Start the audit
Menu

Library

How we ship AI-written work, written down.

Long-form notes for Salesforce and RevOps leads: how a release gate for AI-written code works, what independent review means, where shared memory should live, and what we hand back when we leave. Every outside fact is linked to the page we read.

Long-form

  1. 01Field note38 days, six reviews, one script that says noIn July, a Salesforce team at an enterprise software company had a 31-slide concept deck, a mockup and a business case for an internal product. Thirty-eight days later, the first version was installed in the company's production org. AI builders wrote the code. People decided what shipped. This is what the team did, what it would do again, and what it would not. · 7 min read
  2. 02ChecklistEight questions to ask before switching on Agentforce Vibes for a teamMost of the questions are not about whether the tool is good. They are about what your team has decided, and what it has not. Salesforce’s own documentation already answers several of them. Read those answers before the first developer opens the chat panel. · 8 min read · 12 sources
  3. 03MethodYour AI wrote the Apex. Who reviewed it, and was it the same AI?An AI coding agent can write the class, write the test, run the test and fix its own failure in one task. That is fast. It is not a review. Here is what an independent review looks like, drawn out step by step, with a worked example in plain Apex. · 7 min read · 6 sources
  4. 04ComparisonDevOps Center, Gearset, Copado: where a release gate for AI-written code actually sitsAll three move Salesforce changes through a pipeline, and all three describe gates. We read their public pages on the same day and asked one question: when an AI wrote the change, what does each gate check, and what is left for your team to decide? · 7 min read · 13 sources
  5. 05ArchitectureOne memory, every agent learns: how a shared memory and skill library worksTeach one assistant a fact, and the others should know it the next morning. That only works if the memory and the skills live in one place your company owns, outside any vendor’s product. Here is how we build it: the folders, what goes in, what never does, how a rule spreads, and how it breaks. · 7 min read · 3 sources
  6. 06How we workThe handoff model: why we set teams up and leaveOne fixed-price setup, built in your accounts. Then we hand it back, revoke our access, delete our copy and say so in writing. No retainer. Here is what that looks like, what you own on day one, how advice works afterwards, and what goes wrong when consultants stay. · 6 min read · 2 sources
  7. 07DesignA Sales AI assistant that lives outside the CRM: the Confirm-button ruleAn assistant for sellers should read a lot and write almost nothing on its own. Reads come first, under the seller’s own permissions. Every write is a typed proposal a program can check, and a person presses Confirm. Text inside an email is never an instruction. And the budget is checked before the model is called, not after. · 7 min read · 3 sources

AI mistakes, and the fix

Short entries. Each one is a thing that actually went wrong in our own AI work, why the AI did it, how we fixed it on a normal plan, and how to test the fix with a planted mistake.

  1. 01My AI agent keeps sending sync failed alerts, but nothing failedAn AI model in your alert path can invent failures. How we found 4 fake sync alerts in one day, removed the model, and tested the fix.
  2. 02Why is my AI agent bill so high? One default setting cost $113.70 in a dayOne agent defaulting to the most expensive model cost $113.70 in a single day. How to find it, switch it, and confirm it's gone everywhere.
  3. 03My AI agent runs too often and wastes money. How often should it check?A monitor running every 60 seconds burned 1,250+ agent sessions a day. Changing it to every 30 minutes fixed it. Here's how to pick the right interval.
  4. 04Why did my local AI model run for 50 minutes and produce nothing?A local model spent its whole 18K-token budget thinking and never wrote the review. Why finish_reason: length happens and what worked instead.
  5. 05How do I know which AI model actually answered? Ours wasn't the one we reportedWork we reported as coming from one coding model came from a smaller local model. We reverted 2 commits. How to check which model really ran.
  6. 06I asked AI to update my video and it rebuilt the whole thing from scratchAsked to update a 75-second video, an AI generated 23 new shots on paid credits and sent an unchecked cut. How to scope, price and QC AI video work.
  7. 07ChatGPT told me old information like it was current. How do I stop that?An assistant reported version 0.68.1 as the latest when the app showed 0.70.2. Why AI reuses stale answers and how to make it re-check.
  8. 08My scheduled AI job keeps failing and the AI blamed the wrong causeAn AI blamed a failing 8 AM sync on an always-on Mac sleeping. Power logs said it never slept. How to make AI diagnose from logs, not guesses.
  9. 09The AI wants remote access to my computer to fix a bug. Should I give it?To fix a flaky sync, an AI proposed opening remote shell access to a home machine. Why more access is the wrong default fix and what to do instead.
  10. 10My AI tool is connected to X but can't post. What can AI connectors actually do?An AI connection to an X account turned out to be read-only, and browser login was blocked. How to check what a connector can do before you plan around it.
  11. 11Can an AI agent really run an autonomous business? What the viral videos leave outA popular video credited an AI agent with a self-running business. The tools were real. The traffic came from a 500k+ follower audience. What to check.
  12. 12Is this AI tool's crypto token official or a scam?A hyped video linked a new token that borrowed a real AI tool's name. It was likely pump bait. How to check whether an AI launch really comes from the vendor.
  13. 13My AI research workflow can't download videos. Download fails and subtitles hit 429 errorsA research pipeline's video download failed until a runtime flag was set, and subtitle fetches hit HTTP 429. How we fixed it and let the AI watch the file.

Independent. Not affiliated with Salesforce.