Appendix A: Framework Summary
This appendix provides quick-reference cards for every framework in the book. Use it when you’re mid-workflow and need a refresher without flipping back through chapters. For step-by-step decision flowcharts that connect these frameworks into a single action path, see Appendix D.
The Intern Model
Introduced: Chapter 2 | Core idea: AI is a capable assistant, not an autonomous decision-maker
Four principles:
Clear Tasks—Specific, bounded assignments with defined inputs and expected outputs. Not “make this better.” Instead: “Summarize these three customer complaints into a single paragraph highlighting the common theme.”
Review Before Shipping—Nothing AI produces reaches customers, stakeholders, or public channels without human review. No exceptions.
Incremental Trust—Start with low-stakes tasks. Expand AI’s role only after it demonstrates consistent quality in simpler work.
Feedback Loops—Correct mistakes explicitly, track the patterns, and feed what you learn back into the system. You still own the output—the human remains responsible for decisions and consequences.
Quick test: Would you let a first-week intern do this unsupervised? If not, don’t let AI do it unsupervised either.
SCOPE
Introduced: Chapter 2 | Core idea: Five elements turn a vague request into a well-defined AI task
- S—Specific outcome: What exactly should exist when the task is done?
- C—Constraints: What must the AI not do, assume, or invent? This section prevents the most common failures—give it the most attention.
- O—Output format: Structure, length, and form of the deliverable.
- P—Prior context: The background the AI needs—documents, history, examples of “good.”
- E—Evaluation: How you’ll judge whether the output is acceptable.
Feedback rule: When review catches a recurring problem, it becomes a new constraint in your SCOPE definition—that’s Principle 4 in action.
The Trust Ladder
Introduced: Chapter 2 | Core idea: Autonomy is earned rung by rung, per task type—never granted globally
| Rung | Name | What It Looks Like | Promotion Bar |
|---|---|---|---|
| 1 | Assisted Drafting | AI produces raw material; human reworks everything | Every new task type starts here |
| 2 | Supervised Output | AI drafts need moderate editing (~30–50% changes) | Most teams land here |
| 3 | Spot-Checked Production | Check 1 in 5 outputs; well-understood, lower-stakes tasks only | <5% errors sustained for 2+ weeks |
| 4 | Exception-Based Review | AI handles routine cases; humans review flagged exceptions | Extensive track record |
Calibration rule: Trust is per task type. An AI at Rung 3 on meeting summaries is still Rung 1 on anything it hasn’t done before.
The FFCC Scorecard
Introduced: Chapter 5 | Core idea: Score potential AI workflows to find the best starting opportunities
Rate each dimension 1–5, then multiply:
| Dimension | Question | High Score (5) | Low Score (1) |
|---|---|---|---|
| Frequent | How often does this happen? | Multiple times daily | Once a quarter |
| Forgiving | What if AI gets it wrong? | Internal draft needs revision | Wrong medical advice sent |
| Clear | Are inputs/outputs defined? | “Summarize in 3 bullets” | “Make this better” |
| Contained | How many systems touched? | Single tool, one step | Triggers 5 downstream processes |
Score range: 1–625. Start with your highest-scoring workflows. Anything under 100 should wait until you’ve built confidence with easier wins.
Workflow Structure
Introduced: Chapter 7 | Core idea: Every AI workflow follows the same five-step pattern
Trigger → Input → AI Processing → Human Review → Action- Trigger: What kicks off the workflow? (New email, daily schedule, customer message)
- Input: What does the AI receive? (The email text, calendar data, conversation history)
- AI Processing: What does the AI do with it? (Summarize, classify, draft a response)
- Human Review: What does the human check? (Accuracy, tone, appropriateness)
- Action: What happens after approval? (Send response, file report, escalate to manager)
Design rule: If you can’t name the human review step, you don’t have a workflow—you have a risk.
The Permission Framework
Introduced: Chapter 11 | Core idea: Three questions determine whether an AI application should be permitted
Before expanding AI into any new area, answer:
What’s the specific benefit? Not “efficiency” or “productivity.” What measurable outcome does this deliver? If you can’t name it, you can’t justify the risk.
What’s the actual exposure? What data does the AI see? Who could be affected by errors? What’s the blast radius of a mistake?
What are the failure modes? Not “what if it doesn’t work” but “what specifically goes wrong, and what’s the recovery plan?”
Decision matrix:
| Benefit | Exposure | Failure Mode | Decision |
|---|---|---|---|
| Clear, measurable | Low, internal | Recoverable | Green light |
| Clear, measurable | High, external | Recoverable | Proceed with enhanced review |
| Vague | Any | Any | Stop. Define the benefit first |
| Any | Any | Unrecoverable | Red flag. Require executive sign-off |
The Review Spectrum
Introduced: Chapter 14 | Core idea: Not everything needs the same level of review
Four levels, calibrated to stakes and AI track record:
| Level | When to Use | What You Do | Time |
|---|---|---|---|
| Scan | Low stakes, proven workflow, internal use | Glance for obvious errors | 10–30 seconds |
| Spot-Check | Medium stakes, established workflow | Check 2–3 key claims or numbers | 2–5 minutes |
| Deep Review | High stakes, new workflow, external-facing | Verify every claim, check sources, test edge cases | 15–30 minutes |
| Rewrite | Critical stakes, or AI output misses the mark | Use AI output as raw material, rewrite in your voice | 30+ minutes |
Calibration rule: New workflows start at Deep Review. As AI demonstrates consistent quality, gradually move toward Spot-Check. Scan is earned, never assumed.
Master Prompt Methodology
Introduced: Chapter 24 | Core idea: Build a persistent AI instruction set that compounds your effectiveness
A master prompt has five components:
- Identity—Who you are professionally, what you’re responsible for, what decisions you own. A working description, not a résumé.
- Context—Your industry, organization, and environment. The stable facts that shape your work.
- Preferences—How you want information delivered: format, length, tone, structure. Each line should retire a correction you’ve made repeatedly.
- Constraints—What can’t be done, shared, or assumed. Approval limits, confidentiality boundaries, compliance requirements.
- Priorities—What matters right now. The living section—it changes monthly while the other four barely move.
And two practices that keep it alive:
- The Task Library—Standing variants for recurring work: the base master prompt plus a permanent SCOPE fragment per task type (writing, analysis, communication).
- The Evolution Protocol—The feedback loop pointed at your own instructions. Every instruction you give twice becomes a line item; the master prompt is where line items retire. Review priorities monthly, the whole document quarterly.
Key principle: Your master prompt should make every AI interaction start from your standards, not from zero. A well-maintained master prompt is the difference between “AI assistant” and “AI infrastructure.”
The Five Skills of Managing an Intern
Introduced: Chapter 26 | Core idea: The durable skills are the manager’s, not the tool’s—they transfer across every model and job change
- Briefing—Defining tasks so they come back right the first time. SCOPE is the tool; the skill is knowing what to put in it.
- Reviewing—Catching what’s wrong quickly and knowing what “good” looks like. The Review Spectrum is the tool; calibration is the skill.
- Delegating—Choosing the right rung on the Trust Ladder for each task type, and moving work up as evidence accumulates.
- Situation Recognition—Reading a new task and knowing which pattern it fits: what to check, how much to trust, where it will probably fail.
- Knowing What Not to Delegate—Maintaining the never-list, and revisiting it as capabilities shift. The boundary is a decision, not a default.
Practice rule: These improve with deliberate reps, not accumulated hours—one daily interaction done thoughtfully beats twenty on autopilot.