Session 1: Automate the Signal
Everything from Session 1 of Signal to Strategy. Slides, the working guide, every prompt we ran, the Launchpad data pack, and the actual conversation we walked through.
Enter your cohort code.
Session resources are for Signal to Strategy cohort members. Your code is in the welcome email and pinned in the cohort Slack channel.
Not in the cohort? See the course →
Everything from Session 1.
Session 1 deck
The full 90 minutes, including the Launchpad findings.
Open the slides →Can vs. Should Audit
Twelve jobs, scored on two axes, plotted in a 2x2. Scores stay in your browser.
Open the tool →Working guide
The follow-along guide with every prompt, the framework, and the alternate lenses.
Download the PDF →Launchpad data pack
The exact workspace from the session: 10 interview transcripts, 75 tickets, 120 NPS responses, the company overview, and the prompt pack.
Download the .zip →Launchpad company overview
Who Launchpad is, the plan tiers, the competitive set and the churn picture. Read this before the build.
Download the PDF →The run itself
The actual conversation from the session, prompts and replies.
Open the chat →AI Working Group on Slack
200+ product people. No pitching, no vendors.
Join the Slack →This is the whole build.
The prompt pack we ran live, unchanged. Each one is a move, not a magic string: you're copying the reasoning, not the words. The bold callouts are the judgment moves.
These prompts are written against the Launchpad workspace in the data pack above. Download it and they run as-is. The same six-step pack ships inside the zip as PROMPTS.md.
Set your context
The last line is the one doing the work. It stops the model performing analysis before it has read anything. If the reply could apply to any SaaS company, your context isn't specific enough.
Load the data
Watch it run commands to open the CSVs rather than working from the column names. Then compress before you go on, "summarize that in 10 bullets", because a wall of analysis is not a finding. On your own data later, connectors do the loading: hook up Gong, Zoom, Granola or Drive and transcripts come in directly, no manual exports.
Ask for surprises
The highest-value moment in the whole workflow. Protect it. If the first answer is thin, push back out loud: "That's a pattern, not a surprise. Tell me something I wouldn't have guessed from reading two transcripts." And ask explicitly who isn't represented, churned-and-gone customers, the quietly dissatisfied, smaller accounts, before you trust the picture.
Build your framework
Opportunity Solution Tree is the default. If the OST skill is loaded, run that instead of the prompt and notice the difference: user-need framing, experiments rather than solutions, evidence grounding, all automatic. That gap between a thin prompt and a real skill is the whole point.
The skill: opportunity-solution-tree, built on Teresa Torres' Continuous Discovery Habits. Get it from mcpmarket, or download the SKILL.md and add it to your own Claude. Run the prompt below first, then the skill, and compare.
Jobs-to-Be-Done is the alternate if you'd rather understand why customers hire or fire the product. The designer track, How Might We statements plus a journey map, is a third option. Both are in the working guide, along with a problem-prioritization matrix, a hypothesis brief and a segmentation map.
Pressure-test and own it
Don't lose the human. Before you trust a theme, open two or three of the raw verbatims behind it. Averaging 75 tickets into four bullets is useful, and it quietly erases the specific person who was frustrated. And own it: when this goes to the VP, you own the recommendation, not Claude.
Make it yours
Open a new Cowork space for your own product. Adapt the step 1 context prompt: who you serve, what decision you're making. No customer PII needed. Point it at one folder of your real discovery data, or connect Gong, Zoom or Drive, and run step 3 on it.
What discovery data do you already have sitting unsynthesized, a Granola folder, a Zendesk export, last quarter's NPS? Pick one and run this on it.
Did it actually use the skill?
In the run we walked through, the answer was: "No, I didn't. Honest answer." It had the skill. The skill was relevant. It didn't read it, and it only said so because it was asked. Follow with "just the delta, what changes vs. the current analysis" and you see the size of the gap.
Real signal, or a loud anecdote?
Run every finding through these four before it reaches a stakeholder.
- Frequency. How many times, out of how many? "Six mentions" means nothing. Six out of forty is a pattern; six out of six hundred is noise.
- Spread. How many different sources, and how many different segments? One angry account filing nine tickets is one data point that learned to type.
- Consequence. What's attached? Revenue, churn, deal cycle, support cost. A theme with no consequence attached is a preference.
- Corroboration. Does anything independent agree? Behaviour beating words is the strongest form: they said it's fine and the usage says otherwise.
Worked example from the Launchpad data: self-serve fails frequency in NPS badly, one comment out of 120, and passes everything else, appearing in all ten interviews and fifteen tickets, and named by the 17 accounts that left for a competitor. Mobile passes frequency with ten NPS mentions and thins out fast: six accounts churned on it. Count mentions and you ship a mobile fix. Weigh consequence and you build the flow editor.
These are a floor, not a ceiling.
One sentence from your largest account can outrank fifteen small-tier tickets, and no amount of counting will tell you that. AI counts frequency. You decide what matters, which is exactly the judgment that lands in the trap quadrant.
Score each job twice.
Can AI do this? 1 is confident garbage, 3 is a solid first draft you do real work on top of, 5 is at or above your level. Should it? 1 is never delegate the decision, 3 is delegate the draft and own the call, 5 is delegate freely. Three is the pivot, and the worksheet counts only 4 and 5 as high.
Automate
Hand it over Monday.
Trap
It'll do it, and you'll lose the thing you're accountable for.
Queue
You would, the tools aren't there yet.
Leave it
Not now, revisit.
Should: the hard half
Judgment depth. How much of this is a call rather than a task?
Reversibility. What does a wrong answer cost, and can you undo it?
Accountability. Whose name is on this in a room full of executives?
Presence. Does the value come from a human actually being there?
Can: the easy half
Input verifiability. Can you check whether the answer is right?
Context. Does it need things that live only in your head or your org?
Volume. Is there enough of it that a machine actually helps?
Put in hours per month.
That field is what turns the audit from an interesting chart into a decision: the job sitting in automate with the most hours behind it is the one to hand over first. It's usually the least interesting thing you do, which is exactly why it's the biggest win.
Before next Wednesday.
- Finish all twelve jobs in the audit and post your trap quadrant in the cohort channel. Week 2 assumes you know which of your decisions you're not allowed to delegate.
- Get your signal intake running on one real source and post one surprise it found. One source is enough: a ticket export, a folder of call notes, last quarter's NPS.
- Bring a live decision to Session 2. One you're carrying right now and aren't sure about. We're going to try to break it. The ones you're certain on make for a boring session.
Week 2: Stress-test the bet.
Validation prompts for a real decision, tuned to the people who actually push back on you. Your CEO reads risk. Your CFO reads payback. Your eng lead reads scope. Different lens, different prompt. Session 2 resources →
Momentum Product Co. · Heather Gawel + Babajide Okusanya · Launchpad is fictional. For course use only.