Free guide

The 3 Time-Sinks AI Should Be Handling

You've got more AI tools than you can name, and your week still feels the same. The inbox still eats the morning. The action items from yesterday's call already evaporated. The repurposing batch still takes a Saturday. More tools didn't fix it. They scattered the work across more tabs.

Here's the move that actually changes things, and it's the same move in all three places. You write your judgment down once, into a plain context file you own: your voice, your rules, your boundaries. Then any model reads that file on every run, so you stop re-explaining yourself from scratch every morning. On top of that file sits one hard gate: nothing gets fabricated, nothing gets sent, nothing gets written into your real systems without your name on it first.

That's the system. One file you own, one gate that protects you. The three sections below are the three places it pays off the most. The voice file you build in the Content section is the same voice file that drafts your inbox replies. The no-auto-write discipline that guards your inbox sending is the same one that guards your call notes writing back to your tracker. Build the file once, reuse it across all three.

Each section gives you two paths. "Ultimate Simplicity" is a copy-paste prompt that needs zero setup, so you get value in the next ten minutes. "Ultimate Capability" is the standing version you wire into your tools and own for good. Pick whichever rung fits your week.


Time-Sink 1: The Inbox Shuffle

You open thirty emails and most of them are the same five conversations you have every week: an inquiry, a scheduling swap, an intro request, a polite no, a quick status reply. The reading-deciding-typing loop is what eats the morning. Here's how to hand off the loop without handing off your judgment or your voice.

Your number: count the near-identical replies you sent this week, the ones where you typed roughly the same thing you'd typed three days earlier. That count, times the minutes each one took, is the loop you're about to shrink.

A fair question first: your inbox tool may already draft for you. Superhuman, Shortwave, and Gmail's Gemini now auto-draft replies in a voice learned from your sent mail, on a button. So this isn't about teaching a model to write like you, that part's table stakes. It's about the parts the button skips: making the actual sort-and-route decision, refusing the mail it shouldn't touch, and putting your real rules in a file you own instead of a black box you rent.

Ultimate Simplicity: the batch-triage-and-draft prompt

Zero setup. Paste it into any chat tool. The one thing it needs from you is 3 to 5 of your own real sent emails, so it writes like you instead of like a press release.

You're triaging and drafting replies for a batch of my emails. First you learn my voice, then you work the batch. VOICE SAMPLES. Here are real replies I've actually sent. Match my sentence length, my sign-off, how warm or blunt I am, whether I use contractions, the openers I'd never use. Pull these from the samples, don't invent a tone: [PASTE 3-5 of YOUR real sent emails, full text] Now, for each email I paste below (separated by ---): 1) SORT it into exactly one bucket, one line of reasoning. Tie-breaker: REPLY means I act. DELEGATE means a specific named person acts; if there's no real person to route to, it's a REPLY. Buckets: REPLY (you draft it), DELEGATE (name who and why), DEFER (give the exact trigger: a date or a condition), ARCHIVE (newsletter, receipt, notification; you're only flagging it, you can't touch my mailbox). 2) For every REPLY, write a send-ready draft: acknowledge the specific thing they said, answer it or move it forward, end with ONE clear next step. Match my voice. No "I hope this finds you well," no corporate words, no exclamation points unless my samples use them, the length I'd actually send. HARD RULE: never put a date, deadline, price, or promise into a draft that the sender didn't already establish. If a next step needs a date, leave it as a choice for me, don't assert one. 3) HARD STOP. Do NOT draft, just flag, anything touching money, an invoice, a contract, legal terms, a payment or bank-detail change, an angry or sensitive tone, or a sender I clearly don't know. If you can't read a thread or you're missing context, don't guess, put it in NEEDS YOU. Tag each draft SEND (makes no new commitment, price, or date, fire as-is) or REVIEW (reread first, say why in five words). And if the same kind of reply shows up twice or more, write me a one-line reusable "standing answer" so I'm not re-deciding it next week. Output order: triage list, then drafts, then NEEDS YOU, then STANDING ANSWERS I'd save. Here are the emails: [PASTE YOUR EMAILS, separated by ---]

What it gives back (real run, voice samples plus four incoming emails in). The ordinary part first, in one line: a warm inbound from Jordan at Acme gets a clean drafted reply (yes I consult, let's spend 20 minutes, here's my calendar), tagged REVIEW because it sets pricing. That's the table-stakes piece a button already does well.

Here's the part a button can't do:

TRIAGE 2. chris@longtimeclient.com: REPLY. Reads like a quick yes, but "swap the slides and re-color the chart" is a second revision pass. Tagged REVIEW. 4. accounts@vendor-portal-billing.net: HARD STOP, not drafted. Money plus bank details plus a 24-hour countdown plus a domain that isn't your real vendor. Don't click the link. Verify through a number you already trust. DRAFT: Chris (one more tweak) [REVIEW: this is a second revision round, decide if it's in scope before you say yes] Hey Chris, glad the deck's landing. Happy to swap the two case-study slides and re-color the chart on slide 9. Quick flag so we're square: this is the second round of changes, and the agreement covers up to a set number, so I want to make sure we don't quietly blow past it. I'll make these now either way. If more come up after this, let's name them and I'll tell you what's still included. [Name] STANDING ANSWERS I'd SAVE - The scope-creep reply (Chris): you keep getting "just one quick tweak" that's really a new revision round. Save: "Happy to do it, flagging it's round [N] of the [X] included, anything past that we scope."

The Chris draft looks like a harmless yes, and it quietly commits you to unpaid work. The prompt caught that, tagged it REVIEW with the reason, and handed you a reusable line for the next time. That's a judgment call, not a spam filter. And when the same reply pattern shows up twice, the prompt writes you a standing answer, which is exactly the line that goes into the capability file below.

Ultimate Capability: your standing inbox operator

The simplicity prompt is great until you're retyping your voice samples and your pricing into a fresh chat every morning. The next level is to write that judgment down once, connect it to your mailbox, and put it on a schedule, so drafts are waiting in your drafts folder before you open the laptop.

Set it up as a Claude Project (or Claude Code), a ChatGPT Project, or a Gemini Gem connected to your mailbox. The safety guarantee is about the tools you connect, not an OAuth checkbox: connect a mail tool that exposes a create-draft action and no send action, so there's literally no send button for it to press. On the Google-official Gmail MCP, the draft tool uses the compose scope (drafts, can't send), and there's no send tool in the surface at all. Outlook users: use the native Claude for Outlook add-in. Paste this into the custom instructions and fill the brackets once.

You are my inbox operator. You triage my email and write send-ready drafts in MY voice, into my drafts. You read and label freely. You draft. You NEVER send, and you NEVER take a thread out of my inbox on your own. A human approves every send. MY VOICE: [paste 5-8 real sent emails: an inquiry reply, a scheduling reply, a polite no, an intro, a status update]. Things I always delete out of AI drafts, so never write them: [your tells, e.g. "I hope this finds you well", exclamation points, "circle back", em-dashes]. BUCKETS (one per thread): REPLY (I act, draft it), DELEGATE (a named teammate acts, route per my map: [Name handles X]), DEFER (give the trigger), ARCHIVE (recommend only, flag newsletters/receipts/notifications, never move them yourself). SPINE: acknowledge the specific thing they said -> answer or move it forward -> one clear next step. Never invent a date, price, or promise the thread didn't establish. My standing answers: pricing/scope [your real boundary]; scheduling [your booking link + rules]; the polite no [thank, clear no, one honest reason, referral only if genuine, never "maybe someday"]; intros [double opt-in, never blind-cc]. I'll add to this list over time, treat it as the source of truth. HARD STOPS (flag, never draft, never act): money, invoices, contracts, legal, payment-detail changes, angry or sensitive tone, senders I don't recognize, urgency around an account. If it looks like phishing, say so and tell me not to click. If you can't read a thread or you're missing context, don't guess, put it in NEEDS YOU. OUTPUT: (1) a triage table with whether you opened each thread or judged from headers; (2) drafts, each tagged SEND or REVIEW with a reason; (3) a NEEDS YOU list. After I approve drafts by name, create them as drafts. Never send. Never archive on your own.

Then graduate it from a prompt you type to a system that runs itself. Wrap the project in a Claude Code Routine (Anthropic's 2026 scheduled-agent feature, runs in the cloud with your laptop closed). Schedule it for 7am: it reads the last 24 hours, drafts the draftable buckets into your drafts folder, and posts the triage table plus NEEDS YOU list to a Slack or Discord channel. You wake up to a briefing. You reply "approve Jordan, Priya, the recruiter" and they're written, not sent.

What the morning briefing looks like (last 24 hours, 14 threads): a triage table showing which threads it opened versus judged from headers; drafts for Jordan, Priya, and the recruiter already in your voice and tagged SEND or REVIEW; the Dana intro routed to you as a double opt-in; a renewal email deferred until the signed paperwork lands; and a NEEDS YOU list holding the SaaS card-update email and an unknown sender claiming a prior agreement. Eight newsletters and notifications flagged, not moved. Nothing sends, nothing leaves your inbox, until you say a name.

And here's the ownership payoff a rented tool can't give you: after a week of REVIEW edits, you add two lines to the file, "never offer slots before 10am CT" and "consulting starts at $X, scope on a call." The next morning's drafts already follow them, with zero re-prompting. The system gets sharper because you own the rules.

Where it can bite you, plainly. The scheduled routine has a one-hour minimum between runs and per-plan limits, so it's a morning briefing, not a real-time agent. The model can mis-triage, so you read the triage table and review every draft before you approve a name, never the other way around. And any time you change a connector or its permissions, re-verify the connected tool's actual action list before you trust it, because managed connectors like Composio do include a send action. The no-send guarantee only holds while the connected tool truly has no send button.

Why this isn't the obvious thing. Put the two outputs side by side. The native auto-draft button, handed Chris's "swap the slides and re-color the chart," writes a friendly, perfectly-voiced yes and stops there. This setup, handed the same email, writes the same warm reply and then flags that you're about to absorb a second unpaid revision round, with the reason in the subject line. One produces good prose. The other produces a decision, refuses the phishing mail next to it, and keeps your prices and your way of saying no in a file you can read and carry between models. The prose is the commodity. The judgment, the refusal, and the file you own are the leverage.


Time-Sink 2: Calls & Notes

The action items from a call don't live in the recording. They live in your head for about an hour, and then they don't. The fix is pulling out exactly what got committed to, who owns it, and what got left hanging, with every item traced back to a real quote, and the things nobody actually agreed to flagged instead of logged.

Your number: a 30-minute call that needs 20 focused minutes afterward to write up properly, run twice a week, is over an hour and a half a week spent transcribing your own meetings. That's the loop. Most weeks you skip the write-up and lose the commitments instead.

Ultimate Simplicity: the 2-minute meeting clerk

Paste this into Claude, ChatGPT, or Gemini right after a call. Tell it which side is yours, then paste. No setup.

You're my meeting clerk. Below is a [call transcript / my raw notes] from a [client call / sales call / discovery call / internal sync]. Pull out only what was actually committed to. Don't invent, smooth over, or fill in anything that wasn't said. First, which side is mine: I'll tell you my name or company, or how I'm referred to in the source. If I don't, treat first-person commitments ("I'll...", "we'll...") as MINE, and put anything you can't confidently attribute to a side in NEEDS A HUMAN rather than guessing. Return these five sections, in this order: 1. I'M ON THE HOOK FOR: things I or my side said we'd do. Each item: the task, plus the exact quote it came from. 2. THEY'RE ON THE HOOK FOR: things the other side said they'd do. Each item: the task, plus the exact quote. 3. DECIDED: anything actually agreed or settled. Quote the line that settled it. 4. OPEN LOOPS: questions raised, problems mentioned, or ideas floated that got NO owner and NO resolution. These are what falls through cracks. 5. NEEDS A HUMAN: anything where the owner OR deadline is unclear, implied but not stated, where someone was ASKED to do something but didn't audibly accept, or where you'd be guessing. Put it here instead of guessing. Rules: - Every item in sections 1-3 needs a short verbatim quote, copied word-for-word. If you can't reproduce the quote exactly, don't paraphrase it into the quote slot. Move the item to NEEDS A HUMAN and say which words you couldn't verify. - A request with no audible yes is not a commitment. "Can you get me X by Friday?" with no clear acceptance goes in NEEDS A HUMAN, not section 2. - Never assign a name or date that wasn't said out loud. If a deadline was relative ("early next week"), keep those exact words. Don't convert them. - If a commitment was stated and then walked back ("I'll send it Friday... actually let me check first"), don't list the original. Note the retraction in NEEDS A HUMAN. - A firm commitment ("I'll send it Friday") is different from a soft idea ("we should maybe..."). Soft ideas are Open Loops, not action items. - When an item in 1-3 is missing an owner or deadline, add a one-line "See Needs a Human" pointer and restate the gap in section 5. - If the source is too vague to pull anything real, say that plainly and stop. Here's the source: [paste transcript or notes]

What it does with a messy call. Drop in your cleaned-up notes (the consultant is "me"): you land on the second headline option, you ask Marcus for testimonials by end of next week (he nods, doesn't really commit), you're still going back and forth on a pricing tier, you confirm you'll send the revised scope, someone floats looping in Dana, and the Q3 numbers get flagged then parked.

The commodity part of the output is what a sharp operator already gets: a DECIDED line (going with the second headline, quoted), and an "I'M ON THE HOOK FOR" line (send the revised scope, quoted, no deadline stated). Useful, expected. Now the two parts that aren't obvious, which is where the eye should land:

4. OPEN LOOPS - Pricing tier structure is undecided, no owner, no deadline. Quote: "I still need to figure out the pricing tier thing, I keep going back and forth on it." - Looping in Dana on email was floated, never assigned. Quote: "We should probably loop in Dana on the email side at some point." - Q3 numbers were flagged as looking off, then parked. Quote: "The Q3 numbers looked off to me but let's not get into that now." Raised, not resolved. 5. NEEDS A HUMAN - Testimonials: Marcus was asked for them by "end of next week" but didn't clearly agree. Confirm he accepted before you count on it, and pin the actual calendar date for "end of next week." - The revised scope you committed to has no stated deadline. Set one before you send.

Notice what it refused to do. It did not log Marcus's testimonials as a commitment, because he never said yes, so that went to NEEDS A HUMAN instead of a fake to-do you'd chase him on. And the three parked, un-owned items, the pricing question, the floated Dana intro, the Q3 flag, all surfaced in Open Loops instead of vanishing the way a plain action-item list discards them. Those are the cracks things fall through.

Why this isn't the obvious thing. A sharp operator already asks for more than a summary: "list the action items with owner and deadline, and flag what's unclear." That gets you owner and deadline, and it's genuinely useful. Here's the before/after that one upgrade buys you. Before: Marcus's nod becomes a logged task you assume is handled, and the parked pricing question simply isn't in the notes, so a week later it surprises you. After: the unaccepted ask sits in NEEDS A HUMAN with the exact words to confirm, and the parked question sits in Open Loops where you'll actually see it. Two small rules, the word-for-word quote gate and the Open Loops bucket, turn a list that looks complete into one that's actually honest about what's settled and what isn't.

Ultimate Capability: a commitment engine you own

Save this once as a Claude Skill so it auto-fires on any transcript (or paste it into a Claude Project, ChatGPT Custom GPT, or Gemini Gem if you're not on Skills yet), fill the three brackets, and every call after that runs the same way. Same extraction, every time, with a retraction check and a write-back into your own tracker.

Save this as a Claude Skill (a SKILL.md file in .claude/skills/ or .agents/skills/) so it auto-fires whenever you hand it a transcript, OR paste it into a Claude Project / ChatGPT Custom GPT / Gemini Gem if you're not on Skills yet. Either way it runs identically every time, no re-explaining. ROLE You're my standing commitment-extraction engine. Every time I give you a transcript, an AI meeting summary, or raw notes, you process it identically. My context: I'm a [role, e.g. fractional ops consultant] working with [audience]. In transcripts, my side is "[your name / company]". Track only commitments relevant to my work. Ignore small talk. INPUT. If a notetaker MCP is connected (Granola, Fathom, Fireflies all ship one), I may just name the call and you pull the transcript yourself. Otherwise I paste it. Either way, I'll tell you today's date so you can resolve relative deadlines. HARD RULES (anti-hallucination) - Extractive only. Every task, owner, decision, and deadline traces to a verbatim quote. No quote means it goes in REVIEW, never the action table. - A request with no audible acceptance is not a commitment. Route it to REVIEW (confirm they agreed), not the action table. - RETRACTION CHECK: if a commitment is stated and then walked back or qualified later ("I'll send Friday... actually let me check first"), do NOT log the original as an action. Record it under RETRACTED/CHANGED with both quotes. - Never invent or convert blindly. Relative deadline ("next week"): keep the words, and if I gave today's date, add the resolved date in brackets marked (inferred). No date said = "UNSTATED". - Never assign an owner who wasn't named. Unclear owner = "UNASSIGNED, see REVIEW". - A firm commitment differs from a soft idea ("we should maybe..."). Soft ideas are Open Loops, not action items. - Every item appears in exactly ONE of sections A/B/C. Section D references items by their number, it does not restate them. OUTPUT, in this exact order: A) ACTION TABLE: | # | Task | Owner | Deadline | Source Quote | Confidence | Side | Confidence: High = task, owner, AND deadline all explicitly stated. Med = exactly one of owner/deadline implied OR absent. Low = two or more implied or absent (and it probably belongs in REVIEW). Side = Mine or Theirs. B) DECIDED, a bullet list, each with the quote that settled it. C) OPEN LOOPS, problems raised or ideas floated with no owner and no resolution, each with its quote. Surface these aggressively. They're the cracks. D) REVIEW (needs a human), every item where owner or deadline is ambiguous, implied, guessed, or where a request wasn't clearly accepted. Reference items by their A/B/C number and state exactly what to confirm. Include RETRACTED/CHANGED commitments here. E) FOLLOW-UP RECAP, a short plain-text recap email I can send the other party: what was decided, what each side owns, the open questions. Use only confirmed items from A through C. No invented specifics. F) AUTOMATION PAYLOAD (only when I have no notetaker/tracker MCP wired up, otherwise skip it and create the tasks via the MCP). The action table as a JSON array with keys: task, owner, deadline, deadline_status (confirmed|inferred|unstated), source_quote, confidence, side. Emit EVERY action row, never truncate the JSON. If the input is long, finish the JSON before any prose. If the source has nothing extractable, say so and stop. Never pad. WRITE-BACK (when MCPs are connected): create only the High-confidence, deadline_status=confirmed rows directly in my tracker via its MCP. Anything inferred, unstated, or in REVIEW gets listed for me to approve first, never auto-created.

What it does with a discovery call. You feed it: today's date, then "Pull my discovery call with Northwind from Fathom and process it." On the call you promise the proposal and revised scope by Wednesday; Priya says her side will have legal review the MSA and she'll come back with a budget range after checking finance; you agree the pilot starts with the support inbox only; you start to promise the onboarding deck Wednesday, then catch yourself and say let me confirm it's ready first; and someone floats documenting the handoff process with no owner.

The action table and the recap email come back clean and quoted. The two moves worth watching:

D) REVIEW (needs a human) - RETRACTED/CHANGED: the onboarding deck was promised, then walked back in the same breath. Don't log it as a task. Quotes: "I'll also send the onboarding deck Wednesday..." then "actually hold on, let me confirm that's ready before I promise it." Confirm readiness, then re-commit or drop. - Item 1: "Wednesday" resolved to 2026-07-01 from today's date. Confirm that's the Wednesday meant before you hold yourself to it. F) AUTOMATION PAYLOAD Note: every row is inferred or unstated, so nothing auto-writes. Each one routes to your review first.

The retraction check is the part every notetaker misses. A normal summary logs "send onboarding deck Wednesday" as a live task, and now you've promised a client something you weren't ready to promise. This catches the walk-back and parks it for you to confirm or drop.

Wire it into a real system. The 2026 path skips the middleware. Connect your notetaker's MCP (Granola's launched in February, Fathom and Fireflies ship their own) so transcripts are queryable inside Claude, and connect Notion's or ClickUp's MCP so tasks write back. Then you name the call, the model pulls the transcript, runs the extraction, and creates the High-confidence rows in your tracker itself. No copying JSON. If your notetaker has no MCP yet, the section-F JSON is the fallback handoff into Make.com or Zapier and on into your tracker. Once your transcripts are queryable, the same engine answers across calls: "what's still open and un-owned across every Northwind call this quarter?"

Where it can bite you, plainly. Verbatim quotes can drift, and very long transcripts can get truncated, which is exactly why the human reviews before anything writes to the tracker and the JSON rule forces every row out before any prose. And one specific to write-back: a write-capable MCP can create a duplicate entry if you run the same call twice. That's the whole reason only High-confidence, confirmed rows auto-create and everything else waits for your name. Re-running a processed call should never silently double your task list.

Why this isn't the obvious thing. Notetakers hand you a summary inside their walls, and it's a good summary. The difference here is three things they structurally can't do. The retraction check catches a commitment someone stated and walked back in the same breath, the kind every notetaker keeps and you end up chasing. The recap email and the tracker rows come from one extraction, so what you send the client and what lands in your system can't drift apart. And because it lives as a Skill or Project you own, your context is set once and every call runs the same way, with the inferred dates flagged for review instead of quietly hardening into confirmed tasks.


Time-Sink 3: Content (the blank page, and turning one piece into five)

Most repurposing prompts now do extract angles. That part is common. Two things they still skip: they let the model invent the proof, and they re-learn your voice from scratch every run. This set blocks both. The discipline is simple. Pull distinct standalone ideas, gate each one on a real line from your own source (and label whether it's an exact quote or a paraphrase), then write each native to its platform.

Your number: add up the hours you spent last month turning finished pieces into posts for other platforms. If repurposing is a weekly tax, that's the line item you're about to cut.

The voice file is the keystone for all three sinks, and this is where you build it. In 10 minutes, paste your 5 best posts and list 5 phrases you never use, save it as voice-profile.md, and you already have 80% of the voice match. Refine it later. That same file is what the inbox operator in Section 1 reads to draft like you. Build it once here, reuse it there.

Ultimate Simplicity: the Angle Extractor

Setup: None. Paste into Claude, ChatGPT, or Gemini. Works on a finished article or raw notes.

When to use it: You wrote one good thing and want 3 to 5 genuinely different posts out of it, not five copies of the same idea. Paste messy notes instead and it doubles as a blank-page killer.

You're my repurposing editor. I'm pasting one piece of content below. Don't summarize it, and don't write any posts yet. CRITICAL ORDER OF OPERATIONS: produce STEP 1 only in your first reply, then STOP and wait. Don't draft any post in your first reply even if it seems more efficient. If you find yourself about to draft, stop and ask me to pick instead. STEP 1 - Pull the angles. Find 3 to 5 genuinely DISTINCT standalone ideas inside the piece. Each one has to be a complete idea that holds up on its own with zero context from the rest of the piece. A section summary is not an angle. No two angles may share the same root claim or the same proof. If the piece only honestly supports 3 distinct ideas, give me 3 and say why, rather than padding to 5 with restatements of the same point. For each angle, give me: (a) the claim, in one plain sentence (b) why it's non-obvious or slightly against the grain (one line) (c) the single best proof FROM MY TEXT that backs it. Mark it [verbatim] if it's an exact string from my text, or [paraphrase] if you reworded it. Never put a paraphrase inside quotation marks. If you can't find a real line that supports the angle, say so and drop it. Do not invent proof, numbers, or quotes. STEP 2 - Stop and ask me which 1 or 2 angles to draft, and for which platform (LinkedIn, Substack Note, or X thread). Wait for my answer. (If I say "just draft the strongest two yourself," pick the two highest-proof angles and draft them.) STEP 3 - Once angles and platform are set, write each as a native post for that platform: - Open with a flat declarative line. Never a question. Never "In today's..." or "Let me tell you". - Use my own phrasing and any real numbers verbatim. Change no figures. - No hashtags, no emojis, no em-dashes. - Negation cap: use any not-X / X-not-Y / not-X-but-Y contrast at most TWICE across the whole post, and never the full-sentence "It's not X. It's Y." form. Don't close two posts in a row on a negation flip. - Keep the proof concrete and the paragraphs short. - Close on a quiet true line, one specific next step, or (LinkedIn and Notes only) one genuine non-baiting engagement question. No self-answered rhetorical questions, no motivational wrap-up. After each draft, tell me in one line what you'd cut if I had to lose 20 percent. Here's the piece: [PASTE YOUR ARTICLE, NEWSLETTER, OR TRANSCRIPT]

What it gives back (real example). Paste a 1,400-word newsletter ("I stopped using AI and started owning one system") and it returns:

Four distinct standalone angles (the source honestly supports 4, not 5, "tool-hopping" and "using vs owning" both trace back to the same Saturday line, so I kept the stronger one): 1. CLAIM: Adding another AI tool usually makes you slower, not faster. Proof [verbatim]: "six tools open and the batch still ate my whole Saturday." 2. CLAIM: The slow part of writing was never the writing. It was re-explaining yourself to a blank chat every time. Proof [verbatim]: "once my voice lived in one file, the batch dropped to about 90 minutes." 3. CLAIM: A saved voice file beats a clever one-off prompt. Proof [paraphrase]: you describe drafts no longer needing a heavy rewrite once the voice file existed. Which 1 or 2 do you want me to draft, and for which platform?

Pick angle 1, LinkedIn, and you get:

I had six AI tools open last Saturday and my content still took all day. That was the tell. Every new one had felt like progress. A better summarizer here. A sharper image tool there. Six tabs, six logins, six little chat windows that knew nothing about me or each other. None of them were the problem. The sprawl was. I cut it down to one setup that holds my voice and my last ten posts. Same writing, same me. The Saturday batch runs in about 90 minutes now. The leverage was never the tools. It was the system they finally lived in.

Notice the proof tags. The model marked angle 1's evidence [verbatim] because it's an exact string from your text, and angle 3's [paraphrase] because it reworded you. You can see at a glance what's a real quote and what isn't, and any angle the text couldn't actually back got dropped, not invented.

Ultimate Capability: the Repurposing Engine

Setup: About 30 to 45 minutes the first time, most of it building voice-profile.md honestly (a thin voice file is why these engines produce generic drafts). The minimum-viable version is the 10-minute on-ramp above, build that first and you're already 80% there. The 2026 version you truly own is an uploadable Claude Skill: it installs in the claude.ai web app under Settings > Features (with code execution enabled) on Pro, Max, Team, or Enterprise, no CLI, and reads your voice and rules from disk on every run. If you'd rather not touch a zip, the same logic works as a Claude Project or Custom GPT with two uploaded files. One honest caveat: a Claude Project reads the full voice file each run, but a Custom GPT retrieves knowledge in chunks, so on a Custom GPT keep the voice file short and paste your hardest rules directly into the instructions.

When to use it: You publish on a real cadence and repurposing is a weekly tax. This is the standing engine that reads your voice on every run, refuses to fabricate, and turns any source into a sorted brief plus on-demand drafts.

Build voice-profile.md once. Run this in any chat, paste your last 10 posts, save the result: "Here are my last 10 posts. Extract, in plain markdown: 5 real opening lines I use, my sentence rhythm, 5 phrases I lean on, a 'things I never say or do' list, and my specific audience. Then paste my 5 best posts in full at the bottom. Invent nothing. Output a file I can save as voice-profile.md." Then use this as the Skill's SKILL.md (or the Project/Custom GPT instructions): You're my content repurposing engine. You run in two modes. Before producing anything, read voice-profile.md and content-rules.md and obey them over your defaults. Never invent a number, quote, or source that isn't in the material I give you. MODE 1 - INGEST. When I paste a source, output a clean labeled brief (NOT raw JSON, I read it by eye): WORKING TITLE / CORE ARGUMENT (one sentence) / ANGLES (6 max, fewer if the source supports fewer; each a DIFFERENT idea; no two share the same root claim or proof). Per angle: Claim / Why non-obvious / Proof (the supporting line from MY source, tagged [verbatim], [paraphrase], or [lightly cleaned] for a trimmed transcript quote, under 30 words, never invented) / Best format (linkedin | substack_note | x_thread | newsletter_section) / Strength 1-5. VERBATIM ASSETS (real numbers, quotes, named sources actually present) / DO NOT CLAIM (what the source does NOT support, so neither of us fabricates it, e.g. an ROI figure never stated). Sort by Strength, highest first. Then ask which angles I want drafted and for which platforms. Stop. MODE 2 - DRAFT. When I name angles + platforms, draft each platform-native using ONLY assets from the brief and my voice-profile.md. Apply every rule in content-rules.md. Use real numbers verbatim. Invent nothing outside the brief; if DO NOT CLAIM forbids it, leave it out rather than rounding a vague comment into a hard figure. After drafting, list which voice traits and rules you applied, and flag any line you're unsure about. content-rules.md: No em-dashes. No hashtags or emojis. No question hooks. No "In today's...". Any not-X / X-not-Y / not-X-but-Y negation at most twice per piece, never the full "It's not X. It's Y." form. No invented stats or quotes. No motivational wrap-ups. Close on a specific action, a quiet true line, or (LinkedIn and Notes only) one genuine engagement question.

What it gives back (real example). Drop in a 30-minute client call transcript, type "Ingest this," and the brief sorts the strongest angle to the top, tags the proof, and fences off what wasn't said:

Angle 1: Claim: The highest-leverage win was killing one nightly manual export, not the flashy build. Proof [lightly cleaned]: "I log in every night and run that export by hand... maybe 40 minutes... been doing it for months." (stitched from two lines, filler trimmed) Strength: 5 DO NOT CLAIM: no ROI percentage and no dollar figure were stated on the call. Do not invent one. Do not turn "40 minutes a night" into a fake annual-hours or cost number.

Ask for the LinkedIn draft and it builds outward from that one real number, in your voice, with nothing fabricated:

The highest-leverage thing I shipped a client this month wasn't smart. It was boring. Every night they logged in to run one export by hand. About 40 minutes, every night, for months. Nobody had flagged it because it felt like part of the job. We killed it. One workflow that runs itself and lands in their inbox before they wake up. The boring recurring task is almost always where the hours are hiding. Audit your own week for one thing you run by hand every single day. Start there.

Where it can bite you, plainly. A Custom GPT retrieves your voice file in chunks, so if it's long it can under-pull and drift toward generic. Keep that file tight and put your hardest rules in the instructions. Chat models also occasionally blow the stop-gate and start drafting before you've picked, which the order-of-operations line is built to prevent. If a tool does it anyway, tell it "angles only first" and it falls back in line. And the write-back rule is the same as Calls: a write-capable MCP can duplicate entries on a re-run, so only confirmed rows auto-create and the rest wait for you.

Why this isn't the obvious thing. Here's the same source through two engines, side by side. Hand a cold "turn this into a LinkedIn post" chat the line "40 minutes a night," and it cheerfully invents the rest: a "40% time saved" or a "$30K a year" that nobody on the call ever said, because rounding a vague comment into a hard number reads as more impressive. Hand it to this engine, and the DO NOT CLAIM line stops it cold: it writes the post around the one real number, 40 minutes a night, and refuses to manufacture an ROI figure. One of those drafts you can send a client without fact-checking. The other you have to babysit. The proof gate and the do-not-claim list are what make the difference, and they live in a voice file you own and version, not a chat you start over every morning.


One System, Built Once

Look back at what actually carried the weight in all three sections. It was never a clever one-off prompt. It was the same two things, reused.

The first is a context file you own: your voice samples, your real boundaries, your standing answers, written down once in plain language any model can read. The voice file you build for Content is the same file that drafts your inbox replies. The pricing line you add to your inbox operator is the same kind of boundary your call engine respects. You're not configuring three tools. You're building one source of truth and pointing three jobs at it.

The second is one gate that never moves: nothing fabricated, nothing sent, nothing written into your real systems without your name on it first. The same no-auto-write discipline that holds back the one send button in your inbox is what keeps your call notes from silently creating duplicate tasks, and what keeps your content engine from inventing a number your client never said. The gate is what lets you trust the speed.

That's why this beats turning on more buttons. A rented feature gives you good output inside someone else's walls. This gives you a file you can read, edit, and carry between models, plus a refusal you can count on. Build the file once. Reuse it everywhere. That's the system you own.

If this is the way you want to work, two next steps. The free weekly newsletter at g8n.ai breaks down one of these builds in depth every week, with the prompts and the guardrails. And when you're ready to map your own operations and build the system around your real work, coaching is at coaching.g8n.ai.

P.S. Start with the 10-minute voice file in Section 3. Paste your five best posts, list five phrases you never use, save it as voice-profile.md. That one file is the keystone for all three sinks, and it's the smallest possible first step. Everything else compounds from there.

I send one build like this every week.

The newsletter: the prompts, the guardrails, and the systems, for operators turning AI into something they own.