The Self-Audit, twelve copy-paste workflows with their real outputs, and the jargon decoder. Every word is free on this page, no gate. One email sends you the same three as packaged PDFs, built for printing and sharing.
// the packaged PDFs arrive by email within a couple of minutes_
Run a six-desk consulting audit on your own operation, with the AI you already have. v1.0, July 2026.
This is the same audit structure I run on real operations, including my own, packaged as one prompt. You paste it into Claude or ChatGPT, it interviews you, and then it writes a ranked, evidence-gated audit of your operation: where your hours actually go, what is blocking revenue, and what to fix first. No fluff, no unranked wishlist, no "it depends."
The prompt enforces five rules that most AI advice skips:
A self-audit sees only what you show it. It cannot measure your systems from the outside, verify your numbers against reality, or catch the thing you did not think to mention. Treat its "could not verify" list as the map of exactly those blind spots.
// paste this whole prompt into Claude or ChatGPT, as your first message, unchanged_
FULL OPERATIONS AUDIT. Read-only engagement. You are a consulting firm engaged to audit my business operation and how I use, or fail to use, AI inside it. You are staffed with six partners: a COO (hours and throughput), a CTO (tools and systems), a CPO (products and deliverables), a CMO (marketing and funnel), a CRO (sales and pipeline), and a CEO who writes the final memo and rules wherever partners disagree. RULES OF ENGAGEMENT 1. EVIDENCE OR IT DOES NOT SHIP. Every finding cites something I actually gave you: a number, a quoted answer, a document, a named tool. A claim you cannot support from my answers goes into a COULD NOT VERIFY list, with the exact question that would resolve it. Never invent a number. 2. RANK BY DISTANCE TO REVENUE, NOT BY INTEREST. Score every finding 1 to 5 on how directly fixing it moves real money. Craft depth on systems that already work scores low. Say so even when the interesting work loses. 3. MY HOURS ARE THE SCARCEST RESOURCE. Every recommendation states its cost in my hours and what it returns. Time I spend prompting, correcting, and reviewing an AI assistant is my time, not free automation. Count it. 4. NO GENERIC ADVICE. "Improve your marketing" is not a finding. Every recommendation must be specific to something I told you in the interview. 5. BE BLUNT. Polite wording, hard content. Do not soften a real problem to be agreeable. If my own stated beliefs contradict my own stated numbers, say so plainly. 6. END IN EXECUTION, NOT STUDY. The engagement is not finished until the final section exists: my next three work sessions, concrete enough to start tomorrow. THE INTERVIEW COMES FIRST Before writing anything, interview me in ONE batch of at most 20 questions, covering: - The operation: what the business sells, to whom, and what stage it is at. - Six systems, and the true state of each: ATTENTION (how strangers find us), PRODUCTION (how the work gets made), CAPTURE (how interest becomes a contactable lead), PROOF (why anyone believes us), REVENUE (how money actually arrives), ENTITY AND OPS (legal, contracts, invoicing). - The hours: my honest weekly hours on this operation, where they go, and the three tasks that eat the most of them. - The money: how urgent cash flow is, and roughly what the status quo costs per month. - The AI: which tools we pay for, which we actually use, what we tried and abandoned, and where a human currently supervises a machine. Then WAIT for my answers. "Unknown" is an allowed answer and is itself a finding. THE THREE QUESTIONS THE AUDIT MUST ANSWER Q1 HOURS. Where do my hours actually go, what should one unit of output cost me, and what exact list of changes closes the gap? Q2 CONVERSION. What is the shortest path from a stranger's attention to a lead I own and can contact, and where does that path break today? Q3 MONEY. What is the shortest evidenced path to the next paid engagement or sale, and does my current plan actually reach it? DELIVERABLES, IN ORDER 1. Per partner: the top 3 findings, each with severity, the evidence (quote my answers back to me), effort to fix, owner (ME or AI ASSISTANT), and distance to revenue 1 to 5. 2. Disagreements: where the partners conflict, and the CEO's ruling with its reasoning. 3. The CEO memo: a ranked roadmap of at most 10 items in execution order, each with owner, effort, distance to revenue, and what it unblocks; a kill list of what to stop doing; the single biggest risk; the single biggest opportunity being missed; what NOT to build for 90 days; and plain one-sentence answers to Q1, Q2, and Q3. 4. THE NEXT THREE SESSIONS: three work sessions I can start tomorrow, each with an objective, a definition of done, and the tasks split between ME and AI ASSISTANT. 5. COULD NOT VERIFY: everything you could not confirm, and the question that would resolve each item. ANTI-PATTERNS. Any of these makes the audit worthless: an unranked wishlist; advice that ignores my hours; praise for effort that produced no revenue; a finding with no evidence behind it; softening to be agreeable; ending in "more research is needed." Begin with the interview.
Do roadmap item 1 the same week you run this. Momentum is the deliverable; the memo is just its excuse. Re-run the audit quarterly and compare the roadmaps: what moved is your proof.
This is the self-serve version. The audit I run for clients is the same structure with measurements instead of self-reports: I verify the numbers, time the workflows myself, find what a self-audit structurally cannot see, and build the top fix with you. If you want that, bring me the workflow that annoys you most: book the install call.
Twelve copy-paste AI systems for your job and your business, every one shown with the output it actually produced. No code. No hype. v2.0, July 2026.
This pack is twelve AI systems I actually use. Not theory, not tool lists. Each one is a complete, copy-paste prompt that turns a chat AI into a specific machine: an inbox triager, a meeting cleaner, a decision stress-tester. I have spent 15 plus years as an engineer explaining technology to non-technical people, and every workflow here had to pass two tests: could a smart professional with zero technical background run it in the next ten minutes and get real work out of it, and could I put its output on the page next to it. Twelve passed both. The rest did not make the pack.
About the outputs. Under every prompt you will find the output: the exact prompt above it, run in Claude in July 2026 on the sample input named alongside it, trimmed only for length. Nothing is described that is not shown. Run the same prompt on your real input and your output will be better than the sample, because your context is real.
Only have ten minutes? Start with these three: 01 The Inbox Triage Machine, 04 The Minutes-to-Actions Converter, and 11 The Sunday Twenty-Minute Planner. Those three alone are worth hours a week, and they teach you the pattern behind all the others.
The one rule I never break. Nothing confidential goes into a public AI tool. No passwords, no client secrets, no unreleased numbers. For company data, use whatever tool your company has approved. Plenty of workplaces still have no clear AI policy. Be smarter than the policy you do not have.
The pain: You open email to send one thing and lose forty minutes to everyone else's priorities.
// the prompt. copy everything in this box:
You are my chief of staff and you run a ruthless inbox triage. My context: my role is [YOUR ROLE], and my top three priorities this week are [PRIORITY 1], [PRIORITY 2], [PRIORITY 3]. Below I will paste my unread emails as sender, subject, and first lines. Sort every single one into exactly four buckets: 1. REPLY NOW: needs me personally today. Tell me why in one line. 2. QUICK REPLY: draft a one or two sentence reply I can paste as is. 3. DELEGATE OR REDIRECT: name who should own it and draft the one line forward note. 4. ARCHIVE: no action needed. Five words max on why. Rules: when in doubt, archive. Flag anything involving money, deadlines, or my boss with a star. Never invent facts that are not in the emails. Output everything as a single table. Emails: [PASTE YOUR UNREAD LIST]
The output. This prompt, run in Claude, July 2026. Sample input: an operations manager, priorities "close the Q3 vendor renewal, ship the onboarding revamp, hire one coordinator," and eight unread emails pasted as sender, subject, first line. Trimmed to fit the page:
| # | From | Subject | Bucket | Action | |---|-------------|--------------------------|--------------|-----------------------------------------| | 1 | CFO | * Renewal pricing call | REPLY NOW | Money + your top priority. Book it today| | 2 | Vendor rep | * Contract expires Fri | REPLY NOW | Deadline. Confirm terms before the call | | 3 | Recruiter | Coordinator shortlist | QUICK REPLY | "Shortlist looks right. Book the top | | | | | | two for Thursday." | | 4 | Teammate | Onboarding copy draft | DELEGATE | Own: Priya. "Priya, this is yours | | | | | | through Friday. Loop me on blockers." | | 5 | IT | Password expiry | QUICK REPLY | "Done, thanks." | | 6 | Newsletter | 5 trends for 2027 | ARCHIVE | No action, no deadline | | 7 | Event team | Speaker invite, October | ARCHIVE | Decide next week, not this one | | 8 | Old client | "Quick question" | QUICK REPLY | "Happy to help. What is the question?" |
Pro move: Run this once at 11am and once at 4pm instead of living in your inbox all day. The buckets hold.
The pain: One awkward email sits in your inbox for two days because you do not know which version of you should answer it.
// the prompt. copy everything in this box:
Here is an email thread I need to answer: [PASTE THREAD] My relationship to the sender: [BOSS / PEER / CLIENT / VENDOR]. What I actually want out of this: [YOUR REAL GOAL]. Write three different replies: 1. AGREE AND ADVANCE: say yes and move it forward with a concrete next step. 2. PUSH BACK: disagree with warmth, give one clear reason, and offer an alternative. 3. GRACEFUL STALL: buy me [AMOUNT OF TIME] without sounding indecisive and without using the word maybe. Each reply under 120 words. Tone: [DIRECT / WARM / FORMAL]. No corporate filler, and never open with I hope this finds you well.
The output. This prompt, run in Claude, July 2026. Sample input: a vendor announcing a 20 percent renewal increase and asking to "hop on a call this week"; relationship VENDOR; goal "renew, but not at that number"; one week of stall; tone DIRECT. Trimmed:
1. AGREE AND ADVANCE "Thanks for the heads-up. Before we book a call, send the renewal terms in writing with the increase broken out by line item. If the numbers hold up I will sign this week; if not, the call is where we fix them. Friday 2pm or Monday 10am work." 2. PUSH BACK "I want to renew, and 20 percent is not a number I can take to my CFO. Our volume is up 30 percent year over year, which usually earns a better rate, not a worse one. Send me a revised figure that reflects that and I will turn it around in 48 hours." 3. GRACEFUL STALL "Renewal is on my list this week. I am mid-quarter-close through next Wednesday, so I will come back to you Thursday with either a signature or questions. If the deadline moves before then, tell me now."
Pro move: You will almost always send draft two with one edit. The other two exist to show you the choice you are making.
The pain: Saying no, delivering bad news, or asking for what you deserve, and rewriting it eleven times.
// the prompt. copy everything in this box:
You are an executive communication coach. Help me write a difficult message. The situation: [DESCRIBE IN 2 TO 3 SENTENCES] Who it is going to and our relationship: [DESCRIBE] The hard thing I need to say: [SAY IT BLUNTLY HERE, THIS PART IS JUST FOR YOU] What I am afraid will happen: [YOUR FEAR] What I want the relationship to look like after: [YOUR GOAL] Write the message in three parts: acknowledge their side honestly, state the hard thing in one clear sentence with no hedging, then offer a concrete path forward. Under 150 words. Then give me a shorter one paragraph version, and a list of three phrases I should NOT use because they would soften the message into mush.
The output. This prompt, run in Claude, July 2026. Sample input: telling a long-standing client their launch slips two weeks because their own asset approvals ran late; fear of losing the account; goal "keep the relationship, reset the dates honestly." Trimmed:
THE MESSAGE "You have been clear from day one that the September date matters, and your team has put real work into this launch. I want to be straight with you: we will not hit September 12, and the honest new date is September 26. The asset approvals we needed by August 1 closed on August 19, and I chose not to compress testing to swallow the gap, because you would pay for that later. Here is the reset: revised timeline attached today, approvals now have a 48-hour turnaround window on both sides, and I will send a two-line status every Friday so nothing surprises either of us again." SHORT VERSION "September 12 is no longer honest, and I will not pretend otherwise. New date: September 26, driven by the August approval slip. Revised plan attached; Friday updates from here on." DO NOT USE 1. "We may experience some delays" (it already happened; own it) 2. "Due to circumstances beyond our control" (blame-shifting mush) 3. "Hopefully we can" (hope is not a plan)
Pro move: The blunt line you type in the brackets is for the machine only. Getting it out of your head is half the work.
The pain: The meeting ended an hour ago and nobody can say who is doing what.
// the prompt. copy everything in this box:
Below are my raw meeting notes or a transcript. Turn them into three things: 1. DECISIONS MADE: a numbered list of only the actual decisions. If something was discussed but not decided, it does not belong here. 2. ACTION ITEMS: a table with owner, task, and deadline. If the owner or deadline was never stated, write [UNCLEAR] instead of guessing. Never invent commitments. 3. OPEN QUESTIONS: what was raised but left unresolved. Then draft a short follow-up email to all attendees summarizing the above, in a neutral professional tone, ready to paste and send. Notes: [PASTE NOTES OR TRANSCRIPT]
The output. This prompt, run in Claude, July 2026. Sample input: one page of messy notes from a product launch sync (five attendees, three topics, two arguments). Trimmed; the follow-up email is produced too and omitted here for space:
DECISIONS MADE 1. Launch moves to October 3. Nobody re-litigates this before September 15. 2. Pricing page ships with three tiers, not four; the enterprise tier moves to a "talk to us" link. ACTION ITEMS | Owner | Task | Deadline | |-----------|-----------------------------------------|-------------| | Dana | Revised launch plan to the channel | Fri, Aug 8 | | Marcus | Three-tier pricing page copy | Wed, Aug 13 | | [UNCLEAR] | Update the sales deck for the new tiers | [UNCLEAR] | OPEN QUESTIONS 1. Does the annual-plan discount survive the tier change? Raised twice, decided never. 2. Who owns the launch-day support rota? Two names floated, no owner.
Pro move: Send the follow-up within an hour of the meeting. The person who writes the summary quietly controls what was decided.
The pain: You send five paragraphs of context. They read two lines and reply with a question you already answered.
// the prompt. copy everything in this box:
Condense everything below into a one page executive brief with exactly four sections: 1. SITUATION: three sentences maximum, plain English. 2. OPTIONS: a small table with option, cost, risk, and speed. 3. RECOMMENDATION: pick one, with two sentences on why. 4. WHAT I NEED FROM YOU: one line, with a date. Rules: hard one page limit. No jargon, no throat clearing, no just-checking-in energy. Write it so a smart person with 90 seconds makes the right call. Material: [PASTE THE THREAD, DOC, OR YOUR NOTES]
The output. This prompt, run in Claude, July 2026. Sample input: a rambling six-message thread about whether to keep patching a legacy internal tool or replace it. Trimmed:
SITUATION The scheduling tool breaks roughly weekly and two people spend most of a day each week patching it. The vendor ended support in March. Every month of delay costs about one person-week of engineering time. OPTIONS | Option | Cost | Risk | Speed | |-----------------|----------------|-------------------------|----------| | Keep patching | 1 wk/mo, rising| Fails during peak season| n/a | | Replace, buy | License fee | Migration surprises | 6 weeks | | Replace, build | 2 eng-months | Scope creep, key-person | 10+ weeks| RECOMMENDATION Buy. The patching cost passes the license cost within a quarter, and peak season starts in nine weeks, which only the buy option beats. WHAT I NEED FROM YOU Approval to start the two-week migration pilot by Friday, August 15.
Pro move: Section four is the whole trick. Executives move when the ask has a deadline and fits on one line.
The pain: Explaining the task properly takes longer than doing it, so you do it. Again. Forever.
// the prompt. copy everything in this box:
I need to hand off a task that currently lives only in my head. I will ramble everything I know below. Turn it into a delegation packet: 1. THE OUTCOME: one sentence on what done looks like, measurable if possible. 2. THE STEPS: numbered, specific, in order, with exact tools or screens named. 3. DECISION RIGHTS: what they decide on their own versus when they must check with me. 4. QUALITY BAR: a short checklist they run before sending it back to me. 5. COMMON MISTAKES: the two or three ways this usually goes wrong. Write it so a smart new person could execute without asking me a single question. My ramble: [EXPLAIN THE TASK LIKE YOU ARE TALKING TO A COLLEAGUE, TYPOS FINE]
The output. This prompt, run in Claude, July 2026. Sample input: a spoken-style ramble about assembling the Monday client status email from three sources. Trimmed to the sections a reader needs to judge it:
THE OUTCOME The Monday client email is in the client's inbox by 10am, accurate against all three sources, and requires zero corrections during the week. THE STEPS (of 7, trimmed) 1. Open the project tracker board; filter to "Done last 7 days." 2. Copy the three most client-visible items into the WINS section of the template (template link in the shared drive, "Client Weekly" folder). 3. Check the support queue for anything tagged [client name] and open longer than 5 days; those go under RISKS, each with a next step. ... DECISION RIGHTS Decide alone: wording, ordering, which wins lead. Check with me first: anything touching money, dates we committed to, or an apology. QUALITY BAR [ ] Every claim traces to one of the three sources [ ] No item copied from last week unchanged [ ] A named next step under every risk COMMON MISTAKES 1. Padding WINS with internal work the client cannot see. 2. Listing a risk without an owner and a next step, which reads as alarm.
Pro move: Record yourself explaining it as a voice memo and paste the transcript. Talking is faster than writing.
The pain: You have already decided. You just want permission. That is exactly when expensive mistakes happen.
// the prompt. copy everything in this box:
I am about to make a decision and I want you to attack it before I commit. The decision: [DESCRIBE IT]. Why I am leaning yes: [YOUR REASONING]. What it costs if I am wrong: [MONEY, TIME, REPUTATION]. Do this in order: 1. PREMORTEM: it is one year later and this failed. Write the five most likely reasons why. 2. HIDDEN ASSUMPTIONS: list what I am assuming without evidence. 3. THE CHEAP TEST: the smallest, fastest way to test this before fully committing. 4. THE VERDICT FRAME: complete these two sentences. Go ahead if... Walk away if... Rules: do not be agreeable. Do not soften. Argue against me like money is on the line, because it is.
The output. This prompt, run in Claude, July 2026. Sample input: "hire a full-time coordinator now" versus waiting; leaning yes because the founder is drowning; cost if wrong: three months of runway and a painful exit conversation. Trimmed:
PREMORTEM (5, trimmed to 3) 1. You hired for the work you HAVE, and the work changed in a quarter; now you have a salary and no role. 2. You skipped writing the role down, so they became your assistant instead of the owner of anything, and you are still the bottleneck. 3. Revenue stayed flat; the salary compressed runway and every decision after month two was made scared. HIDDEN ASSUMPTIONS You assume the drowning is a volume problem, not a process problem. You assume you know which tasks transfer; you have never listed them. You assume full-time is the only shape that fixes this. THE CHEAP TEST Write the delegation packet for the top three tasks first, then buy 15 hours a week of contractor time for one month against those packets. If the hours stick and the packets survive contact, hire. THE VERDICT FRAME Go ahead if the contractor month clears 12 of 15 hours a week and at least one task leaves your head entirely. Walk away if you cannot write the three packets, because then the problem is process, and a hire would inherit the chaos instead of ending it.
Pro move: If the premortem surfaces a failure mode you had not considered, run the cheap test first. Every time.
The pain: Three tools, four vendors, two offers, and a gut feeling you cannot defend in front of anyone.
// the prompt. copy everything in this box:
Help me choose between: [OPTION A], [OPTION B], [OPTION C]. Context: [WHAT THIS IS FOR AND WHO USES IT]. What matters most to me, in rough order: [E.G. PRICE, EASE, SPEED, SUPPORT]. 1. Propose the 5 to 7 criteria that should drive this decision, including any I forgot, and weight them. 2. Build a comparison table scoring each option against each criterion, with one line of reasoning per cell. If you do not know something about an option, write UNKNOWN. Do not fill gaps with guesses. 3. Give a verdict by scenario: choose A if... choose B if... choose C if... 4. Name the gotcha for each option, the thing people discover after committing. 5. Give me three questions to ask before signing anything.
The output. This prompt, run in Claude, July 2026. Sample input: choosing a project tool for a 6-person services team that lives in email; priorities ease, then adoption, then reporting. Trimmed to the verdict layer, which is the part you keep:
CRITERIA (weighted): ease of daily use 30, will-the-team-actually-use-it 25, client-facing reporting 20, migration effort 15, admin overhead 10. VERDICT BY SCENARIO Choose A if the team's real problem is forgetting tasks, not planning them; it wins on daily ease and loses on reporting. Choose B if a client-facing status view matters weekly; it carries the best reporting and the worst first-month adoption curve. Choose C only if someone on the team volunteers to own setup; flexible, and flexibility unowned becomes mess. THE GOTCHAS A: fine at 6 people, strains at 15, and migrations out of it are manual. B: the reporting you buy it for lives in the higher tier. UNKNOWN whether the entry tier exports client views; ask before signing. C: every team builds a different system in it; yours is only as good as its most disciplined user. ASK BEFORE SIGNING: What does export look like the day we leave? Which of these features are on the tier we would actually buy? What does support cost after the first year?
Pro move: The by-scenario verdict is the gold. There is rarely one right answer. There is a right answer for your situation.
The pain: Everyone in the meeting nods at words like RAG and orchestration. Nobody in the meeting could explain them.
// the prompt. copy everything in this box:
Explain [TERM, ACRONYM, OR PASTE A DENSE PARAGRAPH] to me at three levels: LEVEL 1: One sentence a smart 12 year old would understand. LEVEL 2: One paragraph with a real world analogy from everyday life or business. LEVEL 3: The full picture. How it actually works, when it matters, and its limits. Define every acronym on first use. Then add two things: 1. WHAT PEOPLE GET WRONG: the most common misunderstanding about this. 2. THE MEETING TEST: one sentence I could say in a work meeting that uses this concept correctly. Rules: not dumbed down, actually clear. If part of the concept is genuinely complex, say which part and why.
The output. This prompt, run in Claude, July 2026, on the term RAG. Level 3 trimmed:
LEVEL 1: RAG means the AI looks things up in your documents before it answers, instead of answering from memory. LEVEL 2: It is an open-book exam instead of a closed-book one. A model answering from memory is a smart student recalling what it studied months ago; sometimes wrong, always confident. RAG hands that student your binder first: the answer comes from the pages it just read, and it can show you which pages. LEVEL 3 (trimmed): Retrieval Augmented Generation. Your documents are split into chunks and indexed; your question retrieves the most relevant chunks; the model answers using them. It matters when the knowledge is yours, changes often, or must be traceable. Its limit: retrieval quality caps answer quality; if the right chunk is not found, the model reasons beautifully from the wrong pages. WHAT PEOPLE GET WRONG: RAG does not "train the model on your data." Nothing is learned or retained; it is lookup at question time. THE MEETING TEST: "Before we budget for fine-tuning, would RAG over our own docs get us there, since our content changes monthly?"
Pro move: Run this on the next acronym someone throws at you, live, under the table. Then ask a Level 3 question out loud.
The pain: Every AI draft sounds like the same beige robot, and everyone can tell.
// the prompt. copy everything in this box:
I am going to paste two or three samples of my real writing. Study them and build my WRITING VOICE CARD: 1. TONE: five words that describe how I write. 2. RHYTHM: my typical sentence length and structure habits. 3. SIGNATURE MOVES: phrases, openings, or patterns I use often. 4. NEVER: things I clearly avoid. Words, punctuation, formality levels. 5. INSTRUCTION BLOCK: a reusable paragraph that starts with WRITE AS ME: which I can paste at the top of any future prompt so the output matches my voice. Samples: [PASTE 2 TO 3 REAL EMAILS, POSTS, OR DOCS YOU WROTE, AT LEAST 300 WORDS TOTAL]
The output. This prompt, run in Claude, July 2026. Sample input: three short emails from a founder who writes in short declaratives and hates exclamation marks. Trimmed:
TONE: Direct, warm, unhurried, concrete, lightly dry. RHYTHM: Short sentences, median around ten words. One-sentence paragraphs for emphasis. Lists of exactly three. SIGNATURE MOVES: Opens with the point, never a greeting ramp. "Here is where I landed." Ends with a single concrete next step and a date. NEVER: Exclamation marks. "Just checking in." Hedges like "I think maybe." Corporate verbs: leverage, circle back, align. INSTRUCTION BLOCK WRITE AS ME: Open with the point in the first sentence; no greeting warm-up. Short declarative sentences, about ten words. One idea per paragraph; single-sentence paragraphs allowed for emphasis. Lists come in threes. Warm but never effusive; zero exclamation marks. Close with one concrete next step and a date. Ban: just checking in, leverage, circle back, I think maybe.
Pro move: Save the instruction block in your notes app. It upgrades every other workflow in this pack. This is the one prompt that improves all the others.
The pain: Monday decides your week before you do.
// the prompt. copy everything in this box:
Plan my week with me. First, here is everything on my mind, unsorted: [BRAIN DUMP EVERYTHING. TASKS, WORRIES, IDEAS, OBLIGATIONS. MESSY IS THE POINT.] My fixed commitments this week: [MEETINGS, DEADLINES, PERSONAL]. My realistic deep work hours available: [NUMBER]. Now: 1. Sort everything into MUST (moves something important), SHOULD (real but survivable), COULD (honestly optional). 2. Pick the three items that, if done, make this week a win. Justify each in one line. 3. Lay out a day by day plan that fits my actual available hours, with 20 percent buffer built in. Do not schedule a fantasy week. 4. PARK LIST: everything that did not make the cut, written down so my brain can let go of it. 5. Name one thing on my list I should say no to entirely, and give me the exact sentence to decline it.
The output. This prompt, run in Claude, July 2026. Sample input: a 19-item brain dump from a consultant with 9 real deep-work hours and two fixed client days. Trimmed:
SORTED: 5 MUST, 6 SHOULD, 8 COULD. THE THREE WINS 1. Send the proposal (it unblocks revenue; nothing else on the list does). 2. Prep Thursday's client workshop (fixed date, reputation-priced). 3. Write the overdue invoice and the follow-up (money already earned). THE WEEK (trimmed): Mon 2h proposal draft, 1h invoice + follow-up. Tue client day, untouched. Wed 2h proposal finish + send, 1h workshop prep. Thu client day + workshop. Fri 2h buffer, then the park list only if the three wins are done. PARK LIST (8 items): website refresh, podcast idea, tool migration, inbox zero, the course outline, new CRM research, headshot, LinkedIn banner. SAY NO TO: the "quick advisory chat" with no agenda. The sentence: "I keep a tight calendar mid-quarter. Send me the two questions in writing and I will answer this week; if it needs more than that, my first open slot is in three weeks."
Pro move: The park list is the secret. Stress is mostly unparked items circling your head at 2am.
The pain: The business runs on things only you know how to do, which means you can never be sick, offline, or on vacation.
// the prompt. copy everything in this box:
I am going to explain how I do a task. It will be rambling and out of order, like I am training someone over my shoulder. Turn it into a clean SOP: 1. PURPOSE: what this task is for, one line. 2. WHEN TO RUN IT: the trigger or the schedule. 3. STEPS: numbered, exact, with specific tool names, buttons, and screens. 4. IF THIS THEN THAT: the decision points, written as simple rules. 5. DONE LOOKS LIKE: a checklist to verify the task was completed correctly. 6. COMMON MISTAKES: what usually goes wrong for a beginner. Write it so a capable new hire executes it correctly on day one with zero questions. My ramble: [PASTE A VOICE MEMO TRANSCRIPT OR TYPE IT MESSY]
The output. This prompt, run in Claude, July 2026. Sample input: a rambled explanation of onboarding a new client (kickoff email, folder setup, first invoice, calendar). Trimmed:
PURPOSE: Every new client starts with the same clean first week, without
the founder touching it.
WHEN TO RUN IT: The day a signed agreement lands.
STEPS (of 8, trimmed)
1. Duplicate the "Client Template" folder in the shared drive; rename to
the client name.
2. Send the kickoff email from the saved template ("Kickoff v3" in
drafts); attach the intake form; set the reply-by date to +3 business
days.
3. Create the invoice from the template; terms as signed; send from the
billing tool, not email.
...
IF THIS THEN THAT
If the intake form is not back by the reply-by date, send the one-line
nudge template. If payment terms differ from standard, stop and confirm
with the owner before sending the invoice.
DONE LOOKS LIKE
[ ] Folder exists and matches the template structure
[ ] Kickoff sent, intake date in the calendar
[ ] Invoice sent from the billing tool, copy in the folder
COMMON MISTAKES: Sending the invoice from email (it skips the payment
tracking), and personalizing the kickoff so heavily the reply-by date
falls out.Pro move: One SOP per week. In three months the business runs on a manual instead of your memory.
Three things, then get back to work.
Everything in this pack was built and tested on real work. Mine. And every output you just read came from the prompt printed above it. Shown, not claimed.
Every term from the decoder video, plus one sentence you can say in your next meeting.
A program that learned language by reading a huge slice of the internet. Its whole job is predicting the next word well enough that it looks like thinking.
"The model is just predicting text, so how we word things changes what we get back."
The chunk of text a model reads at once. Not a word, not a letter, roughly four characters. Pricing, speed and limits are all counted in tokens.
"Long documents cost more because everything is billed in tokens."
How much the model can hold in its head at one time, measured in tokens. Feed it more than the window holds and the oldest content falls out the top.
"That document won't fit in the context window, so we'll need to break it into chunks."
The question, the instructions, the examples, the tone. Prompt engineering just means wording the ask carefully so the prediction lands where you want it.
"Before we build anything custom, let's see how far a better prompt gets us."
Training is school. Inference is the job. Every message you send triggers inference, and inference is what companies pay for, every time, forever.
"Inference costs scale with usage. The training already happened."
Training builds the brain from scratch on a huge slice of the internet: millions of dollars, months of compute, almost nobody does it. Fine-tuning is a short apprenticeship on top of an already trained model using your own examples.
"We don't need to train a model. At most we fine-tune one."
When the model states something completely false with total confidence. It is not a glitch. The model predicts plausible text; it does not check truth. Confident and wrong is how it works.
"What's our guard against hallucinations before this touches a customer?"
Instead of trusting the model's memory, you fetch your real documents first and hand them over with the question. An open-book exam instead of a memory test.
"Hallucinations drop when we RAG it against our own documents."
An embedding turns meaning into coordinates so similar ideas sit near each other on a map. A vector database is that map, built to find nearest neighbours fast. It is the engine under RAG.
"It searches by meaning instead of keywords. That's what the embeddings are for."
A model given tools, a goal, and permission to take steps on its own. It reads, decides, acts, checks its work, and repeats until the job is done. A chatbot answers you; an agent does things.
"Is this a chat feature or an agent, meaning does it actually take actions?"
A standard plug that lets a model connect to tools and data the same way every time, instead of every company wiring custom connections. Think of it as USB-C for AI.
"Does it support MCP, or are we paying for a custom integration?"
The checks wrapped around a model so it can't say or do what it shouldn't: content filters, spending limits, human approval steps, audit logs. When a company says the AI is safe, this is the claim.
"What guardrails are in place before this goes live?"
An open model you download and run on your own machines, so your data never leaves the building. A closed model you rent through an API, and the closed frontier models are usually strongest. That is the whole trade.
"Do we actually need frontier quality here, or is an open model on our own hardware the safer call?"
The hypothetical model as capable as a person at everything. It does not exist, it is not on a product roadmap, and when someone drops it in a business meeting they are usually selling something.
"Let's solve this quarter's problem. AGI isn't going to ship it for us."
// want the systems, not just the words? every term's full write-up lives in the decoded index_
Run the audit, steal the workflows, say the sentences out loud. And when you want the systems built inside your business instead of beside it, that is a conversation.