| Filename | Latest commit message | Latest commit date |
|---|---|---|
Companion to shivaatripathi.com/blog/team-ai-brain-not-personal/ Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> |
||
| channels.md | ||
| decisions.md | ||
| experiments.md | ||
| glossary.md | ||
| icp.md | ||
| LICENSE | ||
| positioning.md | ||
| README.md | ||
| voice.md | ||
Team Brain Template
A shared, versioned context repo for a marketing team working with AI agents. Eight markdown files, one owner each, read by your AI tools on every task - so context gets written once and reused everywhere, instead of five teammates maintaining five drifting personal setups.
Companion to the essay: Your marketing team needs a shared AI brain, not five personal ones.
Why this exists
- Five personal AI setups relearn the same facts five times. ICP, positioning, voice, what was tried and killed - each teammate re-explains them to their own assistant, at different levels of staleness.
- Wikis rotted because nothing complained when they went stale. This repo is read by the model on every task, so a stale line produces visibly wrong output within days, and the fix is one commit.
- Corrections compound. Fix a file once and every future agent run, by every teammate, inherits the fix.
The eight files
| File | What it holds | Owner | Review |
|---|---|---|---|
| icp.md | Who buys, who blocks, per segment | Demand lead | Quarterly |
| positioning.md | The one-paragraph story and proof points | Founder / CMO | Quarterly |
| voice.md | Voice rules and banned words | Content lead | Stable |
| channels.md | What runs where, per market | Channel owners | Monthly |
| experiments.md | Tried, result, killed-why | Everyone | Per experiment |
| decisions.md | What was chosen and why | Whoever decides | Per decision |
| glossary.md | Terms agents must not garble | Anyone | As needed |
| README.md | This file: how to load and edit the repo | Team lead | Stable |
Every file ships with structure, fill-in guidance, and worked example entries for Meridian Analytics, a fictional B2B SaaS company. Replace the example content with your own.
Standing it up: five days
- Mon - Skeleton + owners. Clone this repo (private copy), name an owner per file.
- Tue - Dump, don't polish. Each owner fills their file from decks, docs, and memory. Ugly is fine.
- Wed - Fight about it once. One hour, whole team: resolve every contradiction between files. This is the day that matters.
- Thu - Wire it into the tools. See below.
- Fri - Run one real task through it. Draft something real. Every wrong line is a file to fix, once.
Wiring it into your tools
Claude Code / Claude Cowork: add a CLAUDE.md to the repos or folders where your team works, with a line like:
Before any marketing task, read the team brain at <path-or-url-to-this-repo>.
icp.md, positioning.md, and voice.md are binding. Check experiments.md
before proposing any campaign or test - do not re-propose killed experiments.
Record new decisions in decisions.md.
Claude Projects / ChatGPT Projects: add the eight files as project knowledge; refresh when they change (git history tells you when).
Any other agent framework: the files are plain markdown; load them as system context in file order: positioning, icp, voice, then the rest on demand.
House rules
- One owner per file. Shared ownership is how wikis died.
- Edits happen here, not in personal setups. If an agent got something wrong because a file is stale, the fix is a commit, not a private prompt patch.
- Killing an experiment requires writing the line. The experiments log only works if it is a habit, not an archive.
- Small enough to stay true. If a file grows past what an owner can verify in one sitting, split it or prune it.
License
MIT. Take it, rename it, make it yours.