The complete workbook — Artifacts 1–8
AI Agents for Startup Strategy: Workbook
Living templates from AI Agents for Startup Strategy. Copy each into your notes system or print it. These are operating instruments, not one-time exercises: every artifact carries a review rhythm. Printable versions: the QR at the back of the book or lenvanderhof.com/ai-agents/worksheets.
Artifact 1 — Information-bottleneck map (Chapter 0)
When: One evening to start; then a one-line entry whenever you are surprised by something you could have known.
| Decision (last quarter) | When you decided | When the info became available | When you could have known | What the delay cost |
|---|---|---|---|---|
| 1. | ||||
| 2. | ||||
| 3. |
Trace drill: For the worst gap, write the path the signal took to reach you (page → thread → person → meeting) and circle where it sat waiting for a human. That waiting room is your first agent’s job description.
Surprise log (standing): date · what surprised you · when it became knowable · when it reached you · cost. A month of entries is your prioritized build list.
Artifact 2 — SENSE design canvas (Chapter 2; extended in 5, 8, 12)
One strategic question: ____________________________________________
| Phase | Your one-line design | Failure mode to avoid |
|---|---|---|
| Scan | what is watched, and why those sources | thesis-free collection |
| Evaluate | the claim shape, against which named exposure | summary, not judgment |
| Navigate | what may change without asking | the static watchlist |
| Simulate | the two futures kept warm | confident fortune-telling |
| Execute | which existing meeting; who owns the output | the digest that arrives nowhere |
Cadence ring it feeds: daily / weekly / monthly / quarterly / annual Kill criterion (write it now): ______________________________________
Artifact 3 — Delegation-boundary worksheet (Chapter 2; risk modifiers Chapter 10)
| Decision | Lane (delegate / augment / reserve) | One-line rationale |
|---|---|---|
| 1. | ||
| 2. | ||
| 3. | ||
| 4. | ||
| 5. | ||
| 6. | ||
| 7. | ||
| 8. | ||
| 9. |
Risk re-check (quarterly): Which lanes did hallucination, bias, over-reliance, or regulatory exposure shift this quarter? Move the rows; date the change.
Artifact 4 — Signal map (Chapter 3; assumption register Chapter 4; customer streams Chapter 6)
Three competitive fears, as full sentences:
| Watched surface | Quadrant (core / selective / ambient / ignore) | Latency budget |
|---|---|---|
Ignore list (at least three, with reasons): _____________________________
Assumption register (market edition):
| # | Assumption (falsifiable sentence) | Tier-1 confirmer | Tier-2 interpreter | Tier-3 early rumble | Timestamp |
|---|---|---|---|---|---|
| 1 | |||||
| 2 | |||||
| 3 | |||||
| 4 |
Customer-signal streams: support · sales notes · churn notes · usage · reviews — mark which currently reach roadmap decisions, and which sit unread.
Artifact 5 — Integration cadence plan (Chapters 7, 9, 13)
System: ____________________ Existing meeting its output enters: ____________________ Who would notice if it stopped: ____________________
| Output | Integration pattern (parallel / sequential / human-gate / autonomous+audit) | Gate placement |
|---|---|---|
Load test: the chore this system removes on the user’s first interaction: ______________________ Investor-graph instance (Ch 7): five most-likely leads, each with thesis (quoted), stage fit, timing signal, warm path. GTM loop instance (Ch 13): standing goal · SENSE walk · baseline metric and current value · guardrails (grounding source, voice reference, volume limit, compliance gates).
Artifact 6 — Five-risk guardrail audit (Chapter 10)
| Risk | How it would manifest here | What stands between failure and a decision today | Guardrail to add (class) |
|---|---|---|---|
| Hallucination | |||
| Bias | |||
| Over-reliance | |||
| Security (prompt injection) | |||
| Regulatory |
Evidence-trace test: pick one real claim the system produced and walk it back to its source. If you cannot, the trace is broken — fix that before anything else. Lane shifts: which delegate-lane decisions did this audit push toward augment or reserve?
Artifact 7 — Agent-fluency scorecard (Chapter 11)
Score 1–5, quarterly. Lowest score is the quarter’s development priority.
| Role | Deployment | Judgment | Correction | Governance |
|---|---|---|---|---|
| Founder | ||||
| System owner(s) | ||||
| Fluency lead | ||||
| Team average |
Calibration meeting (monthly, 30 min): What did our systems get right that we acted on, and what did it teach us about trust? What did they get wrong, and what was the root cause? What changes, and who owns it?
System metrics: decision influence · claim accuracy · correction-rate trend · time-to-trust-recovery.
Artifact 8 — Override doctrine + 90-day plan (Chapters 14 and Conclusion)
Never delegated (with one line each on why it is constitutively yours):
Override conditions (when I back my judgment over a confident system): missing context I hold · stakes beyond any track record · conflict with the company we are building.
Sharpness practice: three decisions this month my systems could make that I will make myself: 1. ____________________ 2. ____________________ 3. ____________________
The 90-day plan
Days 1–30 · Diagnose (build nothing): run Artifact 1 for a full month; write one Artifact 2 canvas and one Artifact 3 sheet for the single system the log justifies. Gate: say it in one sentence — watches X, claims Y, feeds decision Z in meeting W.
Days 31–60 · Deploy: one system, simplest architecture, into the existing meeting, instrumented from day one. Gate: it changed at least one real decision, and you can point to which.
Days 61–90 · Evaluate: score against the kill criterion you wrote in week one; decide expand / adjust / kill, in writing.
The day-30 gate goes in your calendar now: ____ / ____ / ____
The layer is not what you bought. It is what you run.
↑ top