One clear spec, one working workflow: how we turned two customer calls into a week of content

Why we ran this

We argue that the process matters more than the tool. This is that idea on one real workflow: turning customer conversations into marketing, which is the core of our Conversation Content Engine. The test question: if the spec is clear, does the output come out checkable and honest on the first pass?

The spec

This is the instruction we gave, unchanged apart from trimming file paths:

OUTCOME  From two customer call transcripts, produce marketing material a
         small B2B team can review and post this week, where every claim
         traces back to something a customer actually said.

STEPS    1 Insight bank: pains, objections, customer language, proof.
           Every item with the call and the exact quote.
         2 Plan: propose 5 pieces (3 LinkedIn posts, 1 blog outline,
           1 follow-up email). STOP and show the plan.
         3 Write the 5 pieces in our brand voice; list the quotes used.
         4 QA: confirm every quote appears word for word in a transcript;
           list unsourced claims as NEEDS PROOF; flag permission needs.

RULES    Never invent quotes, numbers, results or names.
         Missing proof → write [NEEDS PROOF].  Anonymise companies.
         Plain English, no hype words.

DONE     4 files; every quote verifiable; zero invented facts;
         the plan shown before writing.

The inputs

• Two call transcripts Synthetic: an operations manager at a 15-person software company (Monday reporting), and the founder of an 8-person marketing agency (Friday client reports). Written for the test; no real people.
• A six-line voice guide: plain English, the customer's words, honest about limits, end posts with a question, not a hard sell.

What came out, step by step

1. Insight bank

Pains, objections, exact customer phrases and proof, each tied to a quote. Two examples:

• Objection: "it got a total wrong once and I stopped trusting it." (call 1)
• Customer language: "strategy instead of screenshots" (call 2)

And under Proof: "None. Neither customer states a measured result yet." That honesty came straight from the rule.

2. The plan, shown before writing

Five pieces, each naming the insights it would use. The workflow stopped here, as the spec required; we approved it with one line.

3. The pieces

One of the three posts, as produced:

One wrong total is all it takes. A customer told us she tried pasting her numbers into ChatGPT, but "it got a total wrong once and I stopped trusting it." What would earn that trust back? "If it showed me the numbers it used, so I can check them in ten seconds." That's a good test for any AI report: can the person reading it check it in ten seconds? How do you check the numbers in your reports today?

The blog outline ended with "Results: [NEEDS PROOF]: no measured outcome yet from these customers", instead of a made-up statistic.

4. The quality check

What the spec's rules actually did

• "Never invent… write [NEEDS PROOF]" turned the most common AI content failure, a made-up result, into a visible gap.
• "STOP and show the plan" created a cheap approval point before any writing.
• "Every item with the exact quote" + a QA step made the output checkable in seconds, instead of trusting it.
• The voice guide kept the posts plain and ending in a question, without us editing them.

None of that depends on a particular AI tool. Any capable assistant or coding agent can follow the same spec; what changes is speed and convenience.

What still needs a person

• Permission before quoting real customers, even anonymised.
• Judgment on which pieces to post, and when.
• Proof: the [NEEDS PROOF] gap is filled by real results, never by the AI.

Honest limits of this test

This was one run, with synthetic calls, in one tool (Claude Code), and the spec was written by the same team that ran it. It shows what a clear spec does; it is not a comparison of tools, and real calls are messier than ours.

Use it with your own calls

1. Collect two or three recent call transcripts (most meeting tools export them).
2. Write a six-line voice guide.
3. Copy the spec above; change the pieces to what you actually post.
4. Keep the rules, the plan stop and the quote check exactly as they are.
5. Get permission before anything with a customer's words goes out.

Want the full version, with an insight bank that grows every month? See the Conversation Content Engine.

Common questions