| name | ai-systems |
| description | Builds production-grade Claude Projects and CustomGPTs, improves AI output quality using meta-thinking prompts, and architects self-improving AI systems. Use proactively when a founder wants to build a reliable AI workflow, automate a repeatable task with AI, improve the quality of AI outputs they're getting, feels stuck with generic AI responses, or wants to build internal tools powered by Claude. Trigger for "build a Claude project", "my AI outputs are generic", "how do I make Claude more specific", "build an AI workflow", "automate this with AI", or "my prompts aren't working". |
| used_by | ["ceo-thinking-partner","workflow-architect","build-quality-auditor"] |
AI Systems — Founder OS
Builds production-grade AI workflows and improves output quality. Two layers: the Meta-Thinking layer (getting better outputs from any prompt) and the Production Architecture layer (building reliable AI systems).
Layer 1: Meta-Thinking Prompts (Output Quality)
The 3 Stuck Points
| Stuck Point | Symptom | Root Cause |
|---|
| Blank Chat Paralysis | Know the outcome, can't frame the request | Haven't given AI enough context to work from |
| Generic Output Loop | Clear question, vague/textbook response | Right question, wrong altitude or missing constraints |
| Solving the Wrong Problem | Polished output, but wrong answer | AI optimized for what you said, not what you need |
The 9 Meta-Thinking Prompts
For Stuck Point 1 — Let AI interrogate you first:
Prompt 1: Context Extraction
Before answering, ask me 7-10 questions that would help you understand
my situation well enough to give strategic, not generic, advice.
Don't answer yet — just ask the questions.
Prompt 2: First Principles Reset
Break this down using first principles.
What am I assuming that should be questioned?
What's the real problem underneath the problem I described?
Prompt 3: Gap Identifier
What information do you need from me to move from a generic framework
to something implementable for my specific situation?
List exactly what's missing.
For Stuck Point 2 — Fix the altitude:
Prompt 4: Altitude Diagnostic
Am I asking for strategy when I need tactics?
Or tactics when I need strategy?
What altitude should this conversation be at, and why?
Altitude levels:
- 30,000 ft: Vision ("What should we become?")
- 20,000 ft: Strategy ("How will we win?")
- 10,000 ft: Tactics ("What specific actions?")
- Ground: Execution ("Step-by-step instructions")
Prompt 5: Missing Context Probe
What would make your last response completely wrong for my situation?
What assumptions did you make that I should validate?
Prompt 6: Inversion Technique
What would make this the worst possible plan?
List every way this could fail, then help me address each failure mode.
For Stuck Point 3 — Stress-test the output:
Prompt 7: Usefulness Audit
Rate what you just gave me 1-10 for actual usefulness — not theoretical quality,
but practical usefulness for a founder at [my stage] trying to [specific goal].
What specific changes would make it a 10?
Prompt 8: Devil's Advocate
Play devil's advocate on your own response.
What are the 3-5 strongest criticisms of this approach?
Don't defend it — genuinely challenge it.
Prompt 9: Rejection Test
Imagine I'm the CEO reviewing this output. I'm about to reject it.
What are the most likely reasons I'd reject it?
Then rewrite it to survive that rejection.
Power Stack: Run in Sequence
For any high-stakes output (investor pitch, pricing strategy, positioning):
- Start with Prompt 1 (Context Extraction) — let AI ask you questions first
- After answering, run Prompt 4 (Altitude Diagnostic) — make sure you're at the right level
- After getting output, run Prompt 9 (Rejection Test) — harden the output
This sequence turns a 20-minute prompt session into a 45-minute session that produces work worth 10x more.
Layer 2: Production Architecture (Building AI Systems)
The 7-Step Production Framework
For building reliable Claude Projects, CustomGPTs, or any AI workflow that runs repeatedly.
Step 1: Brainstorm-to-Blueprint (4 phases)
Phase A — Problem Mapping:
What specific task does this system need to do?
What's the input? What's the expected output?
What does success look like (measurable)?
What does failure look like?
Phase B — Workflow Prototyping:
Map the steps a human expert would take
Identify decision points (where would a human choose?)
List what information the system needs at each step
Phase C — Refinement:
Where will AI hallucinate or go generic?
Where do you need deterministic output?
Where can AI be creative vs. where must it be precise?
Phase D — Instruction Generation:
Write the system prompt based on the above
Step 2: Workflow Architecture — Choose Your Type
| Type | Use When | Example |
|---|
| Single-Step Agent | One clear input → one clear output | "Classify this customer email" |
| Multi-Step Agent | Sequential process with dependencies | "Research → analyze → draft" |
| Broader Use-Case Agent | Handles multiple related task types | "Head of Growth assistant" |
Step 3: The 10-Point Testing Protocol
Run all 10 before deploying any AI system:
| Test | Target | What to Check |
|---|
| 1. Happy Path | 3 perfect scenarios | Does it work when everything is right? |
| 2. Edge Cases | Missing/unusual inputs | What breaks when input isn't clean? |
| 3. Failure Modes | Contradictions, impossible constraints | How does it fail? Gracefully or badly? |
| 4. Format Compliance | 90%+ across 5 executions | Does it follow output format consistently? |
| 5. Knowledge Retrieval | 10 queries with known answers | Does it know what it should know? |
| 6. Consistency | Same input x3 → 85%+ similarity | Same question = same answer? |
| 7. Cross-Reference Integrity | Multi-part outputs | Do parts of the output contradict each other? |
| 8. Speed | Benchmark | Is it fast enough for the use case? |
| 9. Self-Correction | Prompt it to check itself | Does it catch its own errors? |
| 10. User Acceptance | 3 people x5 executions | Would real users trust this output? |
Step 4: 5 Deterministic Output Techniques
When you need the AI to output the same structure every time:
- Structured Output Formats — Specify JSON schema or exact template in system prompt
- Explicit Constraints — "Exactly 3 bullet points. Each bullet: 10-15 words. No exceptions."
- Built-In Validation — "Before responding, check: did I include X, Y, Z? If not, add them."
- Few-Shot Examples — Include 2-3 perfect output examples in the system prompt
- Constraint-Based Generation — "Output must: [5 absolute requirements]. Count each one. Confirm all 5 are met."
Step 5: Self-Improvement System
WEEKLY EVOLUTION PROTOCOL
After every 10 executions, run this audit:
1. Performance Analysis:
- Average quality score (1-10) across last 10 outputs
- Which outputs scored 9+? What did they have in common?
- Which outputs scored below 7? What went wrong?
2. Pattern Recognition:
"The 9+ outputs all contained: [patterns]"
"The <7 outputs all lacked: [patterns]"
3. Improvement Proposal:
"Based on this analysis, the system prompt change that would most improve
average quality is: [specific change]"
4. A/B Test:
Run current prompt vs. proposed change on 5 identical inputs
Score both versions
Keep the winner
System Prompt Architecture Template
# [SYSTEM NAME]
## Role
You are a [specific role] for [specific user type].
Your primary job: [one clear task]
You are NOT responsible for: [explicit scope boundaries]
## Input Format
You will receive: [describe exact input format]
If input is missing [X], do: [fallback behavior]
## Output Format
Always output in this exact structure:
[template with explicit labels]
NEVER output:
- [anti-pattern 1]
- [anti-pattern 2]
## Examples
Input: [example input]
Output: [perfect example output]
Input: [edge case input]
Output: [correct edge case output]
## Validation Checklist
Before responding, verify:
[ ] Output follows the exact format specified
[ ] All required sections are present
[ ] No prohibited content included
[ ] Response addresses the specific input (not generic version)
The Virality Prompt Suite
For when you need to engineer shareable content.
Master Prompt: ViralGuru
You are ViralGuru, an expert virality coach.
My inputs:
- Target audience: [who should share this]
- Objective: [awareness / signups / downloads / shares]
- Analytics snapshot: [current engagement data if available]
- Content draft: [paste draft or describe concept]
Deliver:
1. Quick-Glance Summary — 3 bullets: what's working, what's not, top fix
2. Diagnosis Matrix — why this content does or doesn't spread
3. Platform Recommendation — best platform for this content and why
4. Refined Script/Copy — improved version
5. Algorithm Optimizations — platform-specific tactical improvements
6. Cross-Platform Repurposing — how to adapt for 2 other platforms
7. Metrics & A/B Plan — what to test and how to measure
Viral Coefficient Model
Viral Coefficient = Invite Rate x Conversion Rate x Retention Rate
B2B tools target: 0.2–0.6
User-driven platforms: 0.5–1.0+
Diagnose your current VC:
- Current invite rate: [X]% of users invite others
- Current conversion rate: [X]% of invites convert
- Current retention: [X]% stay active at 30 days
Which lever has the most room for improvement?
Design one experiment to improve that lever by 20%.
Viral Hook Formula
Generate 10 headline/hook variations for this content:
[paste content brief or draft opening]
For each variation, tag:
- Emotion triggered (curiosity / fear / aspiration / surprise / belonging)
- Psychological mechanism (FOMO / social proof / authority / novelty / utility)
- Predicted engagement type (save / share / comment / click)
Rank by estimated shareability for [target platform].
Quick Wins: Prompts That Fix Common AI Problems
When AI gives generic frameworks:
Stop giving me frameworks. Give me the specific answer for a [stage]
[business type] with [specific constraint].
If you don't have enough information, ask me what you need.
When AI is too positive:
Be a skeptical investor reviewing this.
What are the 3 biggest red flags or weaknesses?
Don't soften them.
When AI gives theory instead of action:
Assume I'm executing this starting tomorrow morning.
What are my first 3 actions, in order?
What do I need before I can start each one?
When AI output is too long:
Distill your last response to the 20% that contains 80% of the value.
Cut everything that doesn't directly affect what I do next.
Integration with Other Skills
- AI Systems improves → every other skill (better prompts = better outputs everywhere)
- Use for →
skills/public/head-of-growth/SKILL.md (build automated weekly audit system)
- Use for →
skills/public/content-trust/SKILL.md (build content audit workflow)
- Foundation for → any repeatable founder task that runs weekly or monthly
