RSS Feed
Spotify 
Apple Podcasts %22/%3E%3Crect%20x=%2217%22%20y=%2217%22%20width=%2230%22%20height=%2230%22%20rx=%229%22%20fill=%22none%22%20stroke=%22white%22%20stroke-width=%224%22/%3E%3Ccircle%20cx=%2232%22%20cy=%2232%22%20r=%227%22%20fill=%22none%22%20stroke=%22white%22%20stroke-width=%224%22/%3E%3Ccircle%20cx=%2241%22%20cy=%2223%22%20r=%223%22%20fill=%22white%22/%3E%3C/svg%3E)
Instagram Context Engineering for AI Agents: The 3-Layer System
In this episode of the GTM AI Podcast, I sit down with returning guest Justin Parnell, CEO of Justin GPT and a forward-deployed engineer who builds AI agents inside startups, local businesses and publicly traded companies. He walks through the context engineering system he runs his own consultancy on, live on screen. We cover:
Why the model is never the problem "You're not really ever going to have the model be the problem." The four context failures behind bad AI output: facts asserted from memory, documents that rot, pages that restate stale numbers, and more than one source of truth.
The three layers of context Operating files (the harness), the canon (declared context) and live data. What belongs in each layer, where each one lives, and how often it refreshes.
Harness vs. context layer Where context engineering sits next to model harnesses, loop engineering and graph engineering, and why a harness without context underperforms.
Business vs. consultancy context engineering One company with one canon, versus a practice that has to keep every client's context isolated so data never lands in the wrong CRM.
The campaign agent that works for a day and a half to two days One campaign brief in. Ad assets, landing pages, sales enablement decks, a CRM loaded with leads and 1:1 copy for around 1,000 people out. Plus why that agent never works without the context layer underneath it.
Staying model-agnostic How to keep context portable in Notion, Obsidian or GitHub so a switch between model providers does not leave your team starting from zero.
Starting from zero Justin's plan for getting scattered files into a context layer in two to three days of weekend work, and his estimate of 20 to 30% off your token budget once it is running.
Live demo: the Friday 8pm routine The cloud routine that refreshes his canon every week, plus the call-ingest and invoicing routines. His recipe for "poor man's context engineering": Notion, Google Drive, GitHub, Supabase and cloud routines.
FREE: JUSTIN OPEN-SOURCED THE SETUP FROM THIS EPISODE
Template repo (click "Use this template" for a private copy): https://github.com/justin-gpt/context-engineering-template
Setup runbook (single-session path with a done-when test for every step): https://github.com/justin-gpt/context-engineering-template/blob/main/docs/07-setup-runbook.md
Sample Notion workspace (duplicate any page into your own workspace): https://app.notion.com/p/3f1dbc7f21e08149a8fed4399166cc3f
Setup skill for Claude Code or Cowork. It interviews you, recommends a stack from what you already run, then walks the build step by step: claude plugin marketplace add justin-gpt/context-engineering-template claude plugin install context-engineering-setup@context-engineering Then say: set up context engineering
Not on Claude? The skill is a plain SKILL.md any agent can run: https://github.com/justin-gpt/context-engineering-template/tree/main/skills/context-engineering-setup
From Justin: "Justin Parnell (Justin GPT) open-sourced the context engineering setup from this episode: a template repo for a governed context layer for AI agents. A thin canon with an owner and a review date on every page, a machine-readable contract agents read, a gap tracker they write to, a metrics catalog that holds definitions but never values, and a nightly sync through a pull-request gate. It covers a single company and a consultancy serving many clients, ships a setup skill that interviews you and recommends a stack from what you already run, and includes a sample Notion workspace to duplicate."