One Person, Four AI Roles: How 7 Marketing Skills Powered a 41M-View Workflow

Core Experiment: 41M Views, 7 Skills, 4 Roles

A solo creator disclosed a full experiment: a content method that yielded over 41 million views in 30 days was broken into 7 reusable marketing Skills, then handed to 4 long-lived AI roles to execute a real social-media carousel planning, creation, review, and scheduling pipeline. The 41M figure comes from the author's own disclosure; no third-party audit is provided, so it should be treated as a workflow case study, not a guaranteed result.

Planning → Writing → Independent Review → Rewrite → Human Copy Approval → Publish Check → Human Final Confirmation → Scheduling

Why 4 Roles Instead of 1 or 7

The author compared a single generalist bot versus specialized bots. A generalist bot becomes just another chat window — tasks change daily, context doesn't accumulate. Specialized roles handle the same output type repeatedly, so corrections and context concentrate. But 7 roles (one per Skill) would over-fragment; Skills are capabilities, roles are responsibilities. One role can invoke multiple Skills if they serve the same outcome.

Final Role–Skill Allocation

Content Planner Bot — owns brand context, strategy, topics, approved briefs — uses brand-brief, content-coach Content Creator Bot — produces drafts, hooks, cross-platform versions — uses post-writer, viral-hooks, repurpose Content Reviewer Bot — independent scoring, flags issues, gatekeeps quality — uses post-grader Content Publisher Bot — final checks, schedules approved content — uses post-scheduler Writing and review must be separate; letting the Creator grade its own drafts conflates intent with quality. The Reviewer actually rejected drafts twice in the live test. Publishing is its own role because it triggers external actions; it only receives approved content and pauses at a human confirmation step.

The 7 Marketing Skills (Markdown-defined, Portable)

1. brand-brief: Brand Fact Source

Asks six categories: product/service, target buyer, single CTA, contrarian industry view, recent real story/result/mistake, expression style. Saved as a shared brief read by all downstream Skills — solves "re-introduce the company every chat" and anchors facts, tone, stance.

2. content-coach: From Blank Page to Writable Topics

When the user says "I don't know what to post", it checks the brand brief and proposes 5 concrete angles: contrarian view, numeric result/evidence, real failure/vulnerable moment, customer before/after, common mistake. Each topic explains the emotional trigger and why people would comment/save/share. User picks one; start with one platform, get one loop smooth before multi-platform sync.

3. post-writer: Platform-Native Drafts

Reads brief, confirms topic and target platform, writes hook-body-CTA. Different platforms have different length limits, first-screen constraints, media requirements, engagement logic — so the same insight must be natively rewritten, not copy-pasted. Hook polished first: at least 3 variants, tested by "do the first three words make you keep reading?" Only one core point, one CTA per post.

4. viral-hooks: 100 Hook Frameworks

13 categories, 100 frameworks (specific result, counter-intuitive, negative angle, borrowed experiment, curiosity gap, vulnerable confession, etc.). Picks framework per topic/platform, fills with real details from brand brief/source material, generates 3 variants, selects strongest. Provides structure only — facts, numbers, cases, results must come from brand brief or source material.

5. post-grader: Independent Scoring, No Favoritism

Scores drafts on 7 dimensions: hook strength, curiosity & specificity, emotional tension, share value, brand voice, stance clarity, platform fit. Hook = 50% of total. Below 8/10 → Reviewer lists top 3 issues: where, why it hurts distribution, exact fix. Author's rule: 7 = decent, 8 = strong, 10 basically doesn't exist.

6. repurpose: One Long Asset, Many Native Outputs

Extracts one central argument, 3–7 supporting points, all numbers, stories, quotes, contrarian takes from blog, newsletter, video transcript, or script. Rewrites per platform. Open-source default yield: 3 professional long posts, 5 short-thread sets, 2 short-video scripts. If source material can't support 10 angles, reduce count — no fluff padding.

7. post-scheduler: Final Gate Before Publish

Checks target account, time, timezone, media order, preview, alt text, platform limits. Schedules via an automation tool (or saves as copy-paste file if tool not configured). Hard rule: no publish without explicit human approval .

Verified against the open-source repo: the 7 Skills' responsibilities match the author's description. A later repo update added a faceless-video module; not part of the original "4 roles, 7 skills" experiment.

Human-in-the-Loop: Where the Person Stays

A broader role map shows what AI can own vs. what the human must keep:

Chief of Staff — AI: prioritize, dispatch, track handoffs; Human: final priority & trade-offs

Marketing — AI: research, content, adaptation, distribution; Human: taste, factual claims, final approval

Sales — AI: prospect research, outreach drafts, CRM prep; Human: relationship & actual send

Support — AI: ticket triage, reply drafts, FAQ; Human: sensitive/exception cases

Operations — AI: repetitive admin, monitoring, weekly reports; Human: exceptions & decisions

Finance — AI: receipt sorting, invoice tracking, report generation; Human: payments & financial sign-off

Start with one role, end-to-end owning one verifiable outcome. Add the next role only when a second major workload appears.

First Role Must Meet 5 Criteria

Recurring — you do it every week

Stable steps — process doesn't change daily

Multi-step — handing it off saves real time

Reversible — mistakes can be fixed, no instant customer harm or money loss

Measurable — you can clearly define "done well" and "done"

"Weekly content brief" fits: Content Brief Bot checks sources, collects ideas, ranks topics, adds research, hands brief to Writer Bot. Human judges in 5 minutes. Another example: daily competitor pricing page monitor — check, log old/new price, source link, screenshot, send to human; no contact, no changes. Safety test: "If it screws up tonight, can I know by morning and undo it?" If no, shrink the role further.

Hiring a Bot: 5 Required Pieces

AI environment supporting long tasks, tool use, handoffs

One real weekly task (≥30 min, clear steps)

Skill files written in Markdown

Actual tools/plugins/web pages the work needs

Explicit human-approval boundaries

Random prompts don't build reliable teammates. Each role must know: the single result it owns, definition of "good", allowed tools, when to stop and ask. "Help me with marketing" is too wide. Executable version:

Per brand brief and performance data, plan next week's content. Give Writer role each piece's topic, hook, source links, bullet points. Do not publish. Do not fabricate results.

The article provides a reusable Bot Job Description template (see source). Rules go in the job description; day-specific tasks go in the chat. Don't stuff the whole org chart into one giant spec — focused roles accumulate same-type corrections into stable capability.

Tool Access: Convenience and Risk from Same Source

When structured plugins exist, grant scoped tools. When not, let the bot open a cloud browser; human takes over at login page, enters password and 2FA, then hands back control. Never paste passwords or OTPs into chat. Multiple bots under one account may share cloud computer, files, browser sessions. Publisher logs into social tool — others could theoretically reuse that session. Therefore, permissions start minimal:

Planner: only brand & product docs

Creator: only approved briefs

Reviewer: only drafts + scoring rules

Publisher: only publishing accounts

Publish, send, spend, delete, change permissions, edit production systems — always after human approval

Demo used a specific social automation platform; product-specific config (API keys, auth, backend links) omitted per policy. General principles: connect per platform's official docs; human enters keys and scopes permissions.

Importing Skills: Keep Rules, Drop Only Incompatible Tool Names

The 7 Skills were written for another AI tool. On import, the author only removed unsupported allowed-tools fields — no compression or rewrite of the Skills themselves. Import prompt used (generalized):

Import all 7 Skills from the specified open-source repo.
Preserve original instructions, decision rules, output formats, validation steps, approval boundaries.
Only remove tool names that don't exist in the current AI environment.
Tell me each modification you made.
Do not run the Skills yet.

All roles see shared Skills, but visibility ≠ authorization. Publisher doesn't need hook library; Creator must not have publish rights. Capability visibility and role authorization must be separated.

Putting 4 Roles in One Collaboration Channel

Author created an "Applied Marketing Team" channel, added the 4 bots, then used an orchestration prompt to establish handoff order (vendor names removed):

Content Planner: use content-coach Skill.
Adapt it into this 4-role channel's playbook.
Keep beginner entry, brand questions, 5-topic brainstorm, platform suggestion, pre-publish explicit-approval rule.
Change single-agent execution to visible specialist handoffs:
- Planner owns brand context, strategy, topic selection, approved brief.
- Creator uses post-writer, viral-hooks, or repurpose to produce drafts.
- Reviewer independently uses post-grader, gives revision notes, only it can approve quality.
- Publisher may use post-scheduler ONLY after Reviewer passes AND I explicitly approve.
Global shared Skills must not erase role boundaries. No bot may execute another role's step just because it sees the Skill.
Do not run workflow, do not connect publishing tool now.
Report all compatibility and orchestration changes, explain how handoffs will appear in this channel.

After this, human only appears at decisions affecting customers or cost; daily prep runs autonomously among roles.

Live Test: Produce a 3-Page Carousel

Chose a real app as case, connected a publishing tool via its official MCP/API docs, gave the team:

Using your Skills, plan a 3-page social carousel for _______.
Iterate until score ≥ 8/10.
Then generate carousel with approved design tool.

Actual steps executed:

Planner locked angle

Creator wrote captions & per-slide copy

Reviewer independently checked facts, tone, structure, CTA

Reviewer scored 7/10, rejected draft, gave revision notes

Creator rewrote per notes

Human approved final carousel

Publisher checked account, asset order, copy, preview

Human gave final publish confirmation

Final scheduling command: Schedule it for tomorrow 8 AM on target social. Publisher invoked post-scheduler, placed content in next-day queue.

What Worked vs. What's Still Clunky

Worked

7 Skills transferred almost unchanged; only incompatible tool fields needed handling

Specialist roles made ownership & accountability clear

Reviewer challenged Creator twice — no rubber-stamp

Shared channel preserved handoff trail & approval boundaries

Shared browser session let Publisher drive real tools

Approval gates paused at alt-text, timezone, etc.

Approved carousel successfully entered publish queue

Clunky

"Skill" and "Plugin" concepts overlap in UI, causing confusion

AI image/video generation takes time; bot must poll repeatedly for completion

Whole bot capability still early — only suitable for reversible, low-risk work first

Multi-role doesn't eliminate waiting; it turns "human refreshing page" into "system polling status". Without timeouts, retries, failure reporting, automation can still stall mid-way.

After It Runs Smooth, Turn Tasks into Routines

Sequence: pick one weekly ≥30-min task with clear steps & clear done criteria → create one role → write verified process as Skill → iterate until repeatable success → then convert to scheduled recurrence. Example routines:

Daily morning: Research bot gathers relevant product launches & audience questions

Monday: Planner hands topics, hooks, source links to Creator

Friday: Reviewer audits next week's scheduled marketing content

Before each sales call: Sales bot prepares prospect research

Month-end: Finance bot organizes receipts, flags missing items

"Runs after laptop shut" is just runtime capability. Long-term viability depends on: fact source, done criteria, handoff format, failure reporting, approval points.

11-Step Checklist (Reconstructed from Source)

Pick a visibly time-consuming job (author chose marketing)

Create 1 bot, give it 1 clear role

Provide fact source, "good" definition, handoff format, approval boundaries

Connect the actual tools the job needs

Test with one real task that has a clear endpoint

Iterate corrections until process repeats successfully

Save verified process as a Skill

Only when a second major job appears, add a second bot

When collaboration needed, put relevant bots in same group/channel

After trusting output, convert to recurring scheduled task

Publish, send, spend, delete, production edits — always after human approval

You don't need an AI employee for every box on an org chart. Start with the task you've been avoiding, give it a clear owner, stabilize it, then "hire" the next.

Author's Verdict: The Real Open-Source Asset Is the Management Method

"41M views" and "4-bot army" grab attention, but the portable assets are five plain management principles:

Role by outcome, Skill by capability. 7 Skills didn't become 7 bots because role count follows responsibility boundaries. Official Agent orchestration docs likewise list "specialist handoff" and "manager calls experts" as common patterns; write-evaluate-rewrite loop is a classic evaluator-optimizer workflow.

Handoffs must carry structure. Deliverable, sources, uncertainties, next action — all four required. Without structured handoff, multi-bot just moves copy-paste from human hands to chat logs.

Independent review beats self-evaluation. Reviewer can say "no", score below threshold triggers return with actionable fixes. It still can't verify facts — only surface expression/structure issues; key numbers/claims must trace back to sources.

Shared login state is a double-edged sword. Enables cross-tool flow but expands blast radius. Each role needs least privilege; critical ops need pause-approve-resume with logs. Current Agent engineering docs already model human approval as interruptible, rejectable, resumable execution state — not a vague "be careful".

Traffic tactics ≠ product & judgment. Hook frameworks, scorers, auto-scheduling boost content throughput but can't guarantee the insight is valuable, nor that 41M views replicate in another account/industry/time window. Human taste, lived experience, customer relationships, business judgment remain the company.

A solo company doesn't get "zero-human ops". It gets a reusable execution layer: human keeps time for non-outsourceable judgment; AI repeats the already-clarified work reliably.

Original: https://x.com/Sabrina_Ramonov/status/2094833331183759612

Source Verification

7 Skill names & responsibilities from author's open-source repo snapshot; original links kept in internal verification, not shown publicly.

Multi-role handoff & two common orchestration patterns: see OpenAI Agents SDK — Agent orchestration.

Sensitive tool call pause/approve/reject/resume: see OpenAI Agents SDK — Human-in-the-loop.

Human control, safe interaction, transparency, privacy: see Anthropic — Trustworthy agents in practice & NIST Generative AI Risk Management Framework.