For each key claim/recommendation, check which sources agree:

## Cross-Validation

| Claim | Internal | ChatGPT | Gemini | Perplexity | Confidence |
|-------|----------|---------|--------|------------|------------|
| Use AudioWorklet over ScriptProcessor | Y | Y | Y | - | HIGH |
| 48kHz→16kHz decimation needed | Y | Y | Y | - | HIGH |
| FIR anti-aliasing filter required | Y | N | N | - | MEDIUM (only internal found this) |
| Cloudflare Tunnel for exposure | Y | Y | Y | - | HIGH |
| HMAC token over simple token | N | N | Y | - | LOW (only one source) |

Confidence levels → Action mapping:

• HIGH: All sources agree (or 3+ in absolute terms) → action per mode table above
• MEDIUM: Majority agrees but minority disagrees or is silent → test first (validate before committing)
• LOW: Only 1 source mentions it → hypothesis (note but don't act yet)
• CONFLICT: Sources actively disagree (contradictory claims) → research more (need /research deep-dive on this specific item)

Note: With only 2 sources, 2/2 agreement = HIGH, not MEDIUM. Confidence is about agreement ratio, not absolute count.

STEP 3: Conflict Resolution

For each item below HIGH confidence (CONFLICT, LOW, or MEDIUM with opposing evidence):

## Conflicts & Uncertainties

### Anti-aliasing filter (MEDIUM)
- **Internal says:** FIR 7-tap filter mandatory — naive decimation causes 8-24kHz aliasing
- **External says:** Just take every 3rd sample
- **Resolution:** Internal is technically correct. Naive decimation does alias.
  Impact depends on use case — for STT, aliased frequencies may or may not
  affect accuracy. Recommend testing both.
- **Action:** Test with and without filter, compare STT word error rate.

STEP 4: Generate Unified Report

Report Template

# [Topic] — Research Synthesis Report

> Synthesized from [N] sources on [date]

## Executive Summary
[3-5 sentences: what we learned, key decisions, confidence level]

## Consensus Findings (HIGH confidence)
[Items all sources agree on — safe to act on immediately]

## Strong Recommendations (MEDIUM confidence)
[Items most sources agree on — act on with awareness of caveats]

## Needs Further Investigation (LOW confidence)
[Items only one source mentioned — worth exploring but don't commit yet]

## Conflicts & Divergences
[Items where sources disagree — include both perspectives and resolution]

## Decision Matrix
[Final comparison table combining all sources]

| Option | Consensus Rating | Key Strengths | Key Risks | Recommended? |
|--------|-----------------|---------------|-----------|--------------|

## Recommended Next Steps
1. [Immediate action based on HIGH confidence findings]
2. [Investigation needed for MEDIUM items]
3. [Defer or test for LOW/CONFLICT items]

## Source Attribution
[Which finding came from which source — for traceability]

## Appendix: Raw Source Summaries
[Brief summary of each source's key points.
Preserve exact numbers (versions, benchmarks, measurements), specific API names,
and direct quotes. Summarize narrative explanations and general observations.]

STEP 5: File Output

Save to:

• docs/reports/{topic}-synthesis-report.md

Korean translation: Generate Korean version (-ko.md) if the user's primary language is Korean. Korean rules:

• Technical terms: first occurrence "한국어(English)", then English only
• Code, file paths, commands, version numbers stay in English

Print executive summary + decision matrix to console.

STEP 6: Actionable Handoff

Based on synthesis findings, route to the appropriate next step:

If findings point to a clear winner (MEDIUM+ complexity — multiple files, external dependencies, or architectural decisions needed):
  → "Decisions ready to lock. Use /spec to define the implementation specification."

If findings point to a clear winner (TRIVIAL/SIMPLE — single file, no external dependencies, no architectural decisions):
  → "Simple enough to implement directly. Use /guide to start."

If 2-3 options need deeper comparison:
  → "Use /research to deep-dive on [Option A] and [Option B]."

If fundamental uncertainty remains:
  → "Need proof-of-concept. Build minimal prototype of [X] to validate."

If a bug or error was discovered during synthesis:
  → "Use /problem to investigate: [specific issue]"

STEP 6 (continued): Circuit Breaker

Track research iteration depth by checking for existing synthesis reports. Detection: If docs/reports/{topic}-synthesis-report.md already exists, read its Source Inventory and confidence levels. If the current round has the same sources and unchanged confidence levels, this is a repeat without new information.

If this is the 2nd /result synthesis for the SAME topic without new information:

⚠️ Research cycle limit reached.
You've synthesized findings on [topic] twice without convergence.

Options:
1. Lock decisions now → /spec (accept current best option with noted risks)
2. Change research angle → /research with a DIFFERENT question
3. Abandon this approach → return to /question for alternative exploration

Signs of research thrashing (stop and escalate):

• Same options keep appearing in successive rounds
• Confidence levels haven't changed between rounds
• No new sources or evidence in the latest round
• User has pasted the same external AI response twice

STEP 7: Quality Checklist

Must-pass

• All available sources inventoried (internal + external)
• Synthesis mode detected (Discovery / Decision / Quality Assessment / Root Cause Analysis)
• Cross-validation matrix complete for key claims
• Confidence levels assigned (HIGH/MEDIUM/LOW/CONFLICT)
• Executive summary is actionable (not just descriptive)
• Routing decision present with next step

Should-pass

• Conflicts explicitly noted with resolution or "needs testing"
• Decision matrix present with clear recommendation column
• Source attribution included for traceability
• Exact numbers, versions, and API names preserved (not summarized away)

Pipeline State Update

If .claude/pipeline-state.md exists, update it before concluding:

1. YAML frontmatter: set delegated_to: to empty, update updated: to today
2. Markdown body: add result to Completed, update Artifacts with generated files, set Recommended Next to /spec, /research, /guide

---## Anti-Patterns

• Cherry-picking: Don't only include findings that support one option
• False consensus: 2 AI tools parroting the same blog post ≠ independent confirmation
• Ignoring internal findings: Internal investigation often finds things external AI misses (e.g., FIR filter)
• No conflicts: If zero conflicts found, verify this isn't false consensus (AI tools parroting the same source). For factual claims with independent convergence, zero conflicts is acceptable.
• Recommendation without confidence: Always state how sure you are and why