Krste Kovaceski just posted a long-form playbook on X about running a direct-response marketing team as a fleet of Grok Bots. His framing is the cleanest I've seen on the operator side, and worth translating for anyone who thinks "AI agents for marketing" still means a Zapier flow with a prompt bolted on.
This is my read of his playbook: the build order, the rules that matter most, and the guardrails I'd put in before any bot touches an ad account.
If you want the architectural view of the same idea at the full engineering end, that's in the marketing agent architecture I'd actually build in 2026. This post is the lightweight cousin: no warehouse, no code, just a team of narrow chat bots and a shared skills library.
The core idea: skill is how, bot is who
Every bot in the team reads from a shared library of skills. A skill is a written playbook for one job (write an advertorial, launch a Meta campaign, caption a video). A bot is a teammate with a persona, memory and routines that must read its skill before it does any work.
Write the skill first, then give it a teammate who lives by it. A bot without a skill is a persona with no method, and you'll feel it the first time you ask it to do anything twice.
That's the entire mental model. Everything else below is just the order to add things in.
Start with one or two bots, not a team chart
The temptation is to sketch out the whole team on day one. Resist it. Build one bot for one job you already do by hand, connect the one or two data sources you touch every day, and talk to it for a week.
Good first bots:
- A niche ad researcher that pulls competitor angles and creatives for one niche (not "all of DTC"). One niche or it's useless.
- A Copywriting Bot that turns research briefs into ad copy in your voice.
Connect data early. In Grok Bot, connectors install once at the account level and every bot can use them. Starting map for a DR stack:
- GetHookd for competitor ad intel across research, copy and image bots
- AdPlexity Native if you run native
- Minea for product research
- The Meta ads MCP for buying
- Shopify + AIDesigner for landers
- TokPortal for TikTok and Instagram publishing
- An image model API key for creative generation
You don't need all of them. You need the two you use every day.
Build the Skill Builder before any specialist
This is the step that's easiest to skip. Skip it and every bot improvises its own method, so the fleet drifts.
Create a bot whose only job is writing skills and spawning the specialists that read them. Call it Skill Builder. In its first message, give it the method:
- A skill is a playbook saved to the shared library. Generic. No channel names, no account IDs, no brand constants.
- A bot is a teammate with persona, memory and routines. It must read its skill before acting.
- Never create a bot without a skill. A bot with no skill is a persona with no method.
- Name skills after the action ("Writing advertorials"). Name bots after the role ("Copywriting Bot").
- Every skill description starts with "Use when…" and lists the trigger phrases.
Build the Skill Builder before any other specialist. Every bot you add after it inherits the method for free.
Turn proven workflows into skills, one at a time
Pick a workflow you've already run successfully three or four times by hand. Hand it to Skill Builder with inputs, steps, rules and output format. The high-value DR skills to write first:
- Static ad images. Copy the packaging exactly, pin one image model, exact on-image text in quotes.
- Video ad scripts. 25 hooks → top 10 → short educational body → soft close. Grade-4 reading level, literal language.
- Advertorial copy. Angle → outline → draft → review pass. Every section stands alone, no fake publishers.
- Media buying via the Meta MCP. Confirm account, page, pixel, URL, budget and creatives by name. Launch paused. Kill rules in bot memory.
- Shopify PDP builder. Section map, callouts, testimonials. One style per page.
- Captioning and export. Trim dead space, burn captions in the safe zone, pass a final format check before export.
- Product research. Filter scaling stores and ads, group by product, shortlist with reasons.
If a skill gets long, split it. A page-builder skill can call a separate copy skill and image skill, not do both itself.
One specialist per job
Once a skill exists, have the Skill Builder create a bot to own it. One job per bot. If the persona has an "and also" in it, that's a second bot.
A full DR team ends up looking roughly like:
- Video Script Builder, Image Generation, KIE Video Ad Builder, Copywriting Bot
- GetHookd Research for swipes and briefs
- Shopify Developer for landers and pages
- Media Buying for Meta
- Native Ads Bot for NewsBreak, MediaGo, SmartNews
- Utility bots: a transcriber, TokPortal
Each persona says who the bot is, which skill it reads, what "great" looks like in its area, and what it saves to memory.
Brand facts in bot memory, not in skills
This is the one that trips everyone up. Account IDs, pages, pixels, prices, claim limits, kill rules: those live in the brand's strategy bot memory, not inside a skill.
Why: the Image Generation skill is generic. If you bake brand constants into it, it serves one brand. If you keep brand constants in memory, one Image Generation bot can stay product-locked for five brands at once. Same for Copywriting, same for Media Buying.
A strategy bot per product holds the brand facts. A product team chat routes the specialists.
Let the bots hand work to each other
Once you have a product team chat, bots start passing work along without you narrating every step:
- Research bots post swipe digests
- Creative bots pick them up and brief Image Generation
- Shopify Developer tells Media Buying when the lander URL is live
- Media Buying keeps a shared test log the strategy bot reads
A primary Grok Bot sits at the top and routes any ask to the right specialist. Ask it for the outcome you want, not the steps.
Routines are what turn this into a team
A single bot with a skill is useful. A team of bots with routines is a company.
Set up one daily routine first (a research digest). Add the next only after the first runs clean for a week. A product team's morning chain, in order:
- GetHookd Research posts the day's swipe digest
- Copywriting Bot writes the day's copy
- KIE Video Ad Builder ships video packs
- Image Generation ships static packs
- Media Buying launches the day's packs
Every step feeds the next. The buyer walks into a desk full of fresh, ready inputs.
When a bot earns the team seat, publish it
Most bots stay personal. When one bot does a job the whole team uses (think a Product Research bot posting daily Minea picks), publish it as a Team Bot.
Three things to get right when you publish:
- Scoped, read-only secrets. Teammates never see the key value, but the bot can use it on anyone's behalf.
- Decide what to copy: chat and memories, or a fresh start. If the bot's history is mostly you, start fresh.
- Routines are personal. They run as the person who set them up and report in that person's chat. To report to the team, end the routine prompt by telling it exactly which shared Slack channel to post to.
Then invite the bot to the channels where it should work. DMs always get answered. In channels it answers on @mention and follows the thread.
The guardrails that keep this from eating your ad budget
A few rules I'd carve into every DR team's skills, non-negotiable:
- Launch paused. Media Buying never launches a live campaign. The human flips the switch.
- Never guess IDs. Account, page, pixel, URL, budget, creatives: all confirmed by name before launch.
- Lock the product in images. Copy packaging exactly from the reference photo. Never redesign the logo. Pin the image model in the skill.
- Read every line of copy yourself before it ships. AI drafts fast enough that reading is now the bottleneck. Don't bypass it.
- Save fixes back into skills. New house rule ("no metaphors in copy") goes into every copy skill at once. Fix the skill once, every bot gets it.
- No secrets in chat. API keys as secrets. Skills never print them.
Starter checklist
- Connect your two most-used data sources
- One narrow bot for a job you already do by hand
- A Skill Builder bot with the method
- One skill from a workflow you've already proven
- One specialist bot from that skill
- Repeat for research, scripts, images, video, copy, landers, buying
- Brand facts in bot memory, never in skills
- Team chat per product + a primary bot that routes
- One daily routine first, then build the morning chain
- Publish team-wide bots to Slack, route reports to a shared channel
- After every fix, update the skill so the whole team gets it
The full version of Krste's playbook has more detail on each piece, especially the Team Bots setup. Read it on X. This post is the condensed operator's sequencing. If you want the architectural end of the same idea with a real data warehouse and a feedback loop, the marketing agent architecture post is the companion.