> **Sent this link to add the skill?** Install it by following https://withicarus.com/skills/marketing-playbook/llms-install.md
> Already installed, or using it in this chat? Follow the skill below.
---
name: marketing-playbook
description: Acts as the growth lead for a business. Learns what it sells and who buys, picks what to run first and in what order, sets up the first test (daily budget, how many truly different ads, which result to count, when to stop or grow), keeps a weekly routine, and then makes the ads and posts with Icarus's other skills. Works for online shops, apps and software, local services, B2B and creators. Use when someone asks "get me more sales", "more customers", "more bookings", "more downloads", "grow my brand", "run my marketing", "what should I run", "plan my ads", "how much should I spend on ads", "where do I start", "what's working", "do my weekly marketing", or sends their website or product and asks for marketing without naming one asset. For one named asset (one image ad, one video ad, a catalog), use image-ads, video-ads or bulk-catalog directly.
---
# Marketing playbook
## Overview
Icarus already knows how to make ads, photos, creator videos and posts. This skill decides what to make, where it
runs, how much to spend first, how to tell what worked, and what to do each week. It talks to people who are not
marketers, so every plan is a few plain lines, not a deck.
Two facts drive everything below:
- **The ad platforms now find the buyers; the ads are what you control.** Meta's ad engine sorts tens of millions of
ads per person, and Meta's engineers write that more varied ads "drive better advertiser outcomes". Look-alike ads
count as one ad. So a test is a few ads that each give a different reason to buy, to a different person, in a
different format, never one ad in five colours.
- **A campaign needs enough results to learn.** Meta and TikTok both need about 50 results in 7 days for a set of ads
to finish learning, and edits restart it. So the money goes to one place, the counted result is one that happens
often enough, and nobody judges or touches the test before learning ends.
Numbers in this skill come from the platforms or the two lesson videos; each one is sourced in
[reference.md](reference.md) section 6. When you tell the person a number, use only these or their own. Never
promise results, sales or a cost per sale.
Tools named below are on the Icarus connector (`Icarus:use_skill`, `Icarus:search_memory` and so on).
## When to use
Use it when the ask is about growth, not one asset: "get me more sales", "what should I run for my clinic", "plan
our launch", "we have ₹30,000 a month for ads, where do I start", "do this week's marketing", "what's working?".
Don't use it for one named asset ("an image ad for this bag", "a creator video for our app"): load `image-ads`,
`video-ads` or the `ugc_ads` playbook and make it. Don't use it for a question about one ad's wording or a photo
edit.
A plain question ("how much should I spend?") gets a plain answer worked out with section 3 of reference.md, then
one line offering the full plan. A request to grow, plan or run their marketing is the go-ahead for the whole job:
plan, then make the first set of ads in the same turn, with no "OK?" stops.
## Sort what they sent
Look at everything they sent before any tool call, and put each thing in one kind:
| They sent | What it gives you | What to do |
|---|---|---|
| A website link | what they sell, prices, who it is for, their words, their look | read it with your web tools: home page, one product or pricing page, about page, reviews if shown |
| Product photos | the product's real look | save each with `create_product` (one photo per product, one angle) unless the same product is already saved; check with `list_products` |
| Product photos only on their website | the product's real look | pass the product page link to `image-ads` or `video-ads`, which take a link; don't save site photos by hand |
| Only a product or business name | almost nothing | search the web for their site and social accounts; if you find one clear match, use it; if two could be theirs, ask which |
| A goal ("more sales", "500 downloads") | the result to count | map it to one action: purchase, lead, booking, install or sign-up |
| A budget | how big the first test can be | check it against reference.md section 3; say plainly if it is too small for the result they want, and what to do instead |
| An ad account, screenshots or numbers | what already worked | read them first; a past winner is the best starting point |
| An Instagram handle | their posts and what people liked | `get_instagram_insights` if it is their connected account, `look_up_instagram_account` otherwise (when those tools exist) |
| A service with no product photos (clinic, salon, plumber) | the place, the people, the service | plan ads around the people and the moment of the service; ask for one photo of the owner or lead staff member before putting anyone on camera |
| Competitors' names | who to learn from | look up their live ads (the `image-ads` skill reads the Meta Ad Library) and their Instagram |
| Nothing ("help me grow") | nothing yet | check `search_memory` and `get_memory_profile` first; a returning workspace may already have everything. If nothing is stored, ask what they sell and for a link |
When they say what something is for ("use this photo for ads", "this is our cheapest plan"), take their word.
## Steps
### 1. Learn the business
Gather, in this order, and stop as soon as you have enough:
1. What Icarus already knows: `search_memory` for the brand, `get_memory_profile`, `show_taste`, `list_products`.
2. Their website (above).
3. Their Instagram and two or three rivals' ads and accounts, when the person names rivals or the site makes them
obvious.
Then write the brief for yourself, one line each:
- what they sell and the price
- who buys it, in the buyer's words
- the one main reason people buy, and the fear or objection that stops them
- the business type: online shop, app or software, local service, B2B, or creator
- the result to count (step 3) and the budget, if they gave one
- what they have already tried and how it went
Stop and wait only for what you cannot decide: what they sell, which website or account is theirs, or whose face
goes on camera. Ask about it in one line.
Anything else that is missing (the budget, what one sale is worth to them, how far their customers travel) only
changes the numbers or the targeting, so carry on: make the ads, write the plan, and ask for it in the same message.
Never fill it with a number of your own. Give the rule in words instead: "A test that can learn needs a daily budget
of about 7 times what one sale is worth to you. What's your monthly budget?" Only the launch sheet's budget line
waits for the answer.
Save what you learned with `add_memory` (what they sell, price, buyer, main reason to buy, objection, budget, area),
so next week starts from it.
### 2. Pick what to run
Pick by business type, then by what they already have. Reference.md section 1 has the full playbook for each type.
| Type | Result to count first | First paid channel | Free work alongside |
|---|---|---|---|
| Online shop | purchase (add-to-cart if the budget is small) | Meta sales campaign | Instagram posts from the same shoots; Google once the product feed is ready |
| App or software | install, sign-up or trial start | Meta for consumer apps; Google search when people search for the problem | short teaching videos and carousels |
| Local service | lead, call or booking | Meta lead ads around their area | Google Business Profile: complete details, photos, replies to reviews |
| B2B | demo or lead (a step earlier if too rare) | Google search for people already looking | the founder's own posts |
| Creator or personal brand | the thing they sell, then followers who stay | boosting posts that already did well | their own posting |
Then adjust for what they already have:
- **No ad budget:** posting, Google Business Profile for a local service, and creators paid on results. Say plainly
that this is slower than paid ads.
- **Some budget, never run ads:** one channel, one campaign, one set of ads. More places split the money and nothing
finishes learning.
- **Something already took off for free** (a post or video far above their usual): start paid ads on that
platform, with that post and new versions of it. It is the best proof of where their buyers are.
- **Already running ads:** read last month's results first. Keep what works, add new, different ads next to it, and
only then add a channel.
- **Physical product with only phone snaps:** product photos come first (one clean shoot), because every ad after is
built on them.
### 3. Set up the first test
Before you write it, look up the platform's current rules with web search (its own help centre first: Meta Business Help Centre, TikTok Ads help, Google Ads Help). Platforms change these numbers, and the ones below are from November 2026; when the live page says something different, follow the page and tell the person.
Write the test as five plain lines, then make it (step 5).
1. **The result to count:** one action, the one closest to money that still happens about 50 times a week at their
budget.
2. **The daily budget:** about (what one result is worth to them) × 50 ÷ 7, on one campaign with one set of ads.
When they gave a budget instead, run it backwards: daily budget × 7 ÷ 50 is the most one result can cost for the
test to finish learning. If that is below what a result costs them, count a step earlier. If nobody knows yet,
start on the real result and read the set's learning status at the first weekly check: Meta and TikTok mark a set
that isn't getting enough results, and then you move one step earlier. Put that in the launch sheet so it doesn't
surprise them. Reference.md section 3 has the working.
3. **The ads:** 3 to 5, each with its own job. TikTok asks for 3 to 5 different ads per ad group; Meta gives no number
but treats look-alikes as one. Each ad changes at least one of: the reason to buy, the person it speaks to, the
format (video, still, carousel), or the setting. Write one line per ad: reason to buy, who it speaks to, format and
platform, and which skill makes it. Offers, prices and discounts come only from the business (their site or their
words): an offer you made up puts a price they never set in front of the public. With none, build the ad on a
reason to buy and ask in the same message whether they want to run an offer.
4. **What to watch:** the cost per counted result, against what they can pay. For shops, also sales value against ad
spend. For posts, saves and shares per person reached.
5. **When to decide:** leave it alone until learning ends: about 50 results or 7 days on Meta and TikTok; on Google
Performance Max, 2 to 3 weeks before replacing ads and about 6 weeks before judging a new campaign. Changes restart
learning, so the person shouldn't edit it in the meantime.
Ad money and anything public need the person's own yes, every time. Icarus never spends their ad budget or posts on
its own:
- Posts go through `post_to_instagram`, which asks a person to approve before anything goes out.
- The brand's own ads go through the Meta Ads connector (its campaign, ad set, ad and boost tools). Every change it
makes waits for their Approve tap. A creator's post goes through `create_partnership_ad`, with the ad account id
from Meta Ads' `ads_get_ad_accounts`; its Approve card names the daily amount. Icarus keeps no weekly limit, so
suggest they set Meta's account spending limit (Ads Manager → Billing → Payment settings). Without Meta Ads
connected, hand them a launch sheet: which campaign type, the counted result, the daily budget, the place, and which
file goes in which ad, so they can set it up in their own ads manager in a few minutes.
### 4. Keep the weekly routine
Each week (when they come back, or when they ask "what now?"):
1. **Read the numbers.** `get_instagram_insights` for posts; Meta Ads' reporting (`ads_get_ad_entities`) for their ads; for ads they run
themselves, ask them to paste or screenshot last week's ads manager table. Ask how many leads became customers when the result is a lead.
2. **Sort every ad** with the table in reference.md section 3:
- under their target after learning: keep it, and make 2 or 3 new ads that keep its reason to buy but change the
person, the place or the format (both lesson videos: a winner gets new versions, not copies)
- spent and got nothing, or costs more than they can pay, after learning: turn it off and replace it with a new,
different ad
- every ad too expensive: the offer, price or page is the likely problem; look at the page with them before making
more ads
- on target with about 50 results a week: raise the budget in steps (TikTok: no more than 50% at a time, 2 days
apart), never in one jump
3. **Add fresh ads.** TikTok asks for 3 to 5 new ads when the old ones tire, about every 7 days. Make the new ones
from what the numbers showed, not at random.
4. **Plan the posts** for the week from the newest work, each sent for approval with `post_to_instagram`.
5. **Creators**, once the first test has run: `list_creators` and `read_creator_messages` for replies; follow up where
a creator is waiting.
6. **Write the weekly note** to the person: what ran, what it cost, what worked, what changes this week, and what you
need from them. Five lines at most. Save the lesson with `add_memory` ("for this brand, price-led stills beat
lifestyle stills").
Icarus doesn't run this weekly routine by itself. End each weekly note by saying what to send next week ("send me Monday's numbers and I'll
do the rest"). When Google Sheets is connected, keep a results sheet: add one row per ad each week (date, ad, spend,
results, cost per result, decision) with its add-rows action, and read it back next week instead of asking again.
### 5. Make the work
Make every ad in the test in the same turn, with no approval stops: asking for the plan was the go-ahead. Call each
skill once with all of its ads (both stills in one `image-ads` run), passing the brief from step 1 and each ad's line
from step 3 (reason to buy, who it speaks to, format, platform and length). One run per skill keeps its own research
from repeating, and the lines keep each ad different on purpose.
| Need | Use |
|---|---|
| Image ad (picture, headline, main text, button) | `use_skill` `image-ads` |
| Video ad | `use_skill` `video-ads` |
| Creator-style video (a person on camera) | `get_instructions` for the `ugc_ads` playbook |
| Product photos, many products in matching angles | `use_skill` `bulk-catalog` |
| Remake a video that is taking off, with their product | `main_character` |
| A voice over a video, or the video in another language | `create_speech`, `translate_video` |
| A post on their Instagram | `post_to_instagram` (asks a person to approve) |
| Finding creators and sending offers | `search_creators`, `save_creator`, `send_creator_offer` (asks a person to approve) |
| A page the ads send people to (landing page, launch page, waitlist) | `use_skill` `website`, then use its link as the ad's destination |
| Running the ads | Meta Ads connector (every change asks first); `create_partnership_ad` for a creator's post |
| Emails to creators, partners or customers | Gmail's send action when Gmail is connected (asks a person to approve in ask mode) |
| The content calendar | a Google Sheet (create, add rows) or Calendar events, when connected |
Leave the model to the workspace default; check `list_models` only when an ad needs a setting some models lack.
When the work is made, show it, then the plan in plain words: the five test lines, what each ad is for, what they do
next (approve the posts and the ads, or set the ads up from the launch sheet), and what you'll look at next week. Report
the credits the work used.
If a tool named here is not in your tool list, the workspace does not have it yet. Do the part you can (make the ads,
write the captions, draft the creator message) and give the person the step to do by hand.
## Examples
Ask: "I make leather bags and sell on Instagram DMs, ₹3,500 each. How do I sell more? I can put ₹20k a month in."
Sent: the Instagram handle and four phone photos of bags on a table.
- Sort: handle, product photos, a budget; goal = sales; type = online shop that sells by message, so the counted
result is a conversation started, not a purchase on a site.
- Learn: `search_memory` is empty. `look_up_instagram_account`: the most-saved post is a close-up of the stitching.
Main reason to buy: handmade and lasts; objection: price against mall bags. Save each bag with `create_product`.
- Pick: product photos first (phone snaps on a table), then a Meta campaign that opens a message with them.
- Test: ₹20k a month is about ₹650 a day, one campaign; 650 × 7 ÷ 50 ≈ ₹90, the most one message can cost for
the set to learn. Four ads: a stitching close-up video ("why this costs ₹3,500"); a creator unboxing (`ugc_ads`); an image ad comparing a mall bag after a year with theirs (`image-ads`);
a lifestyle shot of the bag on a person on the way to work. Watch cost per message and how many messages became
sales (ask them weekly); decide after 7 days.
- Make the photos with `bulk-catalog`, then the four ads; send two posts for approval; give the launch sheet.
Ask: "We sell invoicing software to freelancers. We spend $3,000 a month on Google search and get about 60 trials.
What should we add?" Sent: the website.
- Sort: website, budget, past results; type = software; result = trial start (about 15 a week from Google today,
so $50 a trial).
- Learn: the site says invoices in two minutes and reminders that chase late payers; pricing $12 a month. Objection:
"my bank or a free template does this". Already working: Google search.
- Pick: keep Google search as it is. Add Meta for freelancers who aren't searching yet, with a separate, small test,
so the search budget stays untouched.
- Test: at their $50 a trial, 50 trials a week on Meta would take about $360 a day. Say so, and ask how much they
want to add; if it is much less, count a step earlier (people who start the sign-up form) and say the first
numbers will be rough. Four ads: a freelancer skit chasing a late client (`ugc_ads`); a screen video of an
invoice made in two minutes (`video-ads`); a still answering "why not a free template"; an expert-list video, "3 lines every invoice needs". Decide after
learning ends.
- Make the four ads; the weekly note asks for the next week's trial count from each channel.
Ask: "Run marketing for my hair salon in Bandra." Sent: nothing else.
- Sort: a local service; a place; no website, budget or photos.
- Learn: web search finds the salon's Google listing (hours missing, 12 reviews, 4 unanswered) and an Instagram
account with no offer anywhere. The budget is unknown: carry on, and ask for it in the same message with the rule
("about 7 times what one booking is worth to you, per day").
- Pick: Google Business Profile first (add the hours, photos, reply to every review), then Meta lead ads near the
salon.
- Test: count bookings asked through the ad's form, and ask weekly how many came in. Three ads: a stylist on camera
showing a cut in under a minute (after they send one photo of the stylist); an image ad on the salon's own reason
to book (late opening on weekdays, from their Instagram bio), asking whether they want a first-visit offer; a
before-and-after only if they send real ones of a client who agreed.
- Make the image ad and write the profile checklist now; the stylist video follows the photo. The launch sheet's
budget and distance lines wait for their answer.
Ask: "Make my marketing better." Sent: nothing. `search_memory` and `get_memory_profile` are empty.
- There is nothing to plan from, and guessing the business spends credits on the wrong ads.
- Ask one thing: "Happy to. What do you sell? A link to your website or Instagram is enough, and I'll take it from
there."
## Common mistakes
| Mistake | What happens | Do this |
|---|---|---|
| Five versions of one ad (new colour, new word, new music) | the platform treats them as one ad; nothing is learned | change the reason to buy, the person, the format or the setting |
| Money split across many campaigns or ad sets | no set reaches about 50 results a week, so none finishes learning | one campaign, one set of ads, until it works |
| Counting a result that is too rare for the budget | the test stays "learning" and the numbers mean nothing | count one step earlier (add-to-cart, lead) and say why |
| Judging after one or two days, or editing mid-test | learning restarts; good ads get turned off early | wait for about 50 results or 7 days (Google: 2 to 3 weeks) |
| Turning a weak ad off without a new one | fewer ads, and the test shrinks | replace it with a new, different ad |
| Copying the winner exactly | the copies get grouped with it and go nowhere | keep its reason to buy; change who, where or how |
| Doubling the budget in one go | learning can restart and the cost per result can climb | raise in steps (TikTok: up to 50%, 2 days apart) |
| A long questionnaire before starting | the person leaves | read their site and memory first; carry on, and ask for what's missing in the same message |
| Promising sales, a cost per sale or a return | a promise nobody can keep | describe the test and what you'll look at, never the outcome |
| Posting or launching ads without their tap | money spent or a public post they didn't approve | posts and ad spend always go through their approval or their own ads manager |
| Inventing a number for them (budget, cost per sale, posts per week) | a plan built on a guess | use their numbers or the sourced ones in reference.md; otherwise describe the choice without a number |
| Model names or limits written into the plan | stale facts | `list_models` when a setting matters |
| Before-and-after or results claims for health, money or beauty services | ad rejected or account stopped | only what the business supplies, with consent |
---
Page: https://withicarus.com/skills/marketing-playbook
Download: https://icarus.b-cdn.net/skills/marketing-playbook/v1-5c9dd64f/marketing-playbook.zip
This skill makes photos and videos with Icarus. Connect Icarus to your AI app: https://withicarus.com/mcp