From IT to Influence · AI Cheat Sheet
Prompt to Ship · AI for Your Team · v1
▸ Operations Rubric — Research / Triage
| Dimension | What "Good" Looks Like | Failure Signal |
|---|---|---|
| Triage speed | Classified + owner in <15 min | Still reading source material after 20 min |
| Citation quality | Every claim → dated, linked source | "According to the internet…" no URL |
| Tone fit | Exec-readable in 20 sec; warmth w/o fluff | Reads like a stack dump or a press release |
| Scope discipline | Answer only what was asked | Five pages when the ask was a paragraph |
| Sanitization | Zero PII, zero unreleased data | Anything you wouldn't post publicly |
▸ Development Rubric — Specs / Backlog
| Dimension | What "Good" Looks Like | Failure Signal |
|---|---|---|
| Spec fidelity | AC maps 1:1 to requirement | AC restates the title in different words |
| Test coverage | ≥1 happy path + 1 edge case per AC | "Test: it works" |
| Reproducibility | Peer runs same prompt → comparable spec | Output depends on vibes |
| Technical accuracy | APIs / endpoints / fields actually exist | Invented endpoint that sounds plausible |
| Deploy-readiness | Lands in tracker without a rewrite | PM has to re-spec before groom |
▸ Three Failure Modes — Debug the Prompt
- Missing contextOutput is generic — says nothing specific to your system or project. Fix: attach the source doc, paste the prior spec, link the ticket. The model can't see what you see.
- Vague instructionYou got a three-paragraph essay when you wanted five bullets with owners. Fix: rewrite instructions as a checklist. Count the bullets. Name the columns.
- Model mismatchWrong tool — fast model for nuanced reasoning, text model for video, general model for deep research. Fix: switch lanes. Perplexity for sourced research, reasoning model for logic, agent for multi-step.
▸ Prompt PRD Template — Write This First
Persona: [role the AI should play]
Context: [attach artifacts, not descriptions — paste, don't describe]
Constraints: [3 hard rules — tone, sources, length, format]
Success metric: [what "done" looks like, measurable]
Non-goals: [what to NOT produce]
Context: [attach artifacts, not descriptions — paste, don't describe]
Constraints: [3 hard rules — tone, sources, length, format]
Success metric: [what "done" looks like, measurable]
Non-goals: [what to NOT produce]
Guardrails
- No PII (names, emails, customer IDs, employee data)
- No unreleased pricing, roadmap, or vendor terms
- No data outside your role's authorized scope
- Sanitize exports before pasting into any AI tool
- Use only tools approved by your IT / security policy
▸ The Five-Stage Prompt Pipeline
01 · Export
Saved filter → CSV · strip PII
▶
02 · Classify
Normalize + bucket by severity + owner
▶
03 · Render
Generate HTML card (navy/gold, ≤160 words)
▶
04 · Ship
Paste into exec email or dashboard tile
▶
05 · Checkpoint
Save CSV + HTML · rerun next cycle
▸ Seven Unblocking Moves — When You're Stuck
01
Prompt the prompterAsk AI what prompt it needs from you.02
Switch lanesDownload artifacts; hand to a different tool or thread.03
Many formatsOne session → HTML + slides + explainer doc.04
Constraint as fuelOut of credits = forcing function, not failure.05
Checkpoint itSave CSV/HTML exports like git commits.06
Build the bridgeNo integration? Export → sanitize → reshape → ship.07
Council reviewAsk a 2nd AI: "what's wrong, missing, or overconfident?"