Chili Piper

Claude Code for marketing: how we turned sales calls into a content plan

November 20, 2026•8 min to read•Articles

We checked 96 quotes from our content research against their source transcripts. Every one matched word for word.

Then a separate checker read them in context. Twenty-two didn't support the content ideas we'd attached them to. One argued the opposite.

I was using Claude Code to find content ideas in our sales calls. Buyers were telling us what confused them, what they'd tried, and why they were looking for help. I wanted to use those conversations alongside our own experience to write articles people could learn from.

AI helped me search through that material and propose a plan. The 22 quotes showed where our process went wrong. I wanted to understand how it happened and share what I’d change before doing it again.

Claude Code for marketing: which model should read your sales calls?

Sonnet read the full transcripts, Haiku pulled information from other sources, and Fable, the strongest model in this run, used the findings to build a draft plan.

The idea was token maxxing: getting more useful work from our AI budget. We assigned the reading and information gathering to Sonnet and Haiku so Fable could work with their findings instead of processing every transcript. Then a separate agent checked whether the quotes backed up the proposed ideas.

If you're a content marketer or PMM, sales calls are a gold mine for understanding your readers. This workflow can help you find conversations to draw on when you write, with the original sources close enough to revisit.

First, we chose which calls deserved a read

  1. List

    1,849 Gong calls

    Titles, lengths and attendees only. Nobody read them.
    Haiku
  2. Pick

    40 calls

    Chosen by call type and recency.
    Python script
  3. Read

    303 signals

    4 agents, 10 full transcripts each. Quotes copied word for word.
    Sonnet
  4. Plan

    54 ideas, 18 proposed pieces

    Built from the findings plus Notion, Slack and community sessions. Never saw a transcript.
    Fable
  5. Check

    22 of 96 quotes rejected

    All 96 matched word for word. 22 didn't support their idea once read in context.
    Script + checker agent
  6. Write

    Choose the story

    Bring firsthand experience and shape the article.
    Human writer and editor
How the run worked. Cheap tools did the bulk work, Sonnet did the reading, Fable planned from the findings, and a separate check went back to the source.

On November 13, our team listed 1,849 Gong calls from the previous 60 days. That was a list of titles, lengths, and attendees. We hadn't read 1,849 transcripts.

A short Python script selected 40 calls by type and recency, including demos, scoping calls, strategic meetings, and new use cases. That gave us a specific set of conversations to learn from, not a complete picture of our buyers.

That was the first useful division of work: the Python script could filter the call list without an AI model reading every transcript.

Sonnet read the calls; Fable planned from the findings

Four Sonnet agents each read 10 full transcripts in parallel. Each returned a short list of quotes copied word for word, with the call ID, the speaker's role and industry, and the topic they were discussing.

Fable received those lists rather than the full transcripts. It set the research questions and scoring rules, then used the findings to propose a calendar for me to review.

The four reading agents pulled out 303 signals in about five minutes. That timing covers the parallel reading step, not the entire research and planning process.

Each agent used roughly 230,000 to 260,000 tokens across its batch. That is a usage total, not a dollar cost.

claude · content-matrix

> Read the 40 picked calls: 4 Sonnet agents, 10 transcripts each. Return verbatim quotes with call id, speaker role, industry and theme.

⏺ Running 4 agents in parallel

├✓ Gong deep pass batch 0 · sonnet · 28 tool uses · 246.5k tokens · 5m 20s

├✓ Gong deep pass batch 1 · sonnet · 33 tool uses · 234.8k tokens · 4m 46s

├✓ Gong deep pass batch 2 · sonnet · 25 tool uses · 258.4k tokens · 4m 51s

└✓ Gong deep pass batch 3 · sonnet · 26 tool uses · 235.6k tokens · 3m 28s

⏺ 303 signals returned. Transcripts stayed inside the agents.

Recreated from the run's agent logs. The token counts and run times are the real ones.

We also used Haiku to pull information from Notion, Slack, and community-session transcripts, and to fetch the Gong call list. The workflow covered more than the 40 sales calls.

WorkWhat we used
Select calls by title, date, and typePython script
Gather information and fetch the call listHaiku
Read 40 full sales-call transcriptsFour Sonnet agents
Set research questions and propose a content planFable
Check quotes match the source word for wordScript
Check whether quotes support the ideasChecker agent
Choose the story, bring experience, and shape the articleHuman writer and editor

With the research gathered and ideas proposed, we needed to check the connection between them. Did each buyer quote actually support the article we wanted to write?

Why 22 quotes failed the context check

We checked the quotes in two ways. First, a script searched the original sources for each of the 96 quotes. All 96 were there, word for word.

Then another agent read each quote alongside the proposed article idea and the surrounding conversation. Twenty-two quotes were real, but they didn't support the point the article was trying to make.

For example, one idea argued that sales and support should use separate bots. Its supporting quote came from a buyer concerned that two chat widgets would conflict. The next speaker discussed a widget that let visitors choose sales or support.

Give the checker the quote, the point you want to make, and what was said before and after it.

1 · The point you want to make

Your support bot and your sales bot should not be the same bot.

2 · The quote

Buyer“From my experience, you can only have one, right? Like you can't have two different chats that have two different goals on the site because they run into each other.”

3 · What was said right after

Next speaker, from Chili PiperA widget that asks visitors how it can help, lets them pick support or sales, and starts the right chat.

Does the speaker agree with the point?Agrees✕ DisagreesSomething else

That conversation could help us write about how to handle sales and support questions in chat. It didn't tell us that separate bots were the answer. To write something useful, I'd want to ask someone who's managed those conversations what went wrong and how they dealt with it.

The agent grouping ideas had only seen short summaries. The missing conversation mattered: it explained what the buyer was worried about and what the next speaker proposed.

Carol-Lyn Jardine showed a similar problem in an October GTM Clawllective session. She built a pipeline-risk report using sample data, then asked another model to check it against the original files.

A source said a deal was “low risk.” The report called it “no risk.” The checker caught the difference.

In both cases, the source had been used to make a claim it didn't support. For call research, the checker needs the quote, the proposed article's point, and what was said before and after. Checking the wording alone won't catch that mistake.

Two other jobs needed a different check

In earlier Haiku jobs, an agent reported saving 10 AI answers but had saved one incomplete file. Another counted 47 internal links on a page that had 24.

We moved those jobs to Sonnet and gave the counting to a script. Switching models alone wouldn't be enough to trust the work: we needed to check what the agent had saved and whether its counts were right.

What we got, and what we still can't claim

The wider run involved 18 agent runs: six on Haiku and 12 on Sonnet, plus Fable for planning. It produced 54 content ideas and a proposed nine-week plan with 18 pieces. This article was one of the planned pieces.

We didn't track dollar cost per run or compare the same workload across models. I can't tell you how much this setup saved. The token counts describe usage; they aren't a bill.

For where agents fit in the wider budget, see our budget framework for AI agents.

For a marketing leader reviewing work like this, I'd ask to see one proposed article beside the source that supports it. Then ask how the team selected its calls, what the checker rejected, and what the run cost.

What I'd do differently next time

Alongside the quote checks, I'd make three changes to the next run:

  • Open the files an agent says it saved. Check that every expected file exists and contains the full answer. A message saying “done” doesn't tell you that.
  • Use code for counts. Count files and links with a script, then compare the result with the agent's report. Use the models for work that requires reading and interpretation.
  • Track cost from the start. Record the exact model and token usage for each job. Keep estimated API costs separate from actual charges so there's a basis for comparing runs.

Human experience gives content its soul

As a writer, I bring my own experience to what I hear in a sales call: the things I've tried, the mistakes I've made, and the opinions I've changed because of them.

That's why this article tells you about the 22 quotes that didn't support our ideas and the files an agent said it had saved. Those details come from our work. Sharing what went wrong gives another marketer something to look out for when they try it themselves.

AI can generate hundreds of articles from a list of topics. It hasn't felt the frustration of a project going wrong or the relief of figuring out why. A person has to bring those experiences into the story, whether they're the writer's own or someone else's, shared through a conversation.

For me, that's what gives content its soul. I want to read what someone tried, why they cared, where they struggled, and what they now see differently. AI helps me dig through the research. My experience and the people behind those calls give me something worth writing about.

FAQ

What can marketing teams use Claude Code for?

Our team used Claude Code to read sales-call transcripts, gather information from other sources, and draft a content plan for editorial review. The inputs need to be available through authorized tools or files; Claude Code doesn't automatically have access to your Gong, Slack, or Notion accounts.

Can Haiku analyze sales calls accurately enough?

We can't answer that from this run. Sonnet read the full sales-call transcripts. Haiku pulled information from other sources, and we didn't measure each model's accuracy separately.

Test a small batch against the source before expanding the workload.

How many transcripts should each agent read?

We assigned 10 per Sonnet agent. That was our batch size, not a universal limit. Transcript length and the detail you ask the agent to return will affect the workload.

How should I track Claude Code costs?

Record the exact models and usage, including input, output, and cached tokens. Distinguish API cost estimates from subscription usage and actual charges.

Anthropic's cost-tracking documentation explains the available reporting. Our run recorded token totals, but not actual dollar costs.

Want to compare workflows with other marketing teams? Join the next GTM Clawllective session.

See Chili Piper in action

Book a demo