Capture an expert. Run them in any AI.
I Eat AI for Breakfastis the movement for people who’d rather augment their experts than be replaced. Under the hood: SMETP, an open protocol that turns a 5-minute voice interview into a portable skill.md you can run in Claude, ChatGPT, Gemini, Perplexity, Claude Code or any MCP host.
--- skill: route-exception inputs: [queue, risk_cue] cites: [workflow.step.2] runs_on: [claude, chatgpt, gemini, mcp] --- when the exception queue spikes, check the risk cue first, then …
One file · every model
skill.md is the universal adapter.
Capture once. Drop the same file into Claude as an MCP tool, into ChatGPT as a Custom GPT, into Gemini as a function-calling spec, into Perplexity as a grounded skill, or into Claude Code as a CLI runner. No re-training. No vendor lock-in.
Bring your own keys. The protocol is the product — the providers are interchangeable.
Pick your path
Who are you, and what do you need to capture?
Same protocol, four entry points. Each one pre-loads the right ontology pack and interview style.
I'm an expert
Capture what you know.
5-minute voice interview. The graph builds as you speak. Walk away with a portable skill.md and a Twin that answers in your voice — usable in Claude, ChatGPT or Gemini.
Start the discovery
My expert is leaving
Beat the bus factor.
Run a structured interview before they go. Capture decisions, dependencies, the spreadsheets and tools they touch — before tribal knowledge walks out the door.
Discover the SME
I run a team
Augment every role with their best SME.
One protocol across the company. Top performers become reusable workflows, dictionaries and skill.md files your existing AI stack can call.
Start with one SME
I'm a builder
Use the open protocol.
SMETP ships as MIT-licensed npm packages — @smetp/spec, @smetp/sdk, @smetp/cli, @smetp/runtime. BYO model. Run interviews from a terminal.
Read the specHow SMETP works
From conversation to runnable agent in three steps.
Talk for 5 minutes
The voice agent interviews your SME. As they speak, people · decisions · artifacts · dependencies appear in the live graph on the right.
Compile to skill.md
One click. The capture becomes a SkillDocument v0.4 — a runnable workflow + dictionary + ontology + persona, stored in your knowledge graph.
Plug into any AI
Open /twin/<id> for chat, drop the file into Claude, ChatGPT or Gemini, or have your agents call /api/twin/ask — every reply cites the source it used.
The human behind the protocol
I’m Goker. I’d rather you eat AI for breakfast than the other way around.
I wrote SMETP as a side-project because every “AI transformation” deck I read kept skipping the part that matters: the people who actually know how the work works. The method is open, MIT-licensed, and runs without me — three drops a week land here automatically so the people closest to the work stay sharp.
“Most AI tools want to replace your expert. SMETP captures them so the AI you already pay for finally gets useful.”
SMETP · out in the wild
On the stage, in the studio, in the lecture hall.
SMETP isn’t theory — it’s a method that gets used. The pictures below are from open talks, classes, and a few captured SMEs. None of it is for sale; it all ends up as a Mon / Wed / Fri drop on the Breakfast Club.
Reviewing the protocol
Dilek reviews the engineering — 25 years of turning enterprise data into shipped product.
Dilek Tapucu, PhD sense-checks the AI / data engineering side of the open SMETP method. MIT Sloan GFSA alumna, computer-science PhD on ontology-based databases, and 11+ years at ING Bank Türkiyerunning data strategy & analytics architecture and the bank’s innovation programmes — hackathons, fintech partnerships, R&D centre work. Earlier in her career she helped start SomaTech, a sentiment-analysis project, and published research in NLP and information retrieval.
Translation: when SMETP claims it can ship a skill into a real production system without breaking compliance, Dilek is the person who pushes back on the parts that wouldn’t actually survive a bank’s data pipeline.
- PhD · ontology-based databases
- MIT Sloan GFSA alumna
- 11+ yrs ING Bank Türkiye · data strategy
- Previously co-founded SomaTech (NLP startup)
- Published NLP & sentiment-analysis research
- Hackathons, fintech, innovation R&D
The content engine
The site writes itself, three times a week.
The Breakfast Club is run by a content pipeline, not a person on a deadline. A daily Vercel cron drafts every empty Mon / Wed / Fri slot with the SMETP voice guide. An hourly cron publishes whatever’s due. Both reads come straight out of the same Supabase calendar that powers /club and the admin planner.
- Mon · Tactic of the week
- Wed · Workflow story
- Fri · Future take
/api/cron/social-generate · 04:00 UTC/api/cron/social-tick · every hourNo published drops yet — the calendar will start filling on the next cron tick. The drafts are deterministic, so re-running the cron is a no-op.
See the upcoming calendarThe file is free. The persistence costs.
Try it free. Save your first Twin for $19.
Capture an SME and walk away with a real skill.md on Taste — no signup. Save it forever for $19. Run unlimited Twins on Studio at $49/mo. Or self-host the whole stack with the open SDK.
- One full Twin per browser session
- Live skill.md · copy/download
- Side-by-side test vs vanilla LLM
- No signup, no card
- Save it forever to your account
- Hosted /twin/<id> chat + voice
- Public or private sharable link
- Refine unlimited times
- Unlimited Twins · team workspace
- One-click deploy to Claude/GPT/Cursor/MCP
- Voice included · 25k API calls/seat
- Or $399/yr (save ~32%)
- Run the @smetp/runtime on your stack
- Bring your own LLM + storage
- No license cost, no lock-in
- MIT-licensed forever
The skill.md file is open and MIT-licensed — copy it, host it yourself with @smetp/runtime, no lock-in is even possible.
Created by Goker · open source method
SMETP is the method. I Eat AI for Breakfast is the movement around it.
SMETP defines how expertise is captured, structured, validated and compiled for modern LLM stacks. The hosted product makes that method self-serve for companies and individuals who want practical AI transformation, not slideware.
- Capture expert decisions and artifacts.
- Compile SkillDocument, workflow, dictionary and ontology.
- Run with Claude, ChatGPT, Gemini, Perplexity, Claude Code or any MCP host.