Is the task simple and well-defined?
├── YES: Does it need reasoning or just pattern matching?
│   ├── Pattern matching → GPT-4o-mini, Claude Haiku, Gemini Flash
│   └── Some reasoning → GPT-4o-mini, Claude Sonnet
└── NO: Is it creative, analytical, or requires deep expertise?
    ├── Creative/nuanced → Claude Sonnet/Opus, GPT-4o
    ├── Analytical/research → Claude Sonnet, GPT-4o, Gemini Pro
    └── Coding/technical → Claude Sonnet, GPT-4o, specialized coding models

Weak prompt:
  "Tell me about machine learning"

Better:
  "Explain the difference between supervised and unsupervised
   machine learning with one real-world example of each"

Best:
  "I'm preparing a 5-minute explanation of ML for my marketing team.
   They have no technical background. Explain supervised vs unsupervised
   learning using analogies from everyday life. Keep it under 200 words.
   End with why this matters for our recommendation engine."

Iteration Pattern:

Round 1: Broad request → Get initial output
Round 2: "Good, but make it more [specific adjustment]"
Round 3: "The second paragraph is too technical — simplify"
Round 4: "Add a concrete example for point 3"
Round 5: "Perfect. Now format as [final format]"

Keep refining when:
- The core content is right but needs polish
- You're adjusting tone, length, or format
- Small factual corrections needed

Start a new conversation when:
- The model is stuck in a wrong direction
- Accumulated context is confusing the model
- You've fundamentally changed what you want
- The conversation is very long and responses degrade

The Danger Zone:

                    High Confidence
                         │
         ┌───────────────┼───────────────┐
         │               │               │
         │   CORRECT     │   DANGEROUS   │
         │   AND         │   (confidently│
         │   CONFIDENT   │    wrong)     │
         │               │               │
    ─────┼───────────────┼───────────────┼─────
         │               │               │
         │   UNCERTAIN   │   WRONG AND   │
         │   (says "I'm  │   UNCERTAIN   │
         │    not sure")  │   (at least   │
         │               │    hedges)    │
         │               │               │
         └───────────────┼───────────────┘
                         │
                    Low Confidence

Models are MOST dangerous in the upper-right quadrant:
stating false information with complete confidence.

Common triggers:
- Obscure facts (rare in training data)
- Recent events (after cutoff)
- Specific numbers (dates, statistics, prices)
- Citations (paper titles, authors, URLs)
- Personal/private information

For drafting:
"Write a [type] about [topic]. 
Audience: [who]. 
Tone: [formal/casual/technical/friendly].
Length: [word count or paragraph count].
Key points to cover: [list].
Avoid: [things to exclude]."

For editing:
"Edit this text. Priorities:
1. Cut length by 30% without losing meaning
2. Replace jargon with plain language
3. Make each paragraph start with its key point
4. Flag any claims that need citations

Original: [paste text]"

For summarizing:
"Summarize this [document type] in [format].
Preserve: [key information to keep].
Audience: [who will read the summary].
Length: [constraint].

Source: [paste content]"

Effective AI Research Process:

1: Frame the question precisely
   "What are the tradeoffs between microservices and monoliths
    for a team of 5 engineers building a B2B SaaS product?"

2: Get initial landscape
   "What are the main schools of thought on this topic?
    Who are the key voices? What does recent research say?"

3: Go deeper on promising directions
   "Explain [specific subtopic] in more detail.
    What are the counterarguments?"

4: Synthesize
   "Given everything we've discussed, summarize the key
    decision factors in a comparison table."

5: Verify independently
   Check key claims against primary sources.
   AI is a starting point, not the final word.

Best practices for data tasks:

1: Provide sample data (not just description)
   Show 3-5 rows so the model understands the format

2: Specify edge cases
   "Some dates are MM/DD/YYYY, others are DD-MM-YYYY"

3: Define expected output format
   Show an example of what the result should look like

4: Ask for explanation
   "Write the query AND explain what each part does"

5: Test with known data
   Verify the output against cases where you know the answer

Effective AI Workflow:

1: Identify repetitive cognitive tasks in your day
2: Design a prompt template for each
3: Build a library of proven prompts
4: Chain prompts for multi-step processes
5: Iterate and improve based on output quality

Prompt Library Structure:

/prompts
├── writing/
│   ├── email-professional.md
│   ├── email-followup.md
│   ├── document-summary.md
│   └── meeting-notes.md
├── analysis/
│   ├── swot.md
│   ├── decision-matrix.md
│   └── risk-assessment.md
├── coding/
│   ├── code-review.md
│   ├── test-generation.md
│   └── documentation.md
└── personal/
    ├── weekly-review.md
    ├── learning-plan.md
    └── brainstorm.md

Each file contains:
- The prompt template with [PLACEHOLDERS]
- Usage notes (when to use, which model)
- Example input/output
- Known limitations

Example: Personal Writing Editor

"You are my writing editor. Your job:
- Cut unnecessary words (target: 20% shorter)
- Replace passive voice with active
- Flag vague claims that need specifics
- Maintain my voice (direct, slightly informal, technical)
- Never add flowery language or filler phrases
- If something is unclear, ask rather than guess

Format feedback as:
1. Overall assessment (1 sentence)
2. Specific edits (inline suggestions)
3. Structural suggestions (if any)"

Before sending data to an AI:

1: Check your organization's AI usage policy
2: Classify the data sensitivity
3: Remove or anonymize PII if possible
4: Use enterprise/API tiers (not free consumer tiers)
5: Consider local models for sensitive work
6: Never include credentials, API keys, or secrets
7: Be aware that prompts may be logged for abuse monitoring

AI Usage Policy (Template):

APPROVED TOOLS: [list with data classification levels]
DATA RULES: 
  - Never input [classified data types]
  - Anonymize [data types] before use
  - Use [enterprise tier] for [data types]
QUALITY RULES:
  - AI output must be reviewed before external use
  - Cite AI assistance in [contexts]
  - Human accountable for all AI-assisted decisions
SECURITY:
  - No credentials or secrets in prompts
  - No customer PII without anonymization
  - Report any data leakage incidents

The Right Mental Model:

❌ AI as Oracle:
   "AI said X, so X must be true"
   → Dangerous. Leads to uncritical acceptance.

❌ AI as Threat:
   "AI will replace me, so I won't use it"
   → Counterproductive. Misses real benefits.

✅ AI as Tool:
   "AI helps me think faster and explore more options,
    but I make the decisions and verify the output"
   → Productive. Leverages strengths of both.

The human provides:
- Judgment and taste
- Domain expertise and context
- Ethical reasoning
- Accountability
- Creative direction
- Verification

The AI provides:
- Speed and scale
- Breadth of knowledge
- Pattern recognition
- Tireless iteration
- First drafts and options
- Mechanical transformation

Trust Calibration:

HIGH trust (verify lightly):
- Format transformations (JSON → CSV)
- Grammar and spelling corrections
- Code syntax and formatting
- Summarization of text you've read
- Generating boilerplate/templates

MEDIUM trust (verify key claims):
- Explanations of well-known concepts
- Code for common patterns
- Analysis of data you provided
- Comparisons of well-documented options

LOW trust (verify everything):
- Specific facts, dates, statistics
- Citations and references
- Claims about recent events
- Medical, legal, or financial advice
- Security-sensitive code
- Anything where being wrong has consequences

Effective learning prompt:

"I'm learning [topic]. My current level: [beginner/intermediate/advanced].
Don't give me the answer directly. Instead:
1. Ask me what I already know
2. Present a problem that requires this concept
3. Guide me with hints if I'm stuck
4. Correct my reasoning when I'm wrong
5. Only explain fully after I've attempted it"

"Explain [concept] at three levels:
1. For a 10-year-old (core intuition only)
2. For a college student (mechanisms and tradeoffs)
3. For a practitioner (implementation details and edge cases)"

Email drafting template:

"Draft an email with these parameters:
- To: [role/relationship, not name]
- Context: [situation]
- Goal: [what I want to happen after they read this]
- Tone: [professional/casual/urgent/diplomatic]
- Constraints: [length, things to avoid, cultural notes]
- My draft/notes: [rough points to include]"

Meeting summary prompt:

"Summarize these meeting notes into:
1. DECISIONS MADE (what was agreed)
2. ACTION ITEMS (who does what by when)
3. OPEN QUESTIONS (unresolved, needs follow-up)
4. KEY DISCUSSION POINTS (2-3 sentences max)

Format action items as: [Owner] — [Task] — [Deadline]

Notes: [paste raw notes]"

Prompt: "I'm deciding between [Option A], [Option B], and [Option C] for [context].

Criteria (weighted):
- Cost (weight: 3)
- Time to implement (weight: 2)
- Long-term maintainability (weight: 3)
- Team skill match (weight: 2)
- Risk level (weight: 1)

Score each option 1-5 per criterion. Show weighted totals.
Then tell me what this analysis DOESN'T capture that I should consider."

"I'm leaning toward [option]. Your job is to argue AGAINST it as strongly
as possible. Be specific, cite real risks, and don't hold back. I want to
hear the strongest case for NOT doing this.

After your argument, rate on 1-10 how compelling you find your own
counter-argument."

What happens as context fills up:

Tokens 0-25%:    Peak performance. Model has full attention.
Tokens 25-50%:   Still good. Minor degradation on early details.
Tokens 50-75%:   Middle content may get less attention ("lost in the middle" effect).
Tokens 75-100%:  Early conversation details may be forgotten. Quality drops.
Beyond limit:    Oldest content silently dropped or conversation rejected.

Prompt template format:

---
name: Technical Document Review
model: Any mid-tier+
use-when: Reviewing technical docs before publishing
---

"Review this technical document for:
1. Factual accuracy (flag anything you're uncertain about)
2. Clarity (highlight jargon that needs definition)
3. Completeness (what questions would a reader still have?)
4. Structure (does the flow make sense?)

Audience: [AUDIENCE]
Document type: [TYPE]
Priority concern: [SPECIFIC FOCUS]

Document:
[PASTE CONTENT]"

When a new model releases, test it on YOUR tasks:

1. Keep 5-10 "benchmark prompts" — tasks you run regularly
2. Run them on the new model
3. Compare: quality, speed, cost
4. Switch only if meaningfully better for YOUR use cases
5. Don't chase every release — stability has value

The meta-skill progression:

Beginner:  "AI can do everything" → overreliance, disappointment
           "AI can do nothing" → underuse, missed value

Intermediate: "AI is good at X but bad at Y" → appropriate delegation

Advanced: "For THIS specific task, with THIS context, AI will likely
          [succeed/struggle] because [reasoning about why]"
          → calibrated, adaptive, efficient use

Expert: Automatically routes tasks to AI or human effort based on
        intuition built from hundreds of interactions. Rarely surprised
        by AI failures. Quickly adapts when capabilities change.