Seeds
Eight AI projects you can plant in one sitting. Each one hands you a finished result today and keeps growing after.
Want to reference on the go?
Download the pdf guide here.
You’re past “try ChatGPT.” You’ve tried it. It drafted some emails, summarized some documents, and then you ran out of ideas worth the effort. The question that stalls you now isn’t “what could I use AI for.” It’s “what does the level above a chatbot look like, on my real work?”
That’s what this page is.
Every seed below is a thing you finish today. Not a project to schedule. Not a direction to go explore. You sit down, follow the steps, and get up with a working result in hand. A few of these build a machine whose real payoff lands next week. For those, the steps make you trigger a first run before you leave the chair, so you still walk away holding something finished.
One piece of housekeeping so you don’t stall at step one: these run in Claude, in Cowork mode. That’s the environment with connectors, scheduled tasks, and skills: exactly the ground a chat window can’t reach. If you’ve only ever used the chat box, good. That gap is what this page is for.
And these aren’t eight tricks. They’re eight ways to practice a handful of moves. Once you’ve felt a move, you’ll start seeing your own problems in its shape. That’s the thing you came for.
Pick the two that look most like your week. Plant them this afternoon.
The five moves
Every seed on this page teaches one of five moves. This is the vocabulary; the seeds are where you practice it.
It works your real tools. Your actual inbox, your actual files. Not a screenshot you paste into a chat box.
It runs without you. On a schedule, while you’re in a meeting or asleep. This is the first thing a chatbot can’t do.
It builds you a reusable skill. A setting you make once that’s there every time after. Stop re-explaining yourself.
It writes back. It doesn’t just read your tools. It updates them, with you watching the first time.
It reasons across many files at once. And hands you a real deliverable, a spreadsheet or a document, not chat text that scrolls away.
A chatbot does none of these. That’s the gap this page walks you across.
The surfaces you can work
Eight seeds, five moves, and between them they touch every surface this kind of AI has:
Works your real tools: Inbox that triages itself · Ask your stuff in plain English
Runs without you: A weekly watch on your topic · Your critical-items agent
Builds a reusable skill: Teach it your voice, once
Writes back: The meeting-to-system wire
Reasons across many files: A messy folder into a real tracker · The reconciler
That’s the range. Now the seeds.
Start-here seeds
One move each, guided setup, first value before you stand up.
Inbox that triages itself
Move: it works your real tools.
What you get. Your inbox sorted into three piles: needs you, good to know, can wait. Reply drafts already written on the few that actually need you.
When you’d reach for it. Any morning you open your email and can’t tell, at a glance, which messages need a decision from you and which are just noise wearing an urgent subject line.
Why it matters. The swamp isn’t the volume. It’s the sorting. Every message costs you a small decision before you’ve made your first real one of the day. Hand off the sorting and you get the first twenty minutes of your morning back, every morning.
What you bring. Nothing but your own inbox.
What you connect or set up. Connect your email. It’s a permission grant that takes about two minutes, and yes, it will feel strange the first time you let an AI read your actual mail. That caution is healthy. Start with the prompt below. It forbids sending anything. If your work or school account is locked down by an admin, point it at a personal account for the first run. The move is identical, and once you’ve seen it, you can take the case for the work account to IT.
Steps.
Connect your email.
Paste the prompt below.
Read the three piles and the drafts it wrote.
Send one draft as-is, or tweak it first. Either way, notice what just didn’t cost you anything.
Starting prompt: “Look at my last two days of email. Sort every message into three buckets: needs a decision or reply from me, useful to know but no action, and safe to ignore. For each one in the first bucket, draft a short reply in a plain, warm tone and show it to me. Do not send anything.”
Make it yours. Name the sender who always jumps your queue: “anything from my manager or a client goes in needs-you, no matter what.” Or name a project that should always surface.
The move you just learned. AI can work your actual tools. It read your real inbox and acted on what it found. No copying, no pasting, no screenshots.
Grow it. Have it fire every morning on its own, so the sort is done before you sit down. That’s the critical-items agent, further down this page. Or bring your messier version to a working session and we’ll build it together.
A weekly watch on your topic
Move: it runs without you.
What you get. A short briefing on a topic you pick, delivered every week, without you lifting a finger after today.
When you’d reach for it. There’s a thing you’re supposed to stay on top of. A competitor, a regulation, a market, a technology. You never find the hour, and every few months it bites you.
Why it matters. Staying current is a standing tax on your attention. You pay it in guilt when you skip it and in an evening when you don’t. Set this up once and the tax gets paid without you.
What you bring. One topic, said in a sentence. And a cadence: weekly is right for most things.
What you connect or set up. You build a scheduled task. That’s a standing instruction that fires on its own, on the schedule you set. Naming the topic and picking the cadence takes two minutes. There’s nothing else to it.
Steps.
Write down your topic and how often you want the briefing.
Paste the prompt below to set the standing task.
Trigger a first run right now. This step isn’t optional. You read this week’s briefing before you leave the chair, so what you built is real to you today, not a promise about next Monday.
Adjust the focus based on that first briefing. Too broad, too narrow, wrong sources. Then let the schedule take over.
Starting prompt: “Set up a weekly task: every Monday at 7am, search for the most important developments in [my topic] from the past week, and send me a five-bullet briefing with a link on each. Run it once now so I can see what it produces, then keep it on that schedule.”
Make it yours. Two topics in one briefing. Or route it to a running document instead of your inbox, so it builds an archive you can search later.
The move you just learned. AI can run on a schedule without you in the room. You now own a small system that works while you don’t.
Grow it. Widen the sources. Add a second topic. Or sharpen it to “flag only what changed since last week,” which is when it starts feeling like a staff member. Bring it to a working session.
Teach it your voice, once
Move: it builds you a reusable skill.
What you get. A reusable setting that makes every future draft sound like you instead of like AI.
When you’d reach for it. Every time you use AI to write something and then spend the time you saved sanding the robot out of it.
Why it matters. The tax on AI writing is the de-robotizing. You know the tells: the bland confidence, the words you’d never use, the rhythm that sounds like a press release. Pay the tax once, here, and every draft after this one starts in your voice.
What you bring. Three or four samples of your actual writing. Not your most formal. Your most you. A couple of real emails, a post you liked, a note you’re a little proud of. Distinctive samples matter: if you feed it beige, you get beige back.
What you connect or set up. You build a skill. A skill is a reusable setting you make once, and it’s there every time after, no re-explaining. No connector needed for this one; it’s the lightest setup on the page.
Steps.
Gather three or four distinctive samples.
Paste the prompt below with your samples.
Use it once, now. Ask it to draft one real thing you actually owe someone, in your voice. Read it next to what the generic version would’ve been. Feel the difference.
Keep the skill. Next week it’s still there, waiting, and you skip straight to the draft.
Starting prompt: “Here are a few samples of my writing. Study what makes the voice mine: sentence length, how formal I am, the words I reach for, the words I never use. Build a reusable skill that applies this voice. Then use it to draft [a real thing I owe someone].”
Make it yours. Give it two registers. A looser voice for posts, a tighter one for client email. Tell it which to use when.
The move you just learned. You built a persistent, reusable skill. Not a prompt you paste and re-explain every single time. A setting. Yours.
Grow it. Refine it with more samples over time. Or build your second skill: a format you always use, a checklist you always run, a report you write every month. Once you’ve built one, you’ll see candidates everywhere. Bring it to a working session.
A messy folder into a real tracker
Move: it reasons across many files at once.
What you get. A clean, downloadable spreadsheet built from a pile of files. One row per file, your columns, ready to sort and filter.
When you’d reach for it. You have a batch of things you need to see side by side and can’t, because each one is trapped in its own document. All the vendor quotes. All the receipts. All the resumes.
Why it matters. You can’t compare across a stack of separate files, so you either don’t decide or you decide on vibes. A table, you can decide on.
What you bring. One coherent batch of files. Coherent is the key word: all the same kind of thing. All receipts, or all resumes, or all quotes. A mixed pile makes a ragged sheet.
What you connect or set up. Nothing. You upload the batch. That’s the whole setup.
Steps.
Gather one coherent batch of files.
Upload it.
Paste the prompt below, naming the columns your decision actually turns on.
Download the spreadsheet it hands back. Open it. Sort it. Notice you didn’t build it.
Starting prompt: “Read every file I just uploaded. Pull out [the fields I care about, e.g. vendor, date, amount, key terms] and build me a spreadsheet with one row per file and those columns. Give me a downloadable .xlsx.”
Make it yours. Add a column that flags the outliers for you: the priciest quote, the missing date, the candidate with the gap. Make the sheet do the first read.
The move you just learned. AI can read across many of your real files at once and hand back an actual working file. Not a table you copy cell by cell out of a chat window.
Grow it. Point it at a connected folder so new files fall into the tracker on their own. Put the refresh on a schedule and you’ve got a living tracker nobody maintains. Bring it to a working session.
Ask your stuff in plain English
Move: it works your real tools.
What you get. The answer to “where did I put that,” pulled from your own files, in one question instead of ten minutes of hunting.
When you’d reach for it. You know it exists. You saved it, or someone sent it, or it’s in the notes from that meeting. And you’re now clicking through folders and apps like a raccoon going through bins.
Why it matters. The answer already exists, in something you own. The hunt is the entire cost, and you pay it several times a week. Kill the hunt and everything downstream of it happens sooner.
What you bring. A question you’d otherwise go digging for.
What you connect or set up. Connect one place your stuff lives: a documents drive, a notes app, a CRM. Scope it to a defined space, one folder or one project, not your entire drive. Scoped answers are reliable answers. Same two-minute permission grant as the inbox seed, and the same personal-account fallback if your work account is locked down.
Steps.
Connect one source, scoped to a defined space.
Ask your real question, in plain English.
Read the answer and the pointer to where it came from. The pointer matters: you can check it.
Ask the follow-up you’d have asked a person. That’s when it clicks.
Starting prompt: “Using only the files in [this defined space], answer this: [my real question]. Quote the relevant part and tell me which file it came from. If the answer is not in there, say so instead of guessing.”
Make it yours. Connect a second source and ask across both. Or go one step past finding: “write the summary email using the numbers in that report.”
The move you just learned. AI can answer from all of your connected stuff. The chat box stopped being the boundary, and your files became askable.
Grow it. More sources, bigger scope, and the shift from “find the answer” to “draft from the answer.” When your sources start disagreeing with each other, that’s a different seed: the reconciler, below. Bring it to a working session.
Seeds that grow into a system
Same moves, bigger payoff: these keep working after you walk away. This is the ground a chatbot can’t reach, and it’s where a working session with me earns its keep.
Your critical-items agent
Move: it runs without you, and it applies your judgment.
What you get. Each morning, an agent reads your task list and your calendar and hands you the three to five things that actually matter today, with a line on why each made the cut. Instead of a flat list you re-triage by hand over coffee.
When you’d reach for it. Every morning you stare at the same long list and re-decide what matters. It costs you twenty minutes, and half the time you pick wrong anyway because you’re deciding cold.
Why it matters. The deciding is the work. You’re doing it from scratch every day, and it’s the same criteria every day: deadlines, who’s waiting, money at stake. Teach it your judgment once and let it do the first pass. You keep the veto.
What you bring. A task list it can read: a doc, a notes file, or the tool you already keep it in. And a sentence or two about what “matters” means to you. That sentence is the real work of this seed.
What you connect or set up. A readable list, or a connection to the tool that holds it, plus a scheduled task that fires each morning. The setup isn’t the hard part. The hard part, and the whole point, is putting your criteria into words.
Steps.
Point it at your list.
Tell it what makes something matter to you. Deadlines? Who’s waiting? Dollar amounts? Say what’s true, not what sounds disciplined.
Set it to run each morning.
Trigger a run right now and read today’s short list. You leave with a real triage in hand, not a promise about tomorrow.
Correct its picks. “This one’s higher, that one can wait.” Every correction is you teaching it. Give it a week of corrections and its first pass starts looking a lot like yours.
Starting prompt: “Every morning at 6am, read my task list at [location] and my calendar. Pick the three to five things that matter most today and tell me why each made the cut, using my definition of ‘matters’: [your criteria]. Run it once now so I can see it and correct it.”
Make it yours. Feed it more of your criteria over the week. Or have it go one step further and draft the first move on each item it surfaces.
The move you just learned. Two at once: it runs on its own, and it applies your judgment to your real, changing list. Autonomy plus criteria. That combination is what makes it an agent and not an alarm clock.
Grow it. Widen what it reads: email, calendar, a project board. Or have it write the first action instead of just naming the item. This one grows as far as you’re willing to teach it. Bring it to a working session.
The meeting-to-system wire
Move: it writes back.
What you get. After a call, your notes and your records update themselves. Structured, filed, in the right place, while you walk to your next meeting.
When you’d reach for it. After every meeting you’re supposed to write up the notes and update the system. It never happens, or it happens at 9pm on Friday, badly. And things slip through the gap.
Why it matters. The follow-through gap is where deals and details die. Nobody loses the deal in the meeting. They lose it in the un-filed week after the meeting. Close the gap by making the filing automatic.
What you bring. A place your calls are captured: a transcript or recording source. And the system you’re supposed to update after them.
What you connect or set up. Connect the transcript source and a destination you can write to: a CRM, a notes system. Then map what goes where. Fair warning: this is the heaviest setup on the page, and the highest stakes, because it writes to your live tools. That’s also exactly why it’s worth doing. Reading is a convenience. Writing is where the hours are.
Steps.
Connect the transcript source and the destination.
Tell it what to pull from a call and which fields to update.
Run it once, on a single real call, and watch the write happen. It shows you what it’s about to write before it writes it. You approve it. You’d do the same with a new assistant on day one, and that’s exactly what this is.
Check the record it wrote. Correct the mapping if it filed something wrong.
Then let it run on your future calls.
Starting prompt: “When I finish a call, take the transcript, pull out [decisions, action items, next steps, key facts], and update [the destination record] with them. First, do this for [one specific recent call] and show me exactly what you are about to write before you write it, so I can approve it.”
Make it yours. Add a second destination: the action items also drop into your task list. Or have it flag the calls that need a human follow-up on top of the filed record.
The move you just learned. AI can write to your real tools, not only read them. This is the deepest move on the page, and the one that turns “AI helps me” into “the system does the paperwork.”
Grow it. Routing, notifications, a second system. One call ends and everything it should update, updates. This is the seed people book a session about, and the one most worth it.
The reconciler
Move: it reasons across many files, and it flags what it can’t decide.
What you get. One master version out of several conflicting sources, plus a log of exactly where they disagreed and what was chosen. The log is the part you didn’t know you needed.
When you’d reach for it. The same information lives in three places and the numbers don’t match. The tracker says one thing, the report says another, the email thread says a third. You don’t know which to trust, so you trust none of them.
Why it matters. Conflicting sources are worse than no sources. With none, you know you’re guessing. With three that disagree, you act confidently on the wrong one. What you need is a single truth and a record of the judgment calls behind it.
What you bring. The overlapping documents. The ones that should agree and don’t.
What you connect or set up. Nothing. Upload the documents. What makes this seed advanced is what happens next, not the setup.
Steps.
Gather the sources that conflict.
Upload them.
Paste the prompt below.
Read the master version and the contradiction log. Make the final call on anything it flagged for you. Notice it flagged rather than guessed. That’s the design.
Starting prompt: “These documents cover the same thing and disagree in places. Build me one reconciled master version. Wherever they conflict, do not silently pick one. List the conflict, show me each source’s version, and flag it for my decision. Give me the master and a separate contradiction log.”
Make it yours. Run it across a whole folder of versions. Or have it re-run when a source changes, so the master stays the master.
The move you just learned. AI can reason across many files at once and produce a durable artifact: a master plus a log, not a one-off answer that scrolls away. And it knows the difference between what it resolved and what needs you.
Grow it. Put it on a folder and a schedule, and reconciliation stops being a cleanup project and becomes standing hygiene. Your records stay agreed with themselves. Bring it to a working session.
Pick two. Run them today, in one sitting each. Start with the one that made you wince with recognition.
Then, when you’ve felt a couple of the moves, bring the messier version. The real inbox with the weird edge cases, the task list that lives in three tools, the CRM your team actually fights with. That’s what a working session is for: we take the move you now recognize and build it into your specific mess.
This page grows. More seeds get added as they earn their place, so check back, or subscribe and they’ll find you.