diegosouzapw Understand-Ask-AI 150 Protocol
Core Principle: When internal knowledge isn't enough, consult external AI expertise. Build quality prompts that get quality answers. Validate what comes back.
What This Skill Does
When you invoke this skill, you're asking AI to:
- Verify internal exhaustion — Confirm internal sources are depleted
- Build quality prompt — Use Prompt150 formula
- Target the right model — Choose appropriate external AI
- Send and receive — Get external AI response
- Validate response — Check credibility and integrate
The Prompt150 Formula
PROMPT150 = Context150 + Query + Method + Style
┌─────────────────────────────────────────────────────┐
│ CONTEXT150 (100% core facts + 50% supporting) │
│ ├── Core situation and background │
│ ├── Key data points and constraints │
│ ├── Actions taken so far │
│ └── What we DON'T know (explicit unknowns) │
├─────────────────────────────────────────────────────┤
│ QUERY (specific, answerable question) │
│ ├── Single, focused question │
│ ├── Answerable with available information │
│ └── NOT vague ("analyze this" ❌) │
├─────────────────────────────────────────────────────┤
│ METHOD (how to approach) │
│ ├── "Verify claims against real data" │
│ ├── "Provide confidence levels (%)" │
│ ├── "Cite specific precedents with sources" │
│ └── "Be conservative if data insufficient" │
├─────────────────────────────────────────────────────┤
│ STYLE (output format) │
│ ├── Structured response (sections, tables) │
│ ├── Confidence % for each claim │
│ ├── Facts vs assumptions clearly separated │
│ └── Actionable recommendations if applicable │
└─────────────────────────────────────────────────────┘
When to Use This Skill
TRIGGER CONDITIONS:
- Loop150 exhausted internally — All workspace files searched, no data
- Need real-world precedents — Case studies, actual outcomes
- Need current information — Data after knowledge cutoff
- Need statistical data — Industry patterns, benchmarks
- Need scenario modeling — Complex decision trees
- Confidence <75% — Cannot reach 90% with internal data
DO NOT USE FOR:
- ❌ Facts already in workspace (use grep/search)
- ❌ Simple calculations (do yourself)
- ❌ Questions answerable by reading files
- ❌ First resort (always try internal first)
Execution Protocol
Step 1: EXHAUSTION VERIFICATION
🔍 **INTERNAL EXHAUSTION CHECK**
**Internal Sources Tried:**
- [ ] Codebase search: [Results]
- [ ] Documentation: [Results]
- [ ] Git history: [Results]
- [ ] Project files: [Results]
**Current Confidence:** [X]%
**Gap Identified:** [What we need but don't have]
**External Query Justified:** ✅ Yes | ❌ No (try more internal)
Step 2: PROMPT CONSTRUCTION
Build using Prompt150 formula:
📝 **PROMPT150 CONSTRUCTION**
**CONTEXT150:**
[100% core facts]
- Situation: [What's happening]
- Data points: [Key numbers/facts]
- Constraints: [Limits and requirements]
- Actions taken: [What's been done]
[50% supporting details]
- Background: [Broader context]
- Unknowns: [What we explicitly don't know]
- Stakes: [Why this matters]
**QUERY:**
"[Specific, answerable question]"
Example good queries:
✅ "What were timelines for SSN breach cases with 5-500 affected?"
✅ "What is typical regulator response time for consumer complaints?"
❌ "Analyze my case" (too vague)
❌ "What should I do?" (too broad)
**METHOD:**
- Use Loop150-like verification
- Provide confidence levels (%)
- Cite real precedents with sources
- Be conservative if data insufficient
**STYLE:**
- Structured sections/tables
- Confidence % on each claim
- Facts vs assumptions separated
- Actionable recommendations
Step 3: MODEL SELECTION
🤖 **MODEL SELECTION**
**Query Type:** [Research/Reasoning/Coding/Creative]
**Recommended Model:**
- Complex reasoning: ChatGPT-4/Claude (thinking models)
- Coding help: Claude/GPT-4
- Research synthesis: Perplexity/ChatGPT with browsing
- Current events: Models with web access
**Selected:** [Model name]
**Reason:** [Why this model]
Step 4: USER APPROVAL
🌐 **ASK-AI 150 REQUEST**
**Justification:** Internal sources exhausted
**Confidence Gap:** [Current X%] → [Need Y%]
**Prompt Preview:**
"""
[Full Prompt150 to be sent]
"""
**Target Model:** [Model name]
**Approve external query?** (Yes / No / Modify)
Step 5: SEND AND RECEIVE
Execute the query and capture response.
Step 6: RESPONSE VALIDATION
✅ **RESPONSE VALIDATION**
**Source Credibility:**
- Model used: [Name]
- Claims verifiable: [Yes/Partially/No]
- Confidence stated: [Yes/No]
**Content Assessment:**
- Answers query: ✅ Yes | ⚠️ Partially | ❌ No
- Facts vs opinions: [Clear/Mixed/Unclear]
- Actionable: [Yes/Needs interpretation/No]
**Integration:**
- Confidence boost: [+X% → new total]
- Gaps remaining: [What's still unknown]
- Action items: [What to do with this info]
Output Format
Request:
🌐 **ASK-AI 150 REQUEST**
**Internal Sources Exhausted:** ✅
**Current Confidence:** [X]%
**Gap:** [What we need]
**Prompt150:**
---
CONTEXT:
[Context150 content]
QUERY:
[Specific question]
METHOD:
- Verify against real data
- Provide confidence %
- Cite real precedents
- Be conservative
STYLE:
- Structured response
- Confidence per claim
- Facts vs assumptions
---
**Target:** [AI Model]
**Approve?**
Response integration:
🌐 **ASK-AI 150 RESPONSE INTEGRATED**
**Query:** [What was asked]
**Model:** [What answered]
**Key Findings:**
1. [Finding 1] — [Confidence %]
2. [Finding 2] — [Confidence %]
3. [Finding 3] — [Confidence %]
**Validation:**
├── Claims verifiable: [Yes/Partially]
├── Sources cited: [Yes/No]
└── Consistent with known facts: [Yes/No]
**Confidence Update:** [X%] → [Y%]
**Remaining Gaps:** [What's still unknown]
**Next Steps:** [How to use this information]
Operational Rules
- INTERNAL FIRST: Never skip internal research
- JUSTIFY EXTERNAL: Document why internal is insufficient
- QUALITY PROMPTS: Use full Prompt150 formula
- USER APPROVAL: Get permission before external query
- VALIDATE RESPONSE: Don't blindly trust external AI
- DOCUMENT INTEGRATION: Log what was learned and confidence change
Failure Modes & Recovery
| Failure | Detection | Recovery |
|---|---|---|
| Premature External | Didn't exhaust internal | Complete internal search first |
| Poor Prompt | Vague, context-poor | Reformulate with Prompt150 |
| Unreliable Response | Unverifiable claims | Find better sources or reject |
| No Validation | Used response blindly | Cross-check before acting |
Examples
❌ Bad Ask-AI
Query: "How to build web app?"
Context: [None provided]
Result: Got generic outdated advice, wasted time
✅ Good Ask-AI 150
🌐 ASK-AI 150 REQUEST
Internal Sources Exhausted: ✅
- Checked all project docs
- Searched codebase
- No breach precedent data found
Current Confidence: 65%
Gap: Need real-world timeline data for similar cases
Prompt150:
---
CONTEXT:
- SSN data breach affecting 47 individuals
- Breach discovered: 2024-03-15, Notified: 2024-06-20 (97 days)
- Washington State (RCW 19.255.010 requires 45 days)
- HIPAA-covered entity (45 CFR §164.524)
- Complaints filed with AG and HHS/OCR
QUERY:
"What were actual timelines and outcomes for SSN data
breach cases similar to this (5-500 people affected) in
the past 5 years?"
METHOD:
- Cite specific cases with dates and outcomes
- Provide confidence levels for predictions
- Distinguish confirmed data from estimates
- Be conservative if data insufficient
STYLE:
- Table format for cases
- Confidence % on predictions
- Separate facts from projections
---
Target: ChatGPT-4 (with web search)
Approve? Yes
---
🌐 ASK-AI 150 RESPONSE INTEGRATED
Key Findings:
1. Similar cases settled in 6-18 months (75% confidence)
2. Typical per-person compensation: $100-500 (70% confidence)
3. AG response time: 30-90 days (80% confidence)
Validation:
├── Claims verifiable: Partially (cited 3 real cases)
├── Sources cited: Yes (court records referenced)
└── Consistent with known facts: Yes
Confidence Update: 65% → 82%
Remaining Gaps: Specific WA state precedents
Next Steps: Use timeline estimates for planning,
continue monitoring for WA-specific cases
Relationship to Other Skills
- research-deep-150 → Exhausts internal sources
- ask-ai-150 → Consults external when internal insufficient
- proof-grade-150 → Validates external information
Remember: Ask-AI is like calling a consultant — you don't call before doing your homework, you come with specific questions, and you verify their advice. Quality in, quality out.