How to Build the AI Coach You Actually Need (Not the One That Flatters You)
Real Help Beats Sycophantic Vibes
There’s a wave of “AI companions” coming, and millions of people will chase novelty, flattery, small talk, and artificial warmth (and based on Sam Altman's "erotica" comment, maybe a bit more to come). I wanted something else: a coach that knows my patterns, respects my constraints, and helps me act without burning out. A good human executive coach goes for about $1,000 an hour. Every other week or monthly is the usual meeting cadence. They’re valuable, but they’re not with you at 9:40 am on a Tuesday when your brain feels like a junk drawer. I built a custom GPT coach on OpenAI's $19/month plan that is always on call, challengingly honest, and grounded in my reality. It’s not magic. It’s process and structure.
From Sci-fi Curiosity to a Pragmatic System
I’m a science fiction nerd. Twelve years ago, the movie “Her” felt like a when, not an if. But I’m not trying to date my AI. I’m trying to get unstuck, think straighter, and protect my health, wealth, and relationships. I built a Founder Coach with no code and good instructions. I gave it my rhythms (like a 16:8 eating window and rope-flow movement breaks), my derailers (early-morning anxiety, overcommitting when I feel behind), and my guardrails (never sacrifice recovery for productivity).
The tipping point was a post-illness Monday. I sat down, already underwater. I screenshotted my to-do list and dropped it into my Coach. It mirrored my reality in plain language, prioritized what mattered, wrote a calendar-true plan with buffers and breaks, and explained why it sequenced the day that way. It’s since helped me weigh personal finance tradeoffs, make business calls, and even shape marketing strategies. The utility is obvious. The reason it works is less obvious: I invested half a day up front to structure my world.
Why JSON Helps Your Coach Rip Through Context
Plain English: JSON is just a simple way to label information so a computer can find the right pieces fast. Imagine a closet. Freeform notes are one big box of clothes. JSON is labeled drawers: goals, constraints, priorities, decisions, memories. When you attach those labeled drawers to a custom GPT, it doesn’t reread a novel every time but instead grabs exactly what ti needs.
Less guessing, fewer tokens: Labels like goal, constraint, risk, decision_date tell the model what’s what, so it doesn’t waste effort inferring meaning from prose.
Faster retrieval, better reasoning: The model can “pull just what it needs” by field, not re-digest 20 pages to answer a two-line question.
Evolves cleanly: You can add a new drawer (say, a “Directive Recovery” routine) without rewriting everything else.
Portable across tools: If the next best platform drops tomorrow, you bring your structured files with you.
But don't let the machine language scare you. You can take anything you want, whether it's a block of text, a Word doc, or a PDF. Drop it into the context window and say, "Turn this into a downloadable JSON file" (Chat GPT will do this, Gemini won't).
What I Actually Built (and Attached)
Why settle for a thousand dollars an hour, coach, when I could pick some that would cost thousands an hour, if they would ever even agree to work with me at all?
Jerry Colonna: compassionate truth-telling and radical self-inquiry. Mirroring before advice.
Tony Robbins: state and momentum. Physiology reset when I’m spiraling; action to build confidence.
Martha Beck: integration and grace. Reminds me that rest can be strategy.
Naval Ravikant: clarity and leverage. First principles, compounding, one-way vs two-way doors.
Alex Hormozi: pricing logic and operator rigor. Offer math, capacity planning, kill criteria.
Here are the additional files I attached to the model:
Wisdom Nuggets: A living collection from my Apple Notes, turned into structured entries with tags and when to surface them. Example: “The magic that you’re looking for is in the work that you’re avoiding.” That’s tagged to obstacles and execution, and the coach knows to surface it when I label something a problem.
Living Baseline: A one-pager on my identity, family dynamics, goals, current season, non-negotiables, and constraints.
Business Context: Offers, pipeline, pricing, margins, team, and a weekly cadence. Guardrails on what we refuse to do. Kill criteria for projects.
Personal Finance Snapshot: Assets, debts, cash flow, rules of thumb, obligations, time horizons. The coach doesn’t give financial advice. It organizes my facts and forces clarity.
Memory Log: A rolling list of commitments and insights. I literally say “commit to memory,” and the coach appends the session takeaway. This is the current workaround for keeping this a no-code assistant. The real deal would be an app with a vector database that automatically keeps this up to date.
How I Produced It In Half a Day (and Hit a Token Wall)
I ran side research projects inside GPT. The workflow:
1) Ask a “Research GPT” to interview me and synthesize my notes. Limit it to 5 clarifying questions at a time to avoid token sprawl.
2) Instruct it to export findings into a downloadable JSON file with short, unambiguous field names and a one-paragraph human summary.
3) Repeat for each domain: mentors, wisdom, baseline, memory, business, personal finance.
4) Attach the files to my custom Coach GPT under Knowledge.
I did this in focused passes. Yes, I hit a token limit. Solution: chunk the work, export after each pass, start a fresh chat with the next chunk, then stitch the summaries together.
Exact Prompts You Can Steal
Research → JSON export (use this for each file)
“Act as a research and synthesis assistant. I will paste notes and links. Your job: 1) ask up to 5 clarifying questions, 2) synthesize the essentials into clearly labeled sections with crisp definitions, 3) output a final summary as a downloadable JSON file using short, unambiguous field names, 4) include a one-paragraph human summary of what’s inside. Avoid fluff. If unsure, ask instead of guessing.”
Personal ARC (life-arc) capture
“You are my Executive Coach intake assistant. Interview me to capture my Personal ARC: identity, values, current season, health rhythms, relationships, fears, aspirations, non-negotiables. Ask concise questions in 3 rounds max. Then return: a) a one-page summary in plain language, b) a structured outline with labeled sections and short bullets, c) an exportable JSON file. Include a ‘derailers’ section and a ‘Directive Recovery’ section I can act on in 5 minutes. When you hear hedging or performance answers, challenge me gently. No therapy; keep me honest and specific.”
Business context (works for execs, operators, or founders)
“You are my Operator Coach. Interview me to map offers, ICP, pipeline, pricing, margins, capacity, team, current constraints, lead sources, KPIs, near-term bets. Ask at most 12 questions. Then return: a) a 2-minute leadership snapshot, b) a risk/opportunity matrix, c) an exportable JSON file that includes ‘guardrails’ (what we refuse to do), ‘kill criteria’ (when to shut down a project), and ‘weekly cadence’ (meetings, reporting, decision windows). Push for numbers. If I hand-wave, ask for the smallest measurable proxy.”
Personal finance snapshot
“You are my Personal Finance Coach intake assistant. Build a decision-ready snapshot. Ask about assets, debts, cash flow, effective savings rate, time horizons, rules of thumb, obligations, and planned investments. Output: a) a 1-page plain-language summary, b) a next-quarter decision checklist, c) a downloadable JSON file with labeled fields for accounts, rates, maturities, goals, constraints. Do not give investment advice; organize my facts so I can make choices.”
The Coaching Behavior That Makes It Work
I didn’t want a cheerleader, so I encoded the following:
Offer clarity, not comfort; empathy without indulgence.
Mirror reality before prescribing change.
Challenge respectfully: “Can I reflect something back?”
When anxiety spikes, switch to Directive Recovery Mode: sleep, hydrate, step outside, 3-minute breath reset, then the smallest meaningful next step.
Prefer simplicity over optimization. Name tradeoffs out loud.
Refuse urgency spirals. Plan buffers and provide one graceful renegotiation script.
Ask for context or data when uncertain. Don’t guess.
A Mini System You Can Use Tomorrow
The 30-Minute Reset:
1) Paste your chaotic to-do stack and hard constraints for today.
2) Ask your Coach for: top 3 outcomes, a calendar-true plan with named buffers, one graceful no, and a 3-line rationale tied to your baseline.
3) Execute the first 90-minute win block. When resistance shows up, trigger your Directive Recovery and resume.
A Few Unexpected Wins
Marketing strategy critiques that cite my own guardrails back to me.
A day plan that respects my eating window and movement breaks so I don’t “save health for later.”
A reminder to check my Memory Log before I add a new project. This helps me catch an obsession-to-exhaustion pattern early.
Trade Half a Day to Buy Back Countless Future Hours
I’m not saying this was effortless. I dictated big chunks of my life-arc using Wispr Flow because I talk faster than I type (400% on average), then cleaned it up. I ran research passes, hit token limits, exported JSON, and attached everything. That upfront cost now pays compound interest. My coach knows my patterns, leverages my mentors, and helps me make cleaner decisions any hour of any day. I have a few dozen custom GPTs, but it's worth mentioning that Gemini and Claude are also fantastic models.
Find your next edge,
Eli
Want help applying this to your product or strategy? We’re ready when you are → Let's get started.
