# The Discovery Path ### A multi-stream audit of your working life that ends in a 90-day plan aligned to your ikigai Version 2 · 2026-06-10 · Pairs with the persistent-memory build guide at [eqctrl.io/karpathy+](https://eqctrl.io/karpathy+) **Hand this document to your AI.** It is written for an AI assistant walking a human through their own discovery. Any current frontier model works; the pipeline assumes a human with judgment and an AI with stamina, and neither substitutes for the other. Picture the Venn diagram of your life: things you love, things you are good at, things that pay. Then picture the fourth circle, the honest one: things you actually do. For most working people the fourth circle has drifted away from the other three, quietly, a calendar entry at a time. This path measures that drift with evidence, then closes it with a plan. Your AI runs the streams; you make the calls. The output is not a website or a rebrand or a productivity system. It is a dated, adopted, executable 90-day plan in which everything you do has been given an honest verdict about which circles it lives in. The author ran this on his own practice. The audit, the plan, and the diagnosis lens below are real artifacts from that run, with the specifics stripped so you can pour your own life into the shape. --- ## 00 · How to use this 1. Read "Pick your path," choose your entry, skip what does not apply. 2. Every stage lists **You do** (judgment, minutes), **Your AI does** (the heavy lifting), the **Artifact** it must produce, and the **Gate** that closes it. A stage without its artifact did not happen. A gate without your explicit sign-off does not open. 3. **One chat window is enough.** The stages describe parallel agents because that is fastest, but every fan-out runs fine sequentially: each stream as its own conversation turn, each critic as a fresh prompt. Slower, same shape. 4. If your AI has persistent memory (a wiki, a context file it reads every session), everything compounds. If not, the path still works: keep one running "state" document and re-paste it each session, and consider building the memory layer first (the companion guide above; it takes an afternoon). --- ## 01 · Pick your path **What does the drift feel like?** - **"I'm busy constantly and can't say where it goes."** Run everything, in order. Your telemetry stream (Stage 1) is the one that will hurt and help most. - **"The work is fine but the money is wrong."** Run all streams but weight Stage 4's value-chain lens; your answer likely lives at the front of the chain, not in the work itself. - **"I'm burned out."** Run everything, and tell your AI up front, because it changes the completeness critic's brief: rest, health, and the people you live with become first-class audit lanes, not footnotes. - **"I'm pivoting and don't know what to carry forward."** The kill/park/hobby verdicts in Stage 6 are the whole game for you. Be ruthless in the interview. **Solo or team?** Solo: run as written. Small team: one interview round per principal, decision rules get named owners, and the verdicts in Stage 6 require everyone whose hours are being reassigned to be in the room. **Have time telemetry?** Calendar history, session logs, screen-time data, even one honest week of manual notes. Yes: feed it in raw. No: start a seven-day log today and run Stages 1 through 4 on inventory evidence meanwhile; the plan gets written when the week of truth exists. Recalled time lies; measured time does not. All paths converge on Stage 6, the plan, and section 09, the re-entry rhythm. --- ## 02 · Stage 1: The evidence streams The audit examines the operator, not the tools. Multiple independent streams, gathered before any judging begins, because conclusions formed before evidence are just moods with formatting. **You do (about an hour):** hand over the raw material. Time telemetry. Accounts in rough numbers. The full list of projects, clients, commitments, and the folder of half-finished things you avoid looking at. Do not curate; curation is the drift defending itself. **Your AI does:** fan out one stream per domain of your working life, each producing an inventory with the same fields per item: what is it, which pipeline stage, hours it consumes, revenue attached, papered or unpapered, last touched, owed items dragging on it. The streams that earn their keep: - **Offerings**: everything you sell or could sell, including the things you do free that someone would pay for. - **Time**: where the hours actually went, contrasted against where you believe they go. Expect a gap; the gap is a finding. - **Money**: what each engagement actually nets per hour once the untracked overhead is counted. - **Infrastructure and content**: the systems and public surfaces you maintain, and which ones still earn their maintenance. - **Pipeline**: every live thread that could become work, with its age and last touch. - **The life lane**: household, partner, health, rest. Audited like the rest, because the plan that ignores them gets quietly vetoed by reality within a month. **Artifact:** the stream inventories, evidence attached, no verdicts yet. **Gate:** nothing judged, everything counted. Verdicts rendered during inventory are contaminated; if your AI starts concluding, tell it to get back to counting. --- ## 03 · Stage 2: The adversaries Findings are claims until something hostile has tried to kill them. Three roles, run against the streams, as separate agents or as three fresh prompts in sequence: 1. **The skeptic** re-verifies every claimed finding against the evidence. "You said this client is unprofitable; show the arithmetic." 2. **The operations lens** asks whether a tired human could actually act on each finding. A true finding with no executable response is trivia. 3. **The completeness critic** hunts for what nobody mentioned: the tax deadline, the unsigned agreement everyone forgot, the inbound channel nobody checks, the partner who has opinions about the plan, your own sleep. In the author's run this role out-yielded every other and changed the top recommendation. Budget for it accordingly. **Artifact:** findings tiered by severity (the existential few, then the structural leaks), each marked verified, refuted, or human-input-needed. **Gate:** every finding has survived an attack or been amended by one. --- ## 04 · Stage 3: The interview The streams hold the facts; you hold the meaning. A fixed-length structured interview, your AI asking, around eight questions, recorded as deltas against the findings. It covers two things the evidence cannot: **Facts only you hold.** The contract that is signed but never filed. The thread that looks dead but has a warm relationship behind it. The project you cannot remember the purpose of, which is an archive signal; let it go. **Wants only you hold.** This is where ikigai enters the record, as data: Which of these items would you do for free forever? Which do you dread regardless of pay? What did you do last time you had a free Saturday and nothing owed? If one project had to absorb your best weekday hours for a quarter, which one makes you sit up? What does "enough money" actually mean for the next year, as a number you would accept rather than a fantasy? Answer plainly. The interview is the only stage where being impressive hurts you. **Artifact:** the interview record, appended to the findings as corrections and as the want-data the verdicts will honor. **Gate:** your AI can state, in one sentence each, what you love, what you are good at, what pays, and where your hours actually go, and you agree with all four sentences. --- ## 05 · Stage 4: The second lens One analysis can be wrong with confidence. So run an independent second pass over the same evidence through a classic frame, and require it to reproduce the verdict. Where the two lenses agree, trust them. Where they disagree, the disagreement is the most important finding you have. Adapt the value chain to a one-person practice. Primary activities: **Origination** (how work finds you) → **Scoping and Contracting** (pricing, papering) → **Production and Delivery** → **Retention and Expansion** → **Productization** (turning bespoke time into repeatable offers). Support layer: business and financial ops, capacity (you plus your AI are the team), tooling, inputs. Your AI fills a matrix: every activity gets its cost drivers, its current state with evidence (counts and ages, never adjectives), and its potential. Then read the *shape* before the content. The commonest solo-operator topology: full in the middle, empty at both ends. World-class production, starved origination, leaky papering. If that is you, the diagnosis writes itself: the highest-value hours are not more production, they are the unglamorous front of the chain. Every synthesis point must end in a one-sentence "Move:". Analysis is not allowed to end without an imperative. **Artifact:** the value-chain read, with each offering scored and the shape named. **Gate:** the second lens reproduced the audit's verdict, or you have consciously chosen which lens to believe and written down why. --- ## 06 · Stage 5: Competing plans Do not let your AI write "the plan." Have it write three, as rival candidates with named theses: one revenue-first, one leverage-first, one focus-first. Each commits to a quarter-length horizon, a week-one, and what it kills. Then a synthesis judge, a fresh prompt with no loyalty to any candidate, scores them against the findings and the interview's want-data, picks a winner, and grafts the best ideas from the losers, keeping count of what it grafted and rejected so the merge is inspectable rather than vibes. The want-data matters here as much as the money data. A plan that maximizes revenue while scheduling away everything you love is not a plan; it is a resignation letter with extra steps. **Artifact:** three candidates, the judge's scoring, and the merged winner. **Gate:** you can articulate why the losing plans lost. If you cannot, the judge decided nothing and you are about to adopt a coin flip. --- ## 07 · Stage 6: The plan, compiled to executable The merged winner becomes the 90-day plan, and the difference between a plan and an essay is entirely in the compile. The required parts: - **Two-paragraph TL;DR**: the verdict, then the prescription, closing with the one operating principle the whole plan obeys. - **Week one as a table**: numbered actions, each with an owner split ("AI drafts, you confirm"; "not delegable"), a time budget in *your* hours, and a hard date. Statutory deadlines anchor the week. - **A slip order**: which commitments fold first when the week overloads, and the never-slip items, decided now while calm. - **The ikigai ledger**: every inventoried item gets exactly one verdict, and the verdict is a circle assignment. - **KILL**: lives in no circle. Delete it; do not archive guilt. - **PARK**: real value, wrong season. Dormant behind a written re-entry gate, with the reversal path recorded. - **HOBBY**: you love it and it does not pay, so it is *protected*, fenced to non-prime hours precisely so it stays love instead of curdling into obligation. The fence is a gift, not a demotion. If evenings stop being fun, cut scope, never the fence. - **BET**: circles genuinely stack. Gets prime hours, but only while its next action is critical-path to a paying or decision event, and only one bet holds prime hours per week. Each bet carries a pre-scheduled three-state resolution: scale, niche, or back to hobby. No fourth state, no "let's see." - **THE FLOOR**: the stable earner. Hold it at minimum viable attention; do not over-polish what is already bought. - **Standing decision rules**, numbered and bold-named so future sessions invoke them by name: revenue asks before polish; a hard rate floor with contract-before-work; an hour ratchet on anything unpapered; the new-project intake fence (everything enters as hobby unless it arrives with a paying customer; promotion displaces an existing bet, so sprawl is structurally impossible); a published-numbers gate (no load-bearing number on a public surface without measuring it first); no coverage gaps (never cancel a working structure until its replacement is verified live). - **Hour budgets per lane**, denominated in operator-hours, because your scarce resource is you, not compute. - **Risks, each pre-paired with its mitigation**, with operator burnout treated as a load-bearing constraint, not a wellness aside. **Artifact:** the plan document, dated, with the provenance of how it was made written into its header. **Gate:** adoption, the same day. Into your task tracker, kill-batch executed, partner-of-record informed. A plan that sits unadopted for a week is a rejection you have not admitted to. --- ## 08 · What this is not This path produces the plan. It does not execute it. Execution, shipping the revenue surface, refreshing the identity, building in public, is downstream work with its own craft, and trying to do it inside the discovery corrupts the discovery: you start auditing toward the conclusion you are excited to build. Finish the verdict first. Then go build, with the plan as the referee. Two pointers for the road: the memory layer that makes all of this compound across sessions is the companion guide at the link up top. And the single most portable execution habit is the one this path already taught you: before any substantial plan executes, let hostile critics maul it, because revision after critique is the cheapest fix you will ever buy. --- ## 09 · The re-entry rhythm The path is not a one-time event. Re-enter at smaller scale whenever the fourth circle starts drifting again: - **Weekly**: the plan's week table gets refreshed; slip order applied without renegotiation. - **Quarterly**: a mini-audit, one telemetry pull, verdicts re-checked, bets resolved to their three states, adjust once. No standing scorecard bureaucracy. - **On drift**: when you notice a month of hours went somewhere no verdict authorized, that is not a failure of discipline, it is a signal the ledger is stale. Re-run the interview and re-cut the verdicts. --- ## 10 · Failure modes to preempt Each of these happened in the author's run or its surrounding weeks: 1. **Verdicts before evidence.** Judging during inventory contaminates the count. Count first. 2. **The skipped completeness critic.** The deadline nobody mentioned is the one that detonates week one. 3. **The flattering interview.** Impressive answers produce a plan for the person you wish you were, which the person you are will quietly refuse to execute. 4. **The unledgered number.** An impressive figure shipped unverified and was wrong by a factor of six. Measure or cut. 5. **The unadopted plan.** Written, admired, never entered into the tracker. That is a rejection wearing a compliment. 6. **The starving hobby fence.** Monetizing everything you love until nothing is load-bearing *and* loved. The fence exists so at least one circle stays pure. 7. **The silently dead background run.** A long agent job reported as "still working" had died in minute two. Demand liveness proof, not status optimism, and check on anything that has been quiet too long. --- ## 11 · Trigger phrases Teach your AI these, or your own dialect: - **"count, don't judge"** returns a stream to inventory mode. - **"run the critics"** invokes the three adversaries on whatever is on the table. - **"circle check"** asks, of any item: which circles does this live in, and does its verdict still match? - **"hobby-fence it"** assigns the protective verdict on the spot. - **"adopted"** is the only word that closes Stage 6, and it comes with the tracker entry to prove it. --- ## 12 · The shape of the whole thing Gather the streams without judging. Let adversaries attack the findings. Tell the interview the truth, including what you love. Make a second lens reproduce the verdict. Force rival plans to fight, and adopt the winner the same day, with every piece of your working life assigned to its honest circle. Then re-enter, smaller, whenever the drift returns. The fourth circle never stops drifting. The point was never to pin it permanently; it is to own the instrument that measures the drift and the rhythm that corrects it, with your AI holding the streams together at the center, like a small crab gripping four circles at once. Adapt freely. Credit if you publish a fork. *Based on a real run, June 2026 · [eqctrl.io](https://eqctrl.io) · pairs with the persistent-memory guide at [eqctrl.io/karpathy+](https://eqctrl.io/karpathy+)*