He gave AI a 40-hour Salesforce code review. It finished in an hour Chris Pifer, Salesforce MVP

Most people using AI on a Salesforce build are one bad prompt away from a disaster they won't notice until it's deployed. 500 lines of changed metadata. A security hole. A test that passes because the agent quietly rewrote it to assert that one equals one. Chris Pifer is a Salesforce MVP with 16 and a half years in the ecosystem, self-taught on the platform, and he's spent the last stretch running AI tooling on real enterprise projects with teams of five to ten people building at the same time. Not toy demos. Work where "I just vibe coded it" is not an acceptable sentence. Here's what you'll walk away with: → How to size the blast radius before you type a prompt, and why nobody should ever run this in production → Plan authoring: the written plan you approve before a single piece of metadata changes → Adversarial peer review, where you hand the plan to a second agent and tell it to tear the thing apart → How to catch the lazy agent that makes a test pass by asserting one equals one → Automated browser testing with Selenium or a Puppeteer MCP, and why you should watch it run → The one-command environment setup that provisions the org, verifies it isn't production, and commits a baseline → Sub-agents and context windows: why compacting is what actually ate your guardrails The moment that sticks: Chris was asked to estimate a review of 5,000 lines of Apex sitting on API version 18. He put it at 40 hours for an experienced developer, then gave the AI an hour to help him scope the estimate. He wasn't done with the hour when the actual work was finished. They hadn't even been contracted yet. If you've started using these tools and you're quietly terrified of what they might do to your org, this is the episode that gives you the guardrails. Chapters: 00:00 The disaster you won't notice until it ships 01:05 Meet Chris Pifer, 16 years and self-taught 02:00 What you'll actually learn 04:20 Start with risk tolerance, not the tool 05:45 "500 lines of changed metadata. What did I just do?" 06:00 The five-minute field vs the five-second field 06:50 760 custom fields, and what that math looks like 08:45 The 40-hour estimate that finished in under an hour 10:00 How consulting has to change: teams of three, token budgets 11:50 Cursor vs Claude, and what these tools really cost 15:00 We are still in the subsidized era 16:20 The workflow end to end 16:45 Blast radius: your own scratch org or dev sandbox 18:00 Plan authoring, and why you approve it before any build 21:50 Adversarial peer review: tell the agent to be skeptical 23:00 Lazy agents and the one-equals-one test 24:20 Automated browser testing (Selenium, Puppeteer) 27:20 Git is your life raft 29:20 The one-command environment setup, step by step 33:40 Cursor doesn't even need to talk to your org 34:35 The Salesforce MCP server and the PII problem 38:20 Apex unit tests, even when you built a flow 42:50 Build three plans in parallel and pick the winner 44:20 Dictation: why typing is now too slow 46:40 Human peer review and "let my agent talk to your agent" 49:30 You are responsible for what your agent does 51:20 Audience Q: getting agents to follow your standards 52:40 Sub-agents and managing the context window 54:00 Compacting is why it forgot your guardrails 58:30 Open Source Commons and the day-one-of-a-two-day sprint 01:01:00 Knowing what to build beats knowing how to build it Guest: Chris Pifer, Salesforce MVP LinkedIn:   / chpifer   Get hands-on with the Salesforce Open Source Commons community sprints: https://sfdo-community-sprints.github... Free guide from this episode: the seven-phase build workflow, the adversarial review prompt, and the validation block to paste into every plan, at revenueeng.com The Revenue Engineer Podcast covers people building on Salesforce in the GTM, RevOps, and AI space. New episodes monthly. More at revenueeng.com Thanks for watching. Follow me on Twitter   / thewalters954   Connect with me on LinkedIn   / walters954