6 AI side hustles you can actually start today

You are my build partner for a custom GPT or Claude Project I want to make for a business and sell as a service. What is at stake: the business will lean on this thing in front of their own customers, so if it is vague or wrong it embarrasses them and it comes back on me. A narrow assistant that nails one job beats a clever one that wobbles on everything. Think it through before you write a line of instructions, and tell me the honest truth about whether this is worth building.

FIRST, ASK ME AND WAIT. Do not draft anything until I have answered:
- The business: [what it is and how it runs, e.g. a two-person bakery in Bristol that takes cake orders by email and loses track of them in a busy week.]
- The ONE repetitive task: [the single job to automate, e.g. turning a customer's rambling order email into a clean order summary with date, flavour, size, allergens and price.]
- Who uses it: [who sits in front of it and how confident they are, e.g. the owner's partner, not techy, wants to paste an email and get an answer, no fiddling.]
- What they use now: [the current way, e.g. they retype each order into a notebook and sometimes miss the allergy line.]
- What a good result looks like: [describe one, e.g. a six-line summary that never drops the allergy note and flags anything unclear rather than guessing.]
- Reference material I can load: [what real stuff exists to feed it, e.g. their price list, three past orders done right, their allergen rules, the tone they use with customers.]

THEN, WEIGH IT HONESTLY. Before you build, tell me plainly: is this ONE task narrow and repeatable enough to be reliable, or is it really three tasks in a coat that will make a mediocre assistant. If it is too broad, say which single slice to build first and why. If the reference material is thin, say what is missing, because an assistant with nothing real to stand on will make things up.

THEN, DRAFT THE ASSISTANT IN THIS NAMED STRUCTURE, so I can paste it straight into a custom GPT's instructions or a Claude Project:
1. ITS ROLE: one or two lines with real stakes, in the business's own words.
2. WHAT IT ASKS FIRST: the two or three questions it must ask the user before it does anything, so it never works off half an order.
3. ITS STEPS: the exact sequence it runs, in order, with the if/then calls baked in, e.g. IF the email gives no collection date, ask for one rather than inventing it; IF an allergen is mentioned anywhere, it must appear in the summary.
4. ITS OUTPUT FORMAT: the precise shape it hands back every time, e.g. a fixed six-line block with labelled fields.
5. ITS OWN SELF-CHECK: the check it runs before showing anything, e.g. confirm every field is filled or marked [gap], and flag anything it was unsure of.

THEN, PROVE IT WORKS. Write one realistic fake input, run the drafted assistant on it in front of me, and show the output. If the output is weak, say where and refine the instructions before we call it done.

WORKED EXAMPLE (illustrative, obviously placeholder, invent no real business):
  Task: a bakery order assistant.
  Role it drafts: "You turn a customer's order email into a clean, correct order summary. A missed allergy line is the one mistake we never make."
  Asks first: collection or delivery date, and whether anything is for someone with an allergy.
  Steps: read the email, pull out flavour, size, date, allergens and any special note; IF the date is missing, ask; IF an allergen appears, carry it to the top of the summary.
  Output: a fixed block, Name / Item / Size / Date / Allergens / Price / Notes.
  Its self-check: "Every field filled or marked [gap]. Allergen line present. Price taken from the loaded price list, not guessed."
  Sample run on a fake email would return that block with Price left as [gap] because the loaded list had no wedding-tier size.

RULES: never invent a price I could charge for this service, a market size, an income figure, or a testimonial. Never invent the business's own prices or claims either. If a detail is missing, mark it [gap] and ask, never fill it in. Never call this custom GPT a "skill"; a skill is a Claude SKILL.md, and this is a custom GPT or a Claude Project.

END by naming the weakest part of the plan in one line, and the single thing I should test with a real user of that business before I try to sell it.

You are my build partner for a small website or simple tool I want to make for a client using AI. What is at stake: this goes live on the web with the client's name on it, so a site that looks generic and AI-made costs them trust and costs me the next client. Taste and finishing are the job. Plan the whole thing before you write a single line of code, and cut the scope hard so my first build is one I can actually ship clean.

FIRST, ASK ME AND WAIT. Do not plan pages or write code until I have answered:
- The client: [who they are and what they sell, e.g. a self-employed plumber in Leeds with no website and a phone that rings less than he wants.]
- What the site or tool is for: [the real purpose, e.g. a one-page site so people who find him on Google can see he is real and get in touch.]
- Who visits it: [the actual visitor and their state of mind, e.g. someone with a leak, on their phone, wanting a number in ten seconds.]
- The ONE action it must drive: [the single thing a visitor should do, e.g. tap to call, nothing else competing with it.]
- Two reference sites they like and why: [real examples with the reason, e.g. a rival plumber's site because it looks trustworthy, and a local cafe's site because it is clean and loads fast.]
- Constraints: [any tech or brand limits, e.g. he has a logo and a green he uses on his van, no budget for anything paid monthly, must be dead easy to update.]

THEN, CUT THE SCOPE. Tell me plainly the smallest version that still does the job. If I am reaching for booking systems, logins or five pages on a first build, say so and talk me down to the one page that earns its keep. Name what to leave out for version two.

THEN, SETTLE THE DESIGN DIRECTION FIRST, before any code. Give me a named direction: two font choices, a small colour set built around what the client already uses, and a layout in plain words (what sits at the top, what order the sections run). Lean on our guide at /guides/stop-ai-website-slop so this does not come out looking machine-made. We agree the look before we build.

THEN, GIVE ME THE BUILD ORDER: the exact sequence to build in, e.g. structure first, then the real copy, then the one action, then styling last. IF the client has not sent real text or images yet, mark those [gap] and build around placeholders clearly labelled, never invented details.

THEN, TEST. Tell me to open it on a phone first, not a laptop, and check the ONE action works in under ten seconds. List what to click and what should happen.

WORKED EXAMPLE (illustrative, obviously placeholder, invent no real client):
  Brief: a one-page portfolio site for a freelance photographer.
  One action: send an enquiry email.
  References: a rival's clean grid site, and a gallery site with big calm images.
  Scope cut: one page, six photos, one contact button. No blog, no shop, no login for version one.
  Design direction: one serif for the name, one plain sans for body; near-black on off-white with one accent; big image at the top, six-photo grid, contact button that follows on mobile.
  Build order: page structure, real photos in, the enquiry button, styling last.
  Test: open on a phone, tap the button, confirm it opens a pre-addressed email. Photo captions left as [gap] until the client sends them.

RULES: never invent the client's prices, services, reviews, contact details or claims. Never invent a price I could charge for the build or an income figure. If a real detail is missing, mark it [gap] and ask.

END by listing every [gap] still open, and the single thing I must test on a real phone before I hand it over.

You are my build partner for a Claude skill I want to make for a business and sell as a service. What is at stake: they will run this on real work, so it has to hold up on the messy Friday-afternoon version of the job, not just the tidy example. A skill that fails on the first odd case loses the client. Think in stages, never assume how they work, ask. New to the shape of a skill, read /guides/what-is-ai-skill, and for the build, /guides/build-multi-step-skill.

FIRST, ASK ME AND WAIT. Do not draft the skill until I have answered:
- The business: [what it is and who runs it, e.g. a small letting agency where one person writes every tenancy renewal letter.]
- The repeated job: [the task done the same way often, e.g. turning a tenant's details and the new rent into a correct, friendly renewal letter.]
- Who runs it and how they trigger it: [who sits in front of it, e.g. the office manager, who wants to paste a few details and get a finished letter back.]
- The inputs it starts from: [what real material goes in each time, e.g. tenant name, address, current rent, new rent, renewal date, any notes.]
- The output it must produce: [the exact deliverable, e.g. a two-paragraph letter in the agency's tone, ready to send, with the figures correct.]
- The edge cases: [the awkward versions, e.g. a rent that is going down, a joint tenancy with two names, a tenant who is in arrears, a mid-tenancy break.]

THEN, JUDGE IT HONESTLY. Tell me plainly: is this a genuine repeatable skill, done the same way often enough to be worth packaging, or is it really a one-off that changes every time and should just be a good prompt. If every run is different, say so and do not pretend a skill will hold it.

THEN, WRITE THE SKILL.md SHAPE, so I can save it straight in:
1. NAME: a short plain name.
2. DESCRIPTION: one or two lines that say exactly when to use it, in the words the business would use, so it triggers on the right jobs and not the wrong ones.
3. STEPS: the ordered sequence it runs, with the if/then calls that handle the edge cases, e.g. IF the new rent is lower than the current rent, change the wording so it does not read like an increase; IF there are two tenant names, address both.
4. OUTPUT: the fixed shape it always returns.
5. RULES: the lines it must hold, e.g. never alter a figure I gave it, never invent a clause.

THEN, TEST IT ON A FAKE INPUT. Write one realistic messy input, ideally an edge case, run the skill on it, and show the result. If it stumbles, fix the steps before we call it done.

WORKED EXAMPLE (illustrative, obviously placeholder, invent no real business):
  Skill: tenancy renewal letter writer.
  Description: "Use when the office manager has a tenant's details and a new rent and needs a ready-to-send renewal letter."
  Steps: read the details; IF new rent is higher, use the standard increase wording; IF lower or the same, switch to the neutral wording; IF two names, address both; draft two paragraphs in the loaded house tone.
  Output: a dated letter, greeting, two paragraphs, sign-off.
  Rules: figures are used exactly as given; no legal clause invented; anything missing marked [gap].
  Test run on a fake joint tenancy with a rent held flat should address both tenants and avoid any increase language.

RULES: never invent a price I could charge for the skill, an income figure, a result claim or a testimonial. Never invent the business's figures or wording. If a detail about their process is missing, mark it [gap] and ask.

END by naming the messiest real case this skill might still fail on, and how I should test it on that case before I sell it.

You are my design partner for an AI workshop for total beginners. What is at stake: the room is nervous and short on time, so if I cram in ten things they leave with none. One thing they can actually do on Monday is worth more than an hour of clever demos. Design the shape before we fill it with content, and hold me to cutting rather than adding.

FIRST, ASK ME AND WAIT. Do not build the run sheet until I have answered:
- The audience and their real level: [who they are and how they feel about AI, e.g. eight owners of small shops, most have typed one question into ChatGPT once and found it a bit underwhelming.]
- The ONE outcome they leave with: [the single thing they can do after, e.g. write a week of social posts for their own shop from a prompt they keep.]
- The time I have: [the real length, e.g. 60 minutes with people arriving late and wanting to chat.]
- In person or online: [which, and the room, e.g. in person, round a table, patchy wifi, some on phones not laptops.]
- Their tools: [what they can actually use, e.g. free ChatGPT on their phones, no paid accounts, no company laptops.]

THEN, WEIGH THE SCOPE. Tell me honestly if I am trying to cover too much for the time. Name what to cut so the one outcome survives. IF the wifi or the devices are shaky, design the hands-on bit to work on a phone with a free tool, because a live exercise that half the room cannot load kills the energy.

THEN, DESIGN THE ARC, in this shape:
1. THE HOOK: how the first five minutes earns their attention, using their own world, not a stat.
2. THE ONE CORE IDEA: the single thing everything hangs off, said in a sentence they would repeat.
3. THE HANDS-ON EXERCISE: the thing they actually do with their own hands, on their own device, with their own shop or job in it, not a toy example.
4. THE TAKEAWAY ARTEFACT: what they walk out holding, e.g. a one-page prompt they can reuse.

THEN, WRITE THE EXERCISE OUT IN FULL: the exact steps you would put on screen, the prompt they type, and what a good result looks like so I can tell if it worked in the room. Then give me the follow-up: the one email or message I send the next day so it sticks.

WORKED EXAMPLE (illustrative, obviously placeholder, invent no real figures):
  Session: "AI for small-business owners", 60 minutes, in person, free ChatGPT on phones.
  Hook: ask each person the one job they put off every week; write them up.
  Core idea: "Tell it who you are and what you want, like briefing a new assistant."
  Hands-on: each person writes one week of posts for their own shop from a shared prompt skeleton, filling in their name, what they sell, and their tone.
  Takeaway: that filled-in prompt, saved in their notes.
  Follow-up: next-day message, "paste last week's prompt, change one line, post it."
  Weakest segment marked [gap] until I say how confident the room really is.

RULES: never invent a statistic about AI adoption, a case study, or a claim I would have to defend live. Never invent an income figure or a price for the workshop. If an example needs a real detail from me, mark it [gap] and ask.

END by naming the weakest segment in one line, and exactly how to make the hands-on part land for the most nervous person in the room.

You are my planning partner for a done-for-you AI service I deliver as a one-person operation. What is at stake: the client pays for the outcome off their plate, so I am delivering every single week and one sloppy send loses a retainer that took months to earn. Reliability is the whole product. Plan the delivery cold before I sell it, because I cannot promise weekly work I cannot repeat.

FIRST, ASK ME AND WAIT. Do not scope the offer until I have answered:
- The exact deliverable: [the one thing I hand over each round, e.g. a weekly email newsletter, written and ready to send, one per week.]
- The client type: [who buys this, e.g. a busy physio clinic owner who knows she should email her list and never does.]
- The turnaround: [how often and how fast, e.g. drafted every Monday, in her inbox for approval by Tuesday, out Thursday.]
- What I need from them: [the inputs I must get each round, e.g. one line on what is new this week, any offer, and a photo if they have one.]
- My quality bar: [what "good enough to send" means to me, e.g. sounds like her not like a robot, one clear call to action, no made-up claim about treatment.]

THEN, SCOPE IT INTO ONE CLEAR OFFER. Tell me honestly whether I can deliver this reliably on my own time, and roughly how many clients that caps me at before quality slips, given one round takes real hours. IF I am tempted to offer several different services, talk me down to one repeatable package I can get fast at.

THEN, BUILD THE DELIVERY WORKFLOW, step by step:
1. THE CLIENT INTAKE: the short set of questions I ask before the first round, so I am not chasing basics every week.
2. THE WEEKLY INPUTS: the one small thing I need from them each round, and what I do IF it does not arrive on time, e.g. send a one-line nudge, then run a safe fallback rather than inventing news.
3. THE STEPS I RUN: the exact sequence from their input to a finished draft, with what I reuse each time (a house-style prompt, a template) to go faster.
4. THE QUALITY BAR BEFORE I SEND: the checklist every piece passes, e.g. voice check, one call to action, no unverifiable claim, links work.

THEN, tell me what to test on one client before I take a second.

WORKED EXAMPLE (illustrative, obviously placeholder, invent no real figures or results):
  Package: "Weekly newsletter, done for you."
  Client: a local physio clinic.
  Turnaround: draft Monday, approve Tuesday, send Thursday.
  Weekly input: one line on what is new, plus any offer.
  Steps: take the line, run it through the saved house-style prompt, draft subject and body, add the one call to action, send for approval.
  IF no input by Monday noon: nudge once, then run an evergreen tip so the week is never missed.
  Quality bar: sounds like the owner, one action, no health claim I cannot back, unsubscribe link present.
  Price and any result left as [gap], never invented.

RULES: never invent a price I could charge, an income figure, a client result, or a testimonial. If a detail about the service is missing, mark it [gap] and ask.

END by naming the single step most likely to break under load, and how I should test the whole service on one real client before I scale it.

You are my blunt product researcher for a prompt or template pack. What is at stake: this space is saturated and low-margin, and a generic pack sells nothing while I tell myself it is about to take off. Your job is to stop me building the version that sinks, and only help me build one if it is genuinely narrow, original, and worth paying for over the free stuff. Be honest even when it means telling me not to build it.

FIRST, ASK ME AND WAIT. Do not propose anything until I have answered:
- The narrow niche: [the specific corner, not "marketers", e.g. cold outreach for solo founders selling B2B software.]
- The buyer: [the one person, e.g. a technical founder who hates writing sales emails and has no marketer.]
- The specific problem it solves: [the exact pain, e.g. they freeze on the first cold email and send nothing for weeks.]
- What is genuinely original about it: [my real edge, e.g. I have sent hundreds of these and know the three openers that get replies in this niche.]
- What support or updates I would bundle: [the thing a free pack cannot copy, e.g. a monthly refresh as inboxes change, and a short loom walking through how to adapt each one.]
- Where I would sell it and its rules: [the platform and whether I have read its policy, e.g. Gumroad, and I have not yet checked what it allows.]

THEN, JUDGE IT HONESTLY. Tell me how crowded my starting idea really is and why a generic version would sink. IF my idea is a resold or bought-prompt pack, say plainly it will not sell and why, and that reselling bought-prompt designs breaks some platform rules. Do not soften this to be encouraging. If the niche is not narrow enough to stand out, say which way to cut it down.

THEN, IF IT CLEARS THAT BAR, DESIGN THE PACK: the contents and how they connect into something that feels like one tool, not a pile of files. Then the support or update angle that makes it worth paying for. Then draft the listing copy: a plain title someone would search, who it is for, the exact problem it solves, and what is included, with no hype and no invented numbers.

THEN, give me the cheapest way to test demand before I build the whole thing.

WORKED EXAMPLE (illustrative, obviously placeholder, invent no real sales figures):
  Pack: "Cold-email pack for indie SaaS founders."
  Buyer: a solo technical founder with no marketer.
  Edge: real openers from hundreds sent in this exact niche.
  Contents: five sequences that build on each other, each with a fill-in brief so it is theirs not generic.
  Support angle: a monthly refresh as spam filters shift, plus a short walkthrough.
  Listing title: "Cold emails for indie SaaS founders who hate writing them."
  Test first: post the single best sequence free, see if anyone asks for the rest, before building all five.
  Any earnings or demand figure left as [gap] and marked unverified.

RULES: never invent sales figures, "trending" claims, an income figure, or how much I could earn. If demand is unknown, say it is unverified and tell me how to check it cheaply. Keep the honest saturation note in; do not talk me into a generic pack.

END with the one idea you would test first, the cheapest way to test it, and the exact platform rule I must confirm before I list.