Field Notes
Working notes from inside real AI builds for founders and executives. Each one opens with the short answer, then shows the work.
Subscribe to the Newsletter
How do I check if my live page still matches approved copy?
Compare the approved copy to the live page line by line and sort every difference into BUILT, DRIFTED, MISSING, or UNCONFIRMED. Rank the drifted items by one question: does this line contradict what we sell? Fix those first. An AI prompt can run the comparison in minutes if you forbid it from guessing.
How do I limit what an AI agent can do in my business?
Start every agent with narrower permissions than the task seems to need, and widen them only after it proves reliable. Put a human checkpoint in front of anything touching money, customer data, or outside communication, and log what it decides. Then list every action it can take without your approval and cut that list in half.
How do I handle the 'I already know this' sales objection?
Answer it with a number the buyer can check. When she thinks your offer is an attitude she already holds, replace adjectives with measurable commitments, such as a reply time or a monthly score. She agrees with attitudes and pays for what she has never measured. Audit your sales page for attitude claims and rewrite them.
Why can't ChatGPT see the FAQ answers on my website?
AI crawlers read the raw page text in one pass and never click. Answers hidden inside click-to-open accordions, or stored only in structured code, give an assistant nothing visible to quote. Put each important answer on the page as plain text under a heading written as the question, then confirm it with View Page Source.
How do I fact-check an AI-built presentation before I present?
Have the AI build a sources-and-confidence table with the deck: every claim, where the number came from, and a confidence percentage. Flag anything under 95 percent, cut any claim with no measured baseline, and archive older files that quote conflicting numbers. Read the table before you rehearse a single slide.
Why do my AI-written social posts sound on brand but sell nothing?
Each post is missing a declared belief. Before an AI agent drafts anything, require it to name the one buyer belief the post is meant to change, drawn from a written belief map. A post that cannot name its belief does not get made. That rule makes every caption auditable against what your buyer has to believe.
Why should I audit my documents before AI builds from them?
An AI builds polished output on top of every stale claim in its source files, because it cannot tell a claim has aged. Before you ask for the build, ask for a review: an inventory of what each asset commits you to, flagged contradictions with confidence scores, a promise-to-source table, and a list of open questions.
How do I stop AI from inventing numbers in my business plan?
Label every fact you give the AI as CONFIRMED (with its source), DRAFT (unchecked), or OPEN (unknown). Then add one rule: never produce a total from an OPEN number; write the formula and name the missing variable instead. The model stops guessing and starts asking for the exact input it needs.
How do I stop AI customer service replies quoting old prices?
Keep every price, fee, and policy in one settings file the AI reads, and leave only a placeholder in the reply text. Per-customer details get fill-in slots, and a human approves every reply before it sends. Put the number in one place and the agent can never be more out of date than that file.
How do I turn a training recording into an SOP with AI?
Give AI the transcript and ask for a standard operating procedure with a required structure: purpose, definition of done, buyer profile, positioning, pricing, ordered steps, and a pre-launch checklist. Mark rough numbers as guidance to validate. I converted a 90-minute training into a complete SOP in one working session.
How do I film an online course by myself without a crew?
Finish the thinking in a document before the camera turns on. Ask AI to break your written method into a shot list: one teaching point per clip, a format label on each, arranged in shooting order with checkboxes. One client's full course became a 29-clip checklist that one person can film alone.
How do I prove to clients that my AI system actually works?
Build a dated, verified inventory. List every prompt, workflow, and tool you built, test each against a live task, mark the ones that pass with a verification date, and publish the survivors on one page in business terms. A list of tools is a claim. A verified list with dates is proof.
How can AI turn my expertise into consistent social content?
Write your method on one plain-language page first: its name, the steps, why it works, and one result you can prove. Then ask AI for one format at a time. I turned one page into 50 finished graphics in an afternoon, and every piece stayed consistent because it traced back to the same source.
How do I hand off my AI agents so someone else can run them?
Package them with a one-page Start Here document, a master index, one folder marked CURRENT per system, and a dated, versioned final package. I handed three working agents to another operator this way. An AI system becomes a business asset the day someone else can run it without calling you.
How do I organize a growing library of custom AI tools?
Mark every tool with metadata before you add more. Tag each one with version, date, model compatibility, output quality, how often it gets used, and the job it does. When my Claude skill library grew from 55 to 148, three days of marking turned a scattered pile into a system I can actually use.
Can an AI agent check my YouTube archive for captions?
Yes. Write the job as an SOP, schedule the agent, and require a written report from every run, including the empty ones. My agent checked 38 archive videos for new captions, tested its method on a known-good video first, and reported zero results with a 98% confidence score and a recheck date.
How do I write LinkedIn hooks that make executives stop scrolling?
Borrow the hook structure from a creator who already holds attention in another niche, then rewrite it for your topic. Study the scaffold instead of the content: pick one category, such as a comment trigger, and write five versions in your own voice. I built 30 hooks this way in forty minutes.
How do I get better research results from AI tools?
Give AI a rubric instead of a question. Name five criteria, assign a rough priority to each, and ask for a scored output that flags missing data and the biggest tradeoffs. The structure turns an interesting summary into a comparable decision tool, and the brief takes about eight minutes to write.
What happens to my custom GPTs if the platform changes?
They change or disappear with it, because work built only inside one company's app is rented. Make each assistant portable: copy its instructions and settings into a plain text file, in a folder on a drive you own, named for the job it does. Then your library survives any platform change.
How do I keep track of everything I create with AI?
Build a simple second brain: one home you already open, with a three-line entry for every AI output worth keeping. Write what it is, where the file lives, and one sentence on what it is for. Start today by indexing your last five outputs, then add to it each time you finish something.
Should I save my best AI prompts, and how?
Yes. A prompt that gets the result you want is a small piece of engineering, and discarding it means rebuilding it every time. Keep a prompt journal logging what you asked and why, plus a library of the ten prompts that worked, each labeled with the job it does.
How do I turn a recurring AI task into a repeatable system?
Build an intake: a fixed master prompt plus a short list of scope questions you answer before the AI starts. Write the prompt in six lines (role, context, task, constraints, format, example), add about five questions that set the boundaries of the job, and save it under the task's name.
Why doesn't ChatGPT recommend my business when people ask?
Because the AI cannot read your business as one clear, trusted entity. Brands split across domains with little structured data look like several strangers. Write one bio, put that exact bio on every property you own, and link those profiles to each other so AI reads them as one company.
How do I turn a scattered business strategy into one page?
Give the whole pile to AI with one instruction: put this on one page. Ask for three columns (targets with priority tags, ranked angles, and a dated timeline), move only active items onto it, and close every other tab. The strategy stays the same. The format makes it something you can act on.
How do I make AI-written content sound less like AI?
Run a second editing pass that hunts specific machine patterns and replaces them with plain language. Cut overused tell words, narrator-style drama, stacked hedges, invented "everyone gets this wrong" setups, and sentences that promise value without stating it. Write the checklist once and apply it to every draft.
How do I know if my AI prompts and tools are out of date?
You find out by scoring them on a schedule, because AI prompts, custom GPTs, and automations decay as models change and your standards tighten. Use a short rubric (right model, testable instructions, output quality, versioning, a run trace, current trigger words), rewrite anything below the bar, and repeat the review monthly.