Book Summary
A chapter-by-chapter quick-reference for AI Is Your Intern, Not Your CEO. Each summary tells you when you need the chapter, gives you the core framework or decision in a scannable format, and flags the trap most people fall into. For the connected decision flowcharts, see Appendix D.
Part 1: The Mental Model
Chapter 1: Why Your AI Strategy Is Backwards
When you need this: You’re about to invest time, money, or credibility in an AI initiative and want to avoid the mistakes behind the 95% of organizations getting zero return from AI.
ARE YOU APPROACHING AI IN THE RIGHT ORDER?
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Most people ask: You should ask:
1. Which tool? 1. Where do humans need
2. What to automate? better information?
3. How fast? 2. How will we verify output?
3. How does AI earn expanded
responsibility over time?Quick audit: Look at the last 3 times you used AI output in real work. For each: Did anyone review it? Would you trust this from a new hire without review? What if it had a significant error? If the answers concern you, start there.
The trap: Three cognitive traps cause predictable AI failures — automation bias (professional-looking output gets accepted without verification), the overconfidence effect (higher AI literacy = more overconfidence, not less), and the deploy-now trap (competitive pressure launches pilots without success criteria, and sunk cost keeps them alive in pilot purgatory).
Chapter 2: The Intern Model Explained
When you need this: You want a mental model for working with AI that prevents the common failures — delegation without review, trust without evidence, use without accountability.
THE INTERN MODEL — 4 PRINCIPLES
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
1. CLEAR TASKS → Use SCOPE:
Specific outcome / Constraints / Output format /
Prior context / Evaluation criteria
2. REVIEW BEFORE SHIPPING → Match to stakes:
Scan → Spot-check → Deep review → Rewrite
3. INCREMENTAL TRUST → Climb the Trust Ladder:
Rung 1: Assisted drafting (full review)
Rung 2: Supervised output (detailed review)
Rung 3: Spot-checked production
Rung 4: Exception-based review
4. FEEDBACK LOOPS → Capture / Categorize /
Correct / Confirm → repeatThe trap: People skip straight to Rung 3 or 4. They define tasks clearly but skip review because “the output looks good.” Without review data, there’s no feedback loop. Without feedback, nothing improves. Trust is task-specific — AI at Rung 3 for social posts may be Rung 1 for contracts.
Chapter 3: The CEO’s Information Problem
When you need this: You’re trying to figure out what AI is actually for in your work — not the hype, but the real value.
REFRAME: WHAT PROBLEM DOES AI SOLVE?
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Wrong question: "How can AI do this task for me?"
Right question: "How can I get better information faster?"
Only 30 hours of a typical 40-hour week are
productive (APQC) — the rest lost to email, meetings,
and hunting across ~367 apps (Forrester). AI attacks
that tax.
THREE INFORMATION PROBLEMS AI SOLVES:
1. Volume — too much to process manually
2. Velocity — changes faster than you can track
3. Variety — scattered across formats and sourcesAction: Identify your 3 most important recurring decisions. For each: What information do you wish you had? Where does it exist? How long does it take to get today? The decision with the clearest gap is your starting point.
The trap: The automation mindset (“How can AI do this task for me?”) focuses on replacement — it’s the mindset behind the 95% of organizations getting zero return. The decision-support mindset (“How can AI help me decide better?”) is where the value is.
- Dept leaders: Volume/variety problems — tickets, feedback, reports piling up across your team.
- ICs: Missing context from meetings, channels, other teams you don’t have time to follow.
- CEOs: Decision bottleneck — you have the least time to research and the most decisions to make.
- Senior leaders: Fragmented strategic info across reports, meetings, and regional updates.
Part 2: The Audit
Chapter 4: Mapping Your Decision Landscape
When you need this: You think you know where AI would help, but you haven’t actually tracked where your time goes. Most people are wrong about this.
THE 10-DECISION SPRINT (10 minutes)
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
1. List 10 decisions you made today
2. Note what info each required
3. Note how long getting that info took
4. ★ Star the 3 where better/faster info
would have helped most
Those 3 starred decisions → bring to Ch 5.
WATCH FOR DECISION BLINDNESS:
• Routine blindness — decisions so frequent
they don't register as decisions
• Bundled blindness — multiple decisions
disguised as a single task
• Reactive blindness — decisions hidden
in responses to others' requestsThe trap: I estimated I made maybe 30 decisions a day; one tracking day counted closer to 200. Kenji, an operations director, predicted 20-25 decisions per week — the actual count was 132, and his $45K AI tool targeted strategic planning (8% of decisions) while ignoring the triage and allocation consuming 50%. Map first, invest second.
Chapter 5: Scoring Your Opportunities
When you need this: You have a list of possible AI tasks and need to figure out which one to start with.
IS THIS TASK A GOOD AI CANDIDATE?
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Score each dimension 1–5, then MULTIPLY:
Frequent — How often? (daily = gold)
Forgiving — What if AI's wrong? (catchable = good)
Clear — Inputs/outputs defined? (template = best)
Contained — Can it stand alone? (fewer deps = better)
400–625 → ✅ Start here.
200–399 → ✅ Second wave.
100–199 → ⚠ Proceed with caution.
Below 100 → ❌ Not yet.
Then validate: Does this affect outcomes you care about?
High FFCC + High business value = best starting point.The trap: We chase impressive projects while ignoring daily tasks that score 40x higher. I wanted to build campaign strategy support first (score: 12) while message triage — the boring queue that became the $700K workflow — scored 500. Trust the math, not the glamour.
- Dept leaders: Score your team’s top 5 recurring tasks — the ones eating hours across multiple people.
- ICs: Score your personal daily friction points — tasks you dread that follow a predictable pattern.
- CEOs: Score decisions that repeat but don’t need your judgment — just your time.
- Senior leaders: Score the information-gathering that precedes your real decisions, not the decisions themselves.
Chapter 6: Choosing Where to Start
When you need this: You’ve scored your opportunities and have 2-3 strong candidates. Now you need to stop analyzing and commit.
MAKE THE CHOICE
━━━━━━━━━━━━━━━
Is your top FFCC scorer within 20% of #2?
│
├─ No ──→ Pick the top scorer. Done.
│
└─ Yes ─→ Apply tiebreakers:
1. Energy — which one interests you most?
2. Quick Win — which shows results in ≤2 weeks?
3. Accessible — which can you start with
tools you already have?
THEN WRITE A COMMITMENT CONTRACT:
"My first AI workflow is: ____________
Starting on: [date within 7 days]
Success looks like: ____________
I'll evaluate on: [2-4 weeks out]"The trap: Analysis paralysis disguised as thoroughness. My own scores were 500/400/320 — the top wasn’t even close — and I still lost two weeks second-guessing it. The typical version is six weeks deliberating among close scorers, then committing to the original top scorer anyway, while the person who just picked and started is already on their second workflow.
Part 3: Building Workflows
Chapter 7: Your First Workflow
When you need this: You’ve scored a task and it’s a good candidate. Now you need to turn random AI conversations into a reliable, repeatable process.
BUILD YOUR WORKFLOW
━━━━━━━━━━━━━━━━━━
1. TRIGGER — What kicks it off?
Time-based or Event-based (automatic = reliable) → Manual (weakest)
2. INPUTS — What does AI need?
Content + Context + Constraints + Examples
3. PROCESSING — What does AI do?
Pick ONE verb: Draft / Summarize / Categorize /
Extract / Analyze / Compare / Translate
Define the output format.
4. REVIEW — What do you check?
Accuracy / Tone / Completeness / Appropriateness
Set a time budget: ___ minutes per output.
5. ACTION — What happens after approval?
Send / Publish / Save / Escalate / Assign
⚠ No action = no workflow. Just a conversation.The trap: Most workflow failures come from Step 2 (weak inputs), not Step 3 (bad prompts). If output is consistently off, fix what you feed it. Also: run it 3 times before judging — the first run always feels rough.
- Dept leaders: Triggers are time-based (weekly summaries). Actions flow to your team or upward.
- ICs: Triggers are event-based (new assignment). Actions are personal deliverables.
- CEOs: Triggers are daily (morning briefing). Actions are decisions and delegation.
- Senior leaders: Triggers are cyclical (pre-board meeting). Actions are org communications.
Chapter 8: Role-Specific Applications
When you need this: You’ve built a workflow using Chapter 7’s structure, but it doesn’t feel right for your role. Same five components — different implementation.
WHICH WORKFLOW PATTERN FITS YOUR ROLE?
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Dept Leaders → COORDINATION & VISIBILITY
Focus: Aggregation, pattern detection, exception flagging
Sources: Project tools, standups, cross-functional updates
Patterns: Status aggregation, performance summary,
communication standardization
ICs → PERSONAL LEVERAGE
Focus: Preparation, drafting, quality enhancement
Sources: Project requirements, client comms, research
Patterns: Task prep, draft generation, quality checking
CEOs → DECISION SUPPORT
Focus: Synthesis, priority identification
Sources: Email, calendar, metrics, key project updates
Patterns: Research synthesis, communication drafting,
decision preparation
Senior Leaders → STRATEGIC & ORGANIZATIONAL
Focus: Strategic analysis, communication scaling
Sources: Market intel, org signals, strategic reports
Patterns: Intelligence synthesis, meeting prep,
policy documentationQuick test: Does your action component match your role? If you’re a manager but all actions are personal (saving to your own folders), the workflow is IC-shaped. If you’re an IC but all actions involve distributing to others, you’re doing coordination work.
Chapter 9: Designing Inputs
When you need this: Your workflow exists but the output is generic, inconsistent, or requires heavy editing. The fix is almost always in the input, not the prompt.
DIAGNOSE YOUR INPUT QUALITY
━━━━━━━━━━━━━━━━━━━━━━━━━━━
Output is generic?
→ Missing CONTEXT. Add: who, history, why now.
Output is wrong format/tone?
→ Missing CONSTRAINTS. Add: length, format, tone.
Output varies wildly between runs?
→ AMBIGUOUS inputs. AI fills gaps with guesses.
Expert Test: hand your input to a colleague
with no other context. If they ask clarifying
questions, those reveal your gaps.
Output is close but lacks your voice?
→ Treat the draft as raw material — the Review
Spectrum's Rewrite level. AI drafts, you rewrite.
BUILD BETTER INPUTS IN 4 LAYERS:
1. Context — who, history, constraints
2. Content — actual material to process
3. Constraints — format, tone, inclusions/exclusions
4. Request — what to do with the aboveThe trap: People spend 80% of effort on prompt wording and 20% on input quality. Results come from the inverse. A mediocre prompt with excellent context beats a perfect prompt with thin context. Every time.
- Dept leaders: Biggest gap is usually aggregation — you have data across your team but aren’t feeding enough of it.
- ICs: Biggest gap is implicit expertise — things you know about the client that you haven’t written down.
- CEOs: Biggest gap is signal vs. noise — you have breadth but need to curate what’s relevant.
- Senior leaders: Biggest gap is the strategy-to-operations bridge — connecting high-level context to the specific task.
Chapter 10: Output Quality
When you need this: AI produced something. You need to evaluate it quickly and know whether to use it, edit it, or throw it out.
THE THREE QUESTIONS (before diving in)
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
1. Does it achieve the objective?
2. Is it accurate and appropriate?
3. Does it need my expertise or just my approval?
FOUR QUALITY TIERS:
Tier 1: Use as-is → ✅ Ship it.
Tier 2: Light edit → ✅ Minor fixes, ship.
Tier 3: Substantial revision → ⚠ Major rework needed.
Tier 4: Regenerate → ❌ Start over. Fix inputs.
TIME BUDGET CHECK:
Set a max review time per output type
(email: 2-3 min, summary: 3-5, report: 5-7).
Hit the limit? Stop and decide: use as-is,
regenerate, or escalate.
Review time ≈ creation time → not a
workflow, a hobby.The trap: Two opposite failures — over-editing (polishing every paragraph, eliminating all time savings) and under-reviewing (rubber-stamping everything, creating errors). If you’re consistently at Tier 3-4, that’s an input problem (Ch 9), not an output problem.
Part 4: Permissions & Trust
Chapter 11: Risk Assessment
When you need this: Before expanding AI into a new area — especially one involving sensitive data, external communications, or decisions that affect other people.
THE PERMISSION FRAMEWORK — 3 QUESTIONS
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Q1: What's the SPECIFIC benefit?
Can you say it in one sentence with a number?
Pass: "Saves 10 hours/week on report prep."
Fail: "Improves efficiency."
→ If you can't pass, STOP. Define the benefit.
Q2: What's the ACTUAL exposure?
Internal, non-sensitive ──→ Low
Internal, some sensitive ─→ Medium
Customer-facing, indirect → Medium-High
Customer-facing, direct ──→ High
Regulated / PII / HR ─────→ Very high
Q3: What are the FAILURE MODES?
Name the 3 most likely. For each:
Detected before it ships? → Good
Detected within hours? → Acceptable
Only when someone complains? → ⚠ Danger
Only when sued? → 🛑 Stop
DECISION:
✅ Proceed → Low exposure, recoverable
⚠ Proceed with safeguards → Medium exposure.
Name the safeguards specifically
🔶 Pilot first → High exposure, test limited
🛑 Decline → Exposure too high, failure
modes unacceptable, or benefit insufficientThe trap: “Proceed with safeguards” without naming the specific safeguards is not a decision — it’s a deferral of responsibility.
Chapter 12: Data Access Decisions
When you need this: You’re deciding what data to give AI access to. More data makes it more useful — and increases exposure. You need the right tradeoff.
DATA ACCESS: START WITH MINIMUM VIABLE ACCESS
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
For EACH data source, ask:
1. How much does this data improve utility?
2. How much does it increase exposure?
3. Is the marginal value worth the marginal risk?
FOUR-LEVEL CLASSIFICATION:
Public → Allowed. No restrictions needed.
Internal → Allowed. Standard access logging.
Confidential → Case-by-case. Document justification.
Restricted → Exceptional only. Enhanced controls +
explicit approval required.
DEFAULT STARTING POINTS:
Customer data → Account summaries, not full records
Communications → Specific inputs, not archive access
Financial data → Dashboards, not transaction detail
HR data → Generally avoid. Legal review if needed.The trap: “Give AI access to everything so it can help us” is always the wrong starting point. 80% of value usually comes from 20% of the data. You can always expand access later — clawing it back after an incident is much harder.
Chapter 13: Career Protection
When you need this: You’re using AI regularly and want to make sure it builds your reputation instead of putting it at risk.
THE ACCOUNTABILITY ASYMMETRY
━━━━━━━━━━━━━━━━━━━━━━━━━━━
When AI-assisted work succeeds:
→ Credit is organizational ("We're using AI")
When it fails:
→ Blame is individual ("Why didn't YOU catch this?")
This isn't malicious. AI can't be held accountable.
You can.
THREE TRAPS:
1. Automation Trap — "The AI did it" is never a defense
2. Efficiency Trap — time saved becomes evidence of
insufficient review when problems emerge
3. Precedent Trap — early successes create inflated
reliability expectations
PROTECT YOURSELF: AI DECISION LOG
For each AI-assisted decision, record:
• What AI helped with
• What review you conducted
• What limitations you noted
• What decision you made
Target: 2-3 minutes per entry.Red flags — when to refuse AI: Stakes too high, review time insufficient, documentation impossible, expertise gap too large, or pressure to skip oversight.
Part 5: Sustainable Review
Chapter 14: Calibrating Scrutiny
When you need this: AI produced output. You need to decide how much time to spend checking it before you use it.
CALIBRATE YOUR REVIEW LEVEL
━━━━━━━━━━━━━━━━━━━━━━━━━━━
What are the stakes?
Low (internal, reversible) → SCAN (10-30 sec)
Medium (some external visibility) → SPOT-CHECK (2-5 min)
High (external-facing, reputational) → DEEP REVIEW (15-30 min)
Critical (regulatory, legal) → DEEP REVIEW or REWRITE
(30+ min; AI as raw
material only)
NOW ADJUST:
New workflow or task type? → Start at DEEP REVIEW.
AI less reliable at this task? → One level deeper.
Rushed, tired, distracted? → One level deeper or wait.
10-20 clean iterations? → May step down one level.
⚠ Never: reliability overrides critical stakes.
Even highly reliable workflows producing critical
outputs still require deep review.The trap: The author reviewed 72,000 messages a month word-by-word until review fatigue set in — then started missing errors. The worst of both worlds. The goal isn’t maximum scrutiny on everything. It’s the right scrutiny on each thing.
Chapter 15: Building Review Into Workflow
When you need this: Review keeps getting skipped when people are busy. You need it to be structural — impossible to bypass — not optional.
MAKE REVIEW ARCHITECTURAL, NOT OPTIONAL
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
THREE TYPES OF REVIEW POINTS:
1. Checkpoint — at defined stages within a workflow
2. Pre-Action — immediately before output is actioned
3. Post-Action Audit — after, for learning
THREE DESIGN REQUIREMENTS:
Review must be:
✓ Required — cannot proceed without it
✓ Assigned — to a specific person (+ backup)
✓ Timed — with deadlines and escalation
TEST: Run your workflow 5 times.
Did review happen every single time?
If not, the review is an afterthought, not structure.The trap: Three failure modes of afterthought review — the Discipline Problem (first thing cut under pressure), the Visibility Problem (skipped review is invisible until errors emerge), the Ownership Problem (“someone should review” becomes no one’s job).
Chapter 16: Pattern Recognition
When you need this: You’re reviewing AI output and want to catch errors faster. Experienced reviewers don’t read harder — they know where to look.
5 AI ERROR PATTERNS (and their signatures)
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
1. HALLUCINATION — fabricated but plausible info
Signature: Specific names, dates, citations
that feel too perfect. Verify ALL specifics.
2. OUTDATED INFORMATION — stale training data
Signature: "Current" claims without dates.
Check anything time-sensitive.
3. CONFIDENT UNCERTAINTY — debatable claims
stated as settled fact
Signature: Missing hedging language.
Ask: "Would an expert say this without caveats?"
4. CALCULATION ERROR — math/percentage mistakes
Signature: Numbers. Always verify the math.
5. CONTEXT MISREAD — generic response ignoring
your specific situation
Signature: Output that could apply to anyone.
Check: did it actually use YOUR context?The trap: Reviewing through attention alone (reading every line) vs. reviewing through prediction (knowing where errors cluster). Two reviewers look at the same AI market analysis — one spends 20 minutes and catches 3 errors; the other spends 5 minutes and catches 5. The difference is pattern recognition.
Part 6: Scaling Beyond One
Chapter 17: Workflow to Infrastructure
When you need this: Your AI workflow works great for you. Now you want others to use it — and it’s not transferring.
IS YOUR WORKFLOW READY TO SCALE?
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Four readiness signals — all must pass:
✓ Stable — 4+ weeks without major changes?
✓ Documented — explainable in one page?
✓ Proven — measurable, concrete results?
✓ Pattern-understood — can you predict its
failure modes?
If all four pass, choose a scaling path:
Path A: SHARE & SUPPORT
You own everything. Best for small teams.
Path B: DELEGATE & OVERSEE
Someone else runs it daily. You handle escalations.
Path C: FORMALIZE & TRANSFER
Full handoff: complete documentation, a new
owner, and you step away. Docs = SCOPE brief
written for someone who isn't you + stated
review level + starting trust rung (new users
restart at Rung 1).The trap: Premature scaling. Sharing a workflow before it’s stable poisons not just this rollout but future AI initiatives. People remember failure and become skeptical of the next attempt. Also: the Expert’s Curse — you unconsciously skip steps that feel obvious. Watch a new person try the workflow without helping.
Chapter 18: Team Rollout
When you need this: You’re introducing an AI workflow to your team — and with AI you’re rolling out judgment, not just a tool. 80% adoption with good usage beats 95% adoption with poor usage.
ROLLOUT IS A PROCESS, NOT AN EVENT
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
PHASED WAVE APPROACH:
Wave 1: Pilot (2-5 people, 2-3 weeks)
→ Willing but not champions. Real work. Available
for feedback. Visible to their peers.
Wave 2: Connected early majority
→ Peers who witnessed pilot success.
Wave 3: Remaining team
Wave 4: Adjacent teams
Each wave stabilizes before the next begins.
THE 30-DAY ADOPTION DROP:
Weeks 1-2: ▓▓▓▓▓▓▓▓▓▓ Peak (novelty)
Weeks 3-4: ▓▓▓▓▓░░░░░ Drop (old habits)
Week 5+: ▓▓▓▓▓▓▓░░░ Stabilize (if supported)
▓▓░░░░░░░░ Decline (if abandoned)
Prevent the drop: check-ins, celebrate wins,
keep improving, integrate into standard work.The trap: Choosing the wrong pilot users. Enthusiasts succeed regardless — their workarounds hide your documentation gaps. The ideal pilot is a willing, available, average user with real work, visible to their peers.
Chapter 19: Organizational Considerations
When you need this: AI adoption is succeeding across multiple teams — and now nobody can answer how many workflows exist, what data they access, or who’s responsible.
LIGHT-TOUCH GOVERNANCE — 3 PILLARS
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
1. WORKFLOW REGISTRY
Catalog: owner, purpose, data accessed, users
2. QUALITY STANDARDS
Minimum documentation and review requirements
3. ESCALATION PATH
Clear routing for security, compliance,
cross-functional, and quality concerns
CHOOSE A CAPABILITY MODEL:
Community of Practice → Informal network, voluntary.
Start here — zero headcount, surfaces existing
knowledge.
Champion Network → Identified experts per team.
Best when you need local contact points.
Embedded Specialists → AI-skilled people in functions.
Best for mature orgs with different needs.The trap: The centralization pendulum — swinging between full central control (kills innovation) and zero coordination (creates security gaps and duplication). Engage security/compliance early, not after you’ve built. “Yes, if” beats “No, because you didn’t ask first.”
Part 7: Creating With AI
Chapter 20: Vibe Coding Introduction
When you need this: You’re not a programmer, but you have ideas for tools that would solve real problems in your work. You’re wondering if AI can help you build them.
THE VIBE CODING SWEET SPOT
━━━━━━━━━━━━━━━━━━━━━━━━━━
High domain expertise ✓ You know the problem
+ Lower technical complexity ✓ Internal tool, not enterprise
+ Clear success criteria ✓ You'll know when it works
= Good vibe coding candidate
FIT TEST — the FFCC you already own (Ch 5):
Frequent / Forgiving / Clear / Contained
THE CYCLE — the intern model, applied to code:
Describe (a SCOPE brief where the output is code)
→ AI Generates (small pieces, one at a time)
→ Test & Verify (code that runs is NOT code
that works — check against a hand count)
→ Refine or Ship, then iterate
HARD RULES — OFF THE MAP:
✗ Customer data, payments, or authentication
(not without professional security review)
✗ Mission-critical — failure costs more than
the tool saves
✗ Anything you'd sell or deploy to many users
(production software needs professional
engineering, full stop)The trap: Scope creep — gradually adding “one more thing” until the project is unmanageable. And assuming AI-generated code works because it runs without errors. It can produce plausible-looking but wrong results. Test everything.
Chapter 21: First Projects
When you need this: You want to try vibe coding but aren’t sure what to build first. Project selection is the biggest determinant of success.
PICK THE RIGHT FIRST PROJECT
━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Four criteria — all must be true:
1. Solves a real problem (not hypothetical)
2. Has clear success criteria (you'll know when it works)
3. Stays bounded: hours, not weeks (describable
in a paragraph — no requirements doc needed)
4. Uses data you understand (not unfamiliar domains)
THE CONVERSATION PATTERN:
Set Context → "I'm not a programmer. I need a tool
that does X."
Request ONE thing → not everything at once
Describe outcomes → not implementation
Test immediately, report specifically → symptom,
input, expected vs. actual
Expect 3-5 full cycles of the loop — not one.
BUILD INCREMENTALLY:
V1: Core function only → V2: Most important enhancement
→ V3: Edge cases → V4: Polish → STOP.The trap: Feature creep at the “dangerous moment” — when you have something working. “Just one more thing” destroys more projects than technical difficulty. Set the boundary before you start — “complete when it does X” — and when it does X, stop.
- Analysts: Data transformation, format conversion, duplicate identification.
- Managers: Report generation, status aggregation, metric calculation.
- Executives: Start simpler than you think — a useful daily summary beats an ambitious dashboard that never works.
Chapter 22: Build vs Buy
When you need this: You need a tool for your workflow and aren’t sure whether to build it with AI or buy existing software.
THE BUILD VS BUY DECISION
━━━━━━━━━━━━━━━━━━━━━━━━━
Three factors:
1. REQUIREMENT SPECIFICITY
Generic needs → Buy.
Would ignore ¾ of the features → Consider building.
Would use half of it → Buy, even if imperfect.
2. SCALE & CRITICALITY
High scale / mission-critical → Buy.
Personal / small team → Build is viable.
3. MAINTENANCE BURDEN
Stable requirements → Build can work.
Evolving requirements / external APIs → Buy.
Building? Budget ~¼ of the original build time
every year to keep the tool alive.
Low Scale High Scale
Specific needs │ ✅ BUILD IT │ 🔶 BUILD │
│ YOURSELF │ PROFESSIONALLY │
Generic needs │ ✅ BUY │ ✅ BUY │
HYBRID OPTIONS:
• Buy platform + build customization
• Build prototype + buy production version
• Build the glue between commercial toolsThe trap: The default should be “buy.” Commercial software represents years of professional development. But costs get underestimated in both directions — every build runs at least twice the estimate, and every subscription costs meaningfully more than the sticker once training, setup, and workaround time are counted. And walk the inventory yearly: every tool, built or bought, re-earns its place.
Chapter 23: Interactive Presentations
When you need this: You have a presentation where the audience always asks “what if?” questions. Interactive elements let them explore rather than passively absorb.
FIVE TYPES OF INTERACTIVE ELEMENTS
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
1. Calculators — audience inputs their own numbers
2. Scenario Explorers — decision trees showing consequences
3. Data Explorers — filter/sort/zoom dashboards
4. Simulations — time-based models
5. Guided Narratives — branching stories
SHOULD YOU BUILD ONE? Three questions:
Will the audience ask "what if" questions?
Small, engaged room (4-12) + a real decision?
Will it be reused — or are the stakes high?
→ Three yeses → go interactive.
→ Any no → stay static.
START WITH A CALCULATOR — buildable in an
evening; the other four are later builds.The trap: Overbuilding. One well-executed calculator beats five mediocre ones. And always have a backup — never debut an interactive element without static slides ready for when technology fails.
Part 8: The Long Game
Chapter 24: Your Master Prompt
When you need this: Every AI conversation starts from zero. You re-explain your role, your preferences, your context. A master prompt solves that cold-start problem.
BUILD YOUR MASTER PROMPT — 5 COMPONENTS
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
1. Identity — role, domain expertise, communication style
2. Context — current projects, team, org constraints
3. Preferences — quality criteria, formatting, tone
4. Constraints — what NOT to do, what to avoid
5. Priorities — current focus areas, key goals
YOUR V1 WILL BE FOUR SENTENCES, WRONG IN THREE.
Write it anyway. The mature version runs 200-300
words — earned over years of retired corrections.
TWO PRACTICES THAT COMPOUND:
Task Library — base prompt + a permanent SCOPE
fragment per task type (writing, analysis,
communication)
Evolution Protocol — any instruction you give
twice becomes a line item; glance at Priorities
monthly, full pass quarterly, edit immediately
when a correction repeatsThe trap: Don’t include preferences that restrict critical thinking (“always agree with my conclusions”) — that turns calibration into an echo chamber. Don’t include actual confidential data — state constraint categories instead. And don’t try to build the year-two version in your first week — let it earn its length.
Chapter 25: Your Personal AI System
When you need this: You want AI to be infrastructure you architect, not a tool you pick up and put down randomly. This chapter integrates everything from the book into one system.
FIVE INTEGRATED COMPONENTS
━━━━━━━━━━━━━━━━━━━━━━━━━━
1. Master Prompt — your AI configuration (Ch 24)
2. Workflow Library — your proven workflows (Part 3)
3. Task Library — prompt variants (Ch 24)
4. Reference Documents — context docs AI can access
5. Knowledge Base — patterns, lessons, examples
Start with a folder: /ai-system
Populate with what you already have.
Add one useful item per week.
MAINTENANCE RHYTHM:
Immediately — a repeated correction retires into
the right document that day
Weekly (10 min) — save what worked, delete
what proved wrong
Quarterly (one honest hour) — what did I actually
use? Are references current? Delete the clutter.The trap: System bloat. More components is not better. If you find yourself working around the system (avoiding templates, re-explaining context despite your master prompt), the system isn’t serving you — fix or prune it. A lean system you actually use beats an elaborate one you avoid.
Chapter 26: Developing Your AI Collaboration Skills
When you need this: You’ve been using AI for a while but aren’t getting consistently better at it. Random practice doesn’t build expertise — deliberate practice does.
THE FIVE SKILLS OF MANAGING AN INTERN
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Management skills — they persist across every tool:
1. Briefing — what the intern needs, before the
output proves something was missing
2. Reviewing — fast evaluation + precise direction
(incl. the iterate-or-restart call)
3. Delegating — breaking work into intern-sized pieces
4. Situation Recognition — "I've handled this before"
→ reach for the proven approach
5. Knowing What Not to Delegate — some work was
never AI-shaped
THE DAILY REP:
One interaction per day, run deliberately —
notice before, execute with attention,
30 seconds after: what worked, what didn't.
CADENCE (rides the calendar you already keep):
Daily — one deliberate rep
Weekly — the folder's 10 min (Ch 25) gains one
question: which skill was weakest this week?
Quarterly — the honest hour gains a skills half:
rate yourself on all five, honestlyThe trap: Learning features instead of skills. Features change every update. Skills transfer across every tool. Also: comfort zone persistence — only using AI for tasks you’re already good at doesn’t build new capability.
Chapter 27: Career Positioning in the AI Era
When you need this: You want your AI skills to be a genuine career differentiator — not just another line on a resume everyone claims.
TWO CANDIDATES CLAIM "AI EXPERIENCE":
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Candidate A: "I use AI every day."
Candidate B: "I built a workflow that reduced
our reporting from 2 days to 4 hours."
Only one gets the job.
FOUR DIFFERENTIATING CAPABILITIES:
1. Judgment — knowing when NOT to use AI
2. Reviewing — evaluating AI output fast and accurately
3. Domain expertise enhanced by AI — your field + AI
4. Systems thinking — building infrastructure,
not just executing tasks
YOUR FOLDER (Ch 25) IS YOUR PORTFOLIO — assemble,
don't create:
• Workflow docs → before/after metrics, already written
• Error log → proof of calibration
• Knowledge base → judgment examples, incl. what
you chose NOT to automate and why
Curation is translation: business outcomes, not AI
features. Don't tell them — show the workflow.The trap: Being labeled “the AI person” instead of “domain expert who uses AI effectively.” Feature-specific knowledge decays. Invest in judgment, reviewing, and systems thinking — those persist.
Chapter 28: Continuous Improvement
When you need this: The prompt that worked in January may be obsolete by July. You need a sustainable practice for staying current without chasing every new feature.
THE LOOP, POINTED OUTWARD
━━━━━━━━━━━━━━━━━━━━━━━━━
Monitor → Test → Update → Share
(the intern model's feedback loop, aimed at what
the world changed instead of what the intern
got wrong)
NO NEW CALENDAR — the one you keep, pointed outward:
Daily — the deliberate rep (Ch 26) tests whatever
the scan surfaces
Weekly — the 10 minutes becomes 15 with one added
question: what changed out there this week?
(2-3 curated sources, no more)
Quarterly — the honest hour gains an external
third: what shifted, and what does it obsolete?
Annually — one reset day: the tool audit (Ch 22),
the portfolio refresh (Ch 27), next year's
learning priorities
CORE PRINCIPLE:
Evolve, don't overhaul. Evidence over enthusiasm.
Test before switching — and keep the fallback.The trap: Five improvement pitfalls — chasing every new feature (selective > comprehensive), abandoning working practices for shiny ones (test first), improvement as procrastination (if improvement time > productive use, the ratio is backwards), ignoring regression (new doesn’t always mean better), and isolation (learning alone is slower than learning together).