Our best tech has 0 callbacks on 47 jobs. The culture rubric scored him a 1.

"Enthusiastic communicator who lights up the room."

That line sat in a hiring rubric built from meeting notes at a 6-person HVAC shop. Their best tech texts six words a job. He has 0 callbacks on 47 installs. Under that rubric he scores a 1 out of 4.

Stated values are not the same as who decides.

AI Prompts, Reviewed takes one prompt or workflow from How I AI, Claire Vo's show where guests share their screens and walk through how they really use AI. We run it on a small-business job, score it, and tell you what we'd change.

Credit: Based on "How to Use AI to Define Your Unspoken Company Culture with Granola," a workflow from How I AI (hosted by Claire Vo, written up by the ChatPRD team), shown by Wade Foster, Zapier CEO, in the episode "Zapier CEO Wade Foster's Playbook for AI-Powered Recruiting and Culture Building." Original workflow: https://www.chatprd.ai/how-i-ai/workflows/how-to-use-ai-to-define-your-unspoken-company-culture-with-granola. Full episode: https://www.chatprd.ai/how-i-ai/zapier-ceo-playbook-for-ai-recruiting-and-culture. Guest site: none listed on the How I AI workflow page.

What the original does

Wade runs Granola's unspoken-company-culture recipe across authorized meeting transcripts, compares what shows up with stated values, and turns durable behaviors into interview criteria.

Four steps: pick a representative authorized transcript set, run the culture recipe, compare the draft with stated values and feedback, then convert accepted themes into a hiring rubric with observable evidence and anchored scores. The page already warns that one executive's meetings are not the whole company, and that the model can turn stated values into flattering copy.

The test

Made-up data for Bramblewick HVAC (a made-up HVAC company) in Mesa: owner Mira, 4 techs, 1 office person (6 people). Clock: Thursday, Oct 15, 2026. Role open: a fifth tech at $52,000 a year. We Fed 6 authorized snippets from Mon Oct 5 through Fri Oct 9, 2026, plus the website line "customer is always right," and asked for a culture handbook and a hiring rubric.

1. Refunds (Mon Oct 5). Under $150, techs decide. Over that, Mira decides. A homeowner who cut a line asked for a $220 refund. Mira said no.
2. Overtime (Tue Oct 6). A 4:40 p.m. emergency. Mira: "You three figure out who stays." Two stay; done 8:10 p.m. (3.5 OT hours). Job invoices $680.
3. Vendors (Wed Oct 7). A cheaper condenser saves $95. Lead tech cites 2 callbacks in 3 months. They pay more. At 12 units a year that's $1,140 "saved" vs about $1,600 in callback risk (2 x $800).
4. Mistakes (Thu Oct 8). Tech breaks a thermostat, replaces it for $48, tells Mira after. Own it on the truck; report after.
5. Bad hire (Fri Oct 9). Smiled through interviews. Quit week 4. Training: $22/hr x 40 x 3 = $2,640. Two botched startups: $1,880. Total $4,520.
6. Quiet hire (same day). Luis (made up) texts "en route / on site / done." 0 callbacks on 47 jobs.

Naive run (website values as evidence, no counterexample rule):

1. Theme: "Customer is always right." Contradicts the $220 denial and the $150 threshold.
2. Theme: "Consensus on every decision." Contradicts Mira's refund veto.
3. Rubric: "Enthusiastic communicator who lights up the room." Luis scores 1/4. The polished talker who cost $4,520 would score high.

The page's own runbook already names the fix: "Require examples from meetings, disagreements, tradeoffs, and counterexamples for every theme."

The rerun

Same 6 snippets, with rules: label the sample (owner plus shop-floor meetings only), require a counterexample for every theme, separate "who decides" from website copy, and ban personality-fit language.

Result: 4 themes accepted, 2 cut. That's 4 + 2 = 6 checks. Rubric now scores status texts, same-day ownership of mistakes, evening-emergency willingness, and parts judgment when callbacks exist. Luis clears it. The smile-through-interview profile does not.

Verdict: 8/10

Strengths: evidence, counterexamples, sample limits, and observable hiring criteria are already in the workflow. The runbook calls out flattering copy and personality screens. Useful after a $4,520 bad hire.

Weakness: without the counterexample rule, website slogans win. One owner's meetings still read like "the company." Granola plus ChatGPT is more setup than a one-time hire may need.

Founder tweak: paste who decides refunds, OT and vendors before you ask for themes, and forbid personality language in the rubric.

Copy the MCAFax version

This is our own rewrite, not the original prompt. Copy the original from Wade Foster on How I AI: https://www.chatprd.ai/how-i-ai/workflows/how-to-use-ai-to-define-your-unspoken-company-culture-with-granola

You analyze AUTHORIZED meeting notes for [business name] (label who and what is in the sample).
Do not infer personality. Do not invent values from our website.
For each theme (how we decide, handle disagreement, share context, follow through, fix mistakes):
1. State the observed behavior in one line.
2. Give 1 meeting example with date and quote fragment.
3. Give 1 counterexample or tension from the notes.
4. Say confidence: high / medium / low, and sample limits.
Then list WHO DECIDES for: refunds, overtime, vendor picks, and pricing exceptions. If the notes do not say, write NOT IN SAMPLE.
Cut any theme that rests on one person or one meeting.
After I mark themes ACCEPT or CUT, turn ACCEPT themes into a hiring rubric:
- Observable job behavior only (no "enthusiasm," "culture fit," or "lights up the room")
- One structured question per criterion
- Evidence for strong and weak answers
- Scores 1 to 4 with anchors
- Flag any theme too vague to score fairly

Paste it into any AI assistant with the authorized notes attached, or use it after Granola.

Before you hire from a "culture" rubric

1. Write down who decides refunds, OT, vendors and exceptions.
2. Pull 4 to 6 real meeting notes you are allowed to use. Label the sample.
3. Demand a counterexample for every theme. Cut website slogans with no meeting evidence.
4. Ban personality language from the rubric before anyone scores a candidate.
5. Screen applications against job rules, not vibes: https://mcafax.com/notebook/prompts/job-application-screening

More owner prompts: https://mcafax.com/notebook/prompts