Claude Code in the Wild: GTM Newsletter Digest | Mark Colgan

Guest Speaker

Meet Mark Colgan

Mark Colgan

Founder & GTM Consultant, YellowO

Mark Colgan spent years moving through sales, marketing, and agency leadership before landing on go-to-market strategy as his focus. He runs YellowO, a consultancy doing diagnostics and advisory for B2B revenue teams. Recent clients include HiBob and Moss. He is also a mentor in the Bootprint community, where he comes to learn, help others, and connect clients with people who can support them.

Connect on LinkedIn

Session Overview

What was covered

📰

The newsletter overload problem

Subscribed to dozens of GTM newsletters. Never reading them. No system to surface what was relevant. The background anxiety of falling behind when clients pay for your knowledge.

⚙️

The automated pipeline

A Gmail label system combined with a Claude scheduled task. Every Monday at 10am, it reads all tagged newsletters from the past seven days, processes each one, and writes structured entries to a Markdown file.

🖥️

The searchable dashboard

A Claude Code-built UI with keyword search, content-type filters, sort controls, and previous/next navigation. Designed to look like Hampton. Built to read well on an iPad on a Friday afternoon.

🎯

The client-facing diagnostic

A bonus build: a 20-question Talent Health Check for a fractional CMO client. Users get a scored report with recommendations; the client sees an enriched admin dashboard and an AI-generated roadmap.

The Build

What was built, step by step

Tag newsletters in Gmail

Set up email filters so incoming newsletters from chosen senders automatically get a "digest" label. About 30 newsletters, curated deliberately. This is the only step that requires ongoing human judgment.

Create a scheduled Claude agent task

A scheduled task in Claude runs every Monday at 10am. It searches Gmail for all emails with the "digest" label received in the past seven days, then opens and reads each one in full via Chrome.

Process each email with a structured prompt

For every email, Claude extracts a summary, the two to three insights most relevant to YellowO advisory work, a content type tag (tactic, framework, playbook, or case study), key points, a detailed analysis, and a quality rating. The structure was built by feeding Claude 10 test emails and iterating on the taxonomy until it was right, before locking anything in.

Write structured entries to a Markdown file

Each processed email becomes a numbered entry in a single Markdown file. The file also includes a quick-match table that maps client situations to relevant entry IDs, a tag index, and now holds over 350 entries. Mark copies the updated file to GitHub manually each Monday, which also works as a trigger to open and read the digest.

Build the dashboard in Claude Code

Pointed Claude Code at the Markdown file and asked for a searchable, filterable UI. Provided Hampton's website as the design reference. The result: keyword search across 350+ entries, sort by newest or oldest, content-type filters, and previous/next navigation for reading sequentially.

Add a knowledge-base chatbot

A chat interface built into the same dashboard. Strictly instructed to answer only from entries in the Markdown file and not to hallucinate outside it. When asked about direct mail to enterprise accounts, it surfaced two concrete case studies, cited the source email, and linked to the original for full context.

Bonus: the Talent Health Check

Built for a fractional CMO as a lead generation tool. Users answer 20 questions about their people function and get a scored results page with recommendations and a call to action to book a consultation. The client sees an admin dashboard with each respondent's data, company enrichment, and an AI-generated improvement roadmap. Built entirely in Claude Code through prompting, no manual design work.

Mental Models

How Mark thinks about this

Curate your inputs

The knowledge base is only as good as what goes into it. Thirty newsletters chosen deliberately beats an uncurated firehose. Quality of input determines quality of output.

Schema before code

Feed Claude ten test emails and agree on the categorisation structure before building anything. Decisions first, automation second. Changing the schema after 300 entries is much harder than getting it right at ten.

Build what you will actually use

Mark built an annotate-and-save feature, never touched it, and removed it. The simplest version that fits the real workflow is the correct version. Features you imagine using are not the same as features you use.

Internal and external are the same pattern

A personal intelligence tool and a client-facing diagnostic use the same Claude Code build process. The difference is the audience. Companies sitting on proprietary data can turn it into a lead magnet using the identical approach.

Q&A Highlights

Questions from the session

Q Can YouTube videos be included in the knowledge base?

Yes, but you need the transcript first. Save the video to a private playlist on your own YouTube channel, use an API to pull the transcript, parse and clean it, then add it to the Markdown file. Some YouTube videos include an uploaded transcript; others need the API. LinkedIn posts are trickier because LinkedIn frequently changes their page code and blocks scraping access without warning.

Q What about adding LinkedIn posts by pasting the URL into a form?

Technically possible. The main issue is LinkedIn breaking access when they update their page structure. For Mark, the better content depth lives in newsletters anyway. Most detailed LinkedIn posts are excerpts from something longer, usually a newsletter the author is also running. If you want to work with your saved LinkedIn posts, LinkedMash is worth a look.

Q Were the content-type filters preset or did Claude decide them?

A collaboration. Mark made the decisions; Claude did the heavy lifting. He started with ten emails, asked Claude to suggest a categorisation structure, pushed back on anything that did not fit, and iterated until the taxonomy felt right. Then he locked it in and built the Markdown structure around it.

Q Does the Markdown file auto-push to GitHub?

No, and that is intentional. Mark copies the updated file to GitHub manually each Monday. Keeping it manual works as a psychological trigger: completing the copy reminds him to actually open and read the digest. For a client build, he would automate it.

Q Do you handle deduplication as the library grows?

Not for newsletters, because each edition is a distinct piece of content. If you were building with bookmarked posts or LinkedIn saves and might add the same thing twice, you would add a "deduplicate if this entry already exists" instruction inside the scheduled processing step.

Q Have you added your own playbooks into the knowledge base?

Not in this version. Mark uses it to validate suggestions he is already confident about, and to find inspiration when starting work with a client in an industry he has not worked in before. It is a reference layer and a confidence check, not a replacement for his own thinking.

From a newsletter inbox to a GTM intelligence system

Want to build tools like this for your GTM motion?