AI in GTM School · Q2 2026
Class 3 of 8 · Recap & Takeaways

AI Tools in Action

The three pillars of advanced prompting, four real GTM workflows built live, and the antipatterns that trip everyone up. John Williams's working hour with the cohort.

Instructor: John Williams (Revenue Architect · AI Growth Executive)
Date: May 13, 2026
Focus: Awareness → Essential

TL;DR

Out of "dozens, if not scores" of prompt-engineering methodologies, John distilled the cohort's working set down to three pillars: Role · Examples · Reasoning. The pillars are model-agnostic — they work in Claude, ChatGPT, Gemini, Grok, Copilot. The hour ran four GTM use cases through the pillars, then closed with the antipatterns that quietly degrade everyone's AI output, plus a context-window discipline most of the room hadn't seen before.

The Framework

The Three Pillars of Advanced Prompting

Three things to remember. The model is the same — your prompt is what changes. John kept these in the same order in every demo so the room could feel the pattern.

1

Role (Persona Engineering)

Tell the model who it is. Not "you're a marketer" — say "you're a SaaS VP of Marketing who's run six product launches and lost two." A role with experience and a point of view changes the output more than any other lever.

"There are experts of tasks. Tell the model which one to be."
2

Examples (Few-Shot)

One good example beats 500 words of description. Show the model what "good" looks like — paste in the past output you'd want to match. Skip this and you'll be in the second pillar of "There Be Dragons."

"Skipping examples will hurt. It will, it will, it will cause pain."
3

Reasoning (Show Your Work)

Ask the model to think out loud before it commits to a final answer. "Explain your logic, then give me the output." You'll catch wrong-path answers before they turn into deliverables you have to walk back.

"The three pillars are agnostic. So it doesn't matter which model you're in — one size does fit all in this case. Just pick one workflow and start there."— John Williams

Hands-on Keyboard

Four GTM workflows, four different models

John ran the room through four real use cases — same three pillars each time, deliberately routed through different platforms to surface where each one wins.

Use case 1 · Sales Outreach

New business · 5-touch cadence

  • Scenario: 25 target accounts to build before end of Q2
  • Research: Perplexity (deep-research mode pulling current activity)
  • Drafting: Gemini 3 — strong for Google Workspace shops
  • Three pillars: Role = "experienced SDR with strong outbound POV"; Examples = past winning sequences; Reasoning = explain target-fit before writing
Why two tools: Perplexity for the freshest grounding data; Gemini for clean structured output if your security stack is Workspace-aligned.
Use case 2 · Expansion / QBR

Cross-sell into an existing customer

  • Scenario: healthcare account, usage drop, QBR next week, cross-sell opportunity
  • Tool: Claude (Sonnet) — could also run in Copilot
  • Three pillars: Role = "CSM running a turnaround QBR"; Examples = strong past QBR decks; Reasoning = diagnose the usage drop before pitching the cross-sell
  • Output: QBR checklist, opening-statement script, cross-sell positioning
Note from John: Don't let it skip diagnosis. The model wants to pitch first; the pillars force it to think before it sells.
Use case 3 · Content Brief

Marketing brief + creative direction

  • Tool: Gemini 3 — content briefs, structure, asset-friendly outputs
  • Surfaced both the creative direction AND the metrics to track
  • Showed the model also doing RevOps coding and ideation, not just briefs
What it changed: Most rooms think of Gemini as "the Google one." It's a frontier model with the same range as Claude — just packaged differently. Use it where the data already lives.
Use case 4 · Forecasting

End-of-quarter forecasting narrative

  • Tool: Grok — strong on real-time data, X/Twitter signal, and stepped-thinking output
  • Demo showed the model's chain-of-thought feeding back in real time
  • The pivot John reinforced: ask the model to surface the data it needs, not just feed it
Sticky line: "Once you master the pillars, the question shifts from 'how do I prompt this' to 'how do I automate this for end-of-month.'"

Tool selection

The Platform Matrix

"It's a version, not the version" — John's words. Start here when you don't know which model to pick; revise based on what works in your actual workflow.

Tool Strongest at Where it struggles When to reach for it
Claude (Sonnet / Opus) Long-form reasoning, structured analysis, code, content with voice Real-time data unless connected; image generation Default for GTM operators who want one tool that does most things well
Gemini Workspace-native work, long-context document analysis, NotebookLM-style synthesis Slower iteration UX than Claude/ChatGPT When data already lives in Google Drive or Docs; security-conscious orgs
Perplexity Deep research, current information, multi-source synthesis with citations Generative drafting beyond research The first step of any unfamiliar account, persona, or competitor
Grok Real-time signal from X, stepped-thinking visibility, breaking-news context Smaller ecosystem; fewer integrations When recency matters more than depth
Copilot (M365) Native Microsoft 365 integration, has Claude underneath, security envelope Constrained UX vs. running Claude directly Microsoft-shop enterprises that can't use external models

The Boring-but-Critical Layer

Enterprise data privacy

A theme John kept returning to: most of you are either a Google Workspace shop or a Microsoft 365 shop. Both come with a frontier model already protected under your enterprise's data-privacy agreement.

Workspace shop → Gemini is your default. Your data stays within the security envelope. Notebook LM, agent-building tooling, and free enablement content all live here.
Microsoft 365 shop → Copilot is your default. Copilot has Claude and ChatGPT under the hood. Visual Studio Code + AI Foundry sit alongside. Same envelope, same protection.
Don't bypass the envelope. Even when you strip PII, sensitive context leaking outside the enterprise model is a risk. Use what's already covered before bringing in a new vendor.
Both come with free training. Microsoft and Google publish extensive enablement material others have to pay for. Block time on your enablement calendar for "build an agent in our environment."

Antipatterns

"There Be Dragons" — what trips everyone up

Four common failure modes. Each is fixable in under a minute once you can name it.

1

Accepting the initial output

It may look polished. That doesn't mean it's right. Pressure-test the first answer before you ship it — ask the model to challenge itself, or ask for the version it would have written if it knew you'd push back.

2

Skipping examples

The single biggest quality lever is also the most-skipped step. One past output of "what good looks like" beats any amount of adjective stacking.

3

Reading instead of using

You can't read your way to fluency. The model is the most patient teacher with all the answers — but only if you actually engage. Block 30 minutes after class. Run one of the workflows. Don't just bookmark the deck.

"You can't just come to GTM and at school and be an expert. You have to use it."
4

Letting the context window expire

After 15–20 turns, the model starts to drift. Hallucinations creep in. Use the wrap-up trick: ask for a summary plus an introductory prompt to continue. Paste both into a new window. Keep rocking.

"At the end of each conversation I say: 'Create a summary of what we've discussed and an introductory prompt to continue.' I copy that. Start a new context window. Paste it in. I'm good to go."— John Williams, on context-window discipline

For the deep-end folks

Local and open-source models

A small but growing slice of the cohort wants to run AI without sending data anywhere. John spent the last few minutes here.

Why local

When the laptop is the boundary

  • No internet required — work offline, on a plane, in a SCIF
  • No vendor lock-in, no usage limits, no waiting on token resets
  • Specialized models for photo analysis, video, regulated workflows
  • The hardware bar is real — fast enough to be useful = recent Mac or workstation GPU
John's take: Treat local-model fluency like a second language — it adds value to your career, but it's not where you start.
Why vertical

Specialized = the next wave

  • Frontier models do "anything" reasonably well
  • Specialized open-source models do "one thing" exceptionally well
  • Vertical use cases (inspection video, medical imaging, regulated finance) increasingly need specialists, not generalists
Strategic read: Just like software went vertical, AI is going vertical. Knowing how to deploy a specialist model in your industry is a future-proof skill.

Sticky Quotes

Lines worth saving

"The model is one of the most patient teachers with all the answers you can have. Just engage with it. Ask it: what does this mean? What should I know about this?"— John Williams
"There's a problem with junior people right now. They churn out content and skip the last mile of reviewing it. They don't even know what's in it. Can they ask a follow-on question? They might not even know what's on page three."— John Williams, on AI hygiene
"Practice makes perfect. You're not going to break anything. But you have to use this to get good at it."— John Williams
"Don't ignore the learning opportunity to develop the skills you're going to need in the out-years ahead."— John Williams

The Cohort Asked

Questions worth the replay

The most-engaged threads

  • "What's the sweet spot for prompt length?" Enough context to do the job, no more. The pillars (Role + Examples + Reasoning) are the structure; everything else is fat.
  • "Can I use this format to design A/B testing for ICP messaging?" Yes — the persona engineering step is exactly the same. Iterate the Role pillar across personas; hold Examples + Reasoning constant.
  • "For Gemini's deep-research mode — should I select that for this prompt?" Yes. Deep research mode invokes additional source-grounding; use it any time you'd open Perplexity.
  • "Can I assign personas to projects in Claude?" Yes — Projects hold persistent context; you can bake a Role-pillar persona into the project's system prompt and reuse it indefinitely.
  • "How do I scale 1:1 outreach into batches?" The pillars stay the same. The change is moving from one-account-at-a-time prompting to a chained workflow — which is exactly what Class 2 set you up for.

Before Class 4

What to do this week

John's homework — pick one quick win for your function, run it through the pillars, post the output back to your cohort group.

Pick one quick win. Sales: account research + outreach for 5 target accounts. Marketing: a content brief for next month's campaign. CS: a QBR deck for your most at-risk customer. RevOps: a forecasting narrative for next quarter.
Run it through all three pillars. Role: who is the model acting as? Examples: paste one strong past version. Reasoning: ask it to show its work before delivering the final.
Wrap up your context window when it gets long. After 15–20 turns: "Create a summary and an introductory prompt to continue." Paste in a new window. Don't fight a degrading session.
Post your output to your cohort. Share the input, the output, and one thing that surprised you. The cohort gets smarter together; don't workshop in private.
← Previous
Class 2 · Growth Strategy Design