A chat answers. An agent gets it done
From a clear request to an agent that handles your routine work. You decide where it stops to wait for approval.
Try the first practice task and lesson as a guest. Before the next lesson, create a free account to keep learning with your saved work. The first steps take about 15 minutes.

A fluent report still needs evidence
This example comes from the “Spot the fabrication” trainer in lesson 2.9, “Verify before you publish.” Sáng Coffee is a fictional teaching brand. The numbers and the planted error below are exercise data, not a client result, an industry benchmark or a live model test.
The supplied weekly export
Analytics, last 7 days, Facebook
- Reach 12,400
- Engagement 3.1%
- Link clicks 212
- Best post: Morning ritual reel (reach 5,900)
A sentence the source cannot support
According to industry data, 68% of Vietnamese coffee buyers discover brands on Facebook.
The trainer deliberately adds this unsupported sentence. It sounds like research, but the supplied export contains no industry survey.
Trace each claim to its input
Reach 12,400, engagement 3.1% and the best post’s reach 5,900 are in the source. The 68% industry figure is absent. Mark that sentence as a made-up statistic and remove it; if the report needs an industry comparison, request a dated original source before adding one.
A report grounded in the export
Facebook reach was 12,400 in the last seven days. Engagement was 3.1%, and link clicks totaled 212. The morning ritual reel was the best post by reach, with 5,900. No industry-discovery figure was supplied.
This revision preserves the supported data and makes the missing evidence explicit. It does not turn reach or engagement into a claim about sales.
The first guest exercise is available before registration. In the course, you inspect the source, identify unsupported claims and explain the correction in your own words.
Practice the work you want to delegate


Sort email, messages and comments, then draft replies using the right facts and rules.


Write a useful brief, check drafts and prepare content for your publishing calendar.


Keep product details, photos, inventory and delivery terms consistent across listings.


Compare contracts, invoices and records. Trace numbers to their source, flag differences and leave the decision to a person.
Still exploring? Start with the core lessons. You can add or change a practice track later without losing your progress.
13 modules: from the first request to a team of agents
The exercises use Sáng Coffee, a fictional coffee business with a brand guide, a calendar, customer requests and business documents. The setting makes each task concrete. You can use the same method for your own business, channels and language.
Lift-off
9 lessons · 148 min
Anatomy of an agent
7 lessons · 120 min
Real agents in ChatGPT and Claude
9 lessons · 177 min
Build your own agent
9 lessons · 155 min


What the agent does, and what stays yours
Highlighted steps belong to the agent. The rest stays with you. The course helps you draw that line before handing over the work.
Learn a concept, then try it
Learn to give clear instructions, connect the right tools and decide when an AI agent needs your approval. Practice in a fictional business, then apply the method to your own tasks. 93 lessons and 65 lessons with a trainer, with no coding required.
A look inside the practice
A good draft still needs a check
Your request
Schedule the launch post for the new drip bags for tonight. Use the brief called launch-post.
What a suitable image needs
- Vertical format, 9:16
- The whole product in frame
- No added headline over the image
The agent reads the brief and checks the calendar before making drafts.
Three drafts from the exercise. A convincing image can still miss the brief.




In the practice calendar
One checked image, one scheduled post
The agent saves the third draft and uses it for the evening post.
Scheduled, with nothing published without approval.
Practice inside the course
Write instructions, inspect each action and check the result. 65 of the 93 lessons include a practice exercise. Recorded examples let you practice without connecting real accounts or paying for AI calls.
Check your work and request feedback
The exercises provide objective checks and prepared explanations. If you have a question, submit your work for teacher feedback. Your request and any replies stay together in Feedback; there is no guaranteed reply time.
Four things you keep
An agent you built
Design an agent for one of your own tasks, define its tools and approval points, then test the plan. Export the configuration and an optional code example.
A certificate with a code
Complete the required learning material to earn a certificate. A public verification code shows your name and course completion details.
Your own rules and tests
Keep your instructions, retry limits, test cases and design plan to use on your own tasks.
A one-page cheat sheet
Course terms, common agent failures, pipeline stages and safety rules in a printable reference.
From chat replies to useful workflows
You write a request, read what comes back, and fix it yourself. Everything happens in one window and nothing happens without you.
An AI agent reads the brief, uses the tools you allow and pauses at your approval points. You inspect its actions and change the instructions when needed.
Assign separate research tasks to several agents, combine their results and compare the effort with one agent doing the whole job.
Every lesson, in the open
Explore each module to see the lessons and practice tasks before you sign up.
00Lift-off
- 0.1Chat answers, agents actrequired10 min
- 0.2Watch an agent work, one step at a timerequired17 min
- 0.3Where agents already work in marketingrequired20 min
- 0.4Should this be an agent?required12 min
- 0.5Five lines instead of onerequired18 min
- 0.6Show it, do not describe itrequired18 min
- 0.7Ask for the calculation before the answerrequired18 min
- 0.8Ask for rows, not paragraphsrequired17 min
- 0.9Only from this documentrequired18 min
01Anatomy of an agent
- 1.1Think, act, observe: the loop from the insiderequired17 min
- 1.2Apps and actions: how an assistant uses toolsrequired17 min
- 1.3An app is only as good as its description18 min
- 1.4Working memory: what is on the desk right nowrequired17 min
- 1.5Stopping: done, stuck, or askrequired17 min
- 1.6Planning and subtasks: a five-post series18 min
- 1.7When apps fail: errors, retries, dead endsrequired16 min
02Real agents in ChatGPT and Claude
- 2.1The autonomy dial for content workrequired17 min
- 2.2Brief the assistant: five lines that prevent guessesrequired17 min
- 2.3Projects: the brand book on the desk before every chatrequired20 min
- 2.4Code execution in chat: batch work with a plan first20 min
- 2.5Playbooks: write the procedure once, use it every week21 min
- 2.6Browser agents: research first, approval before action22 min
- 2.7Connectors: one plug for Drive, Canva and the sheet21 min
- 2.8Scheduled routines and the glue that delivers them21 min
- 2.9Verify before you publish: four things the assistant will inventrequired18 min
03Build your own agent
- 3.1Scope and boundaries: the Content Producer's jobrequired17 min
- 3.2Instructions: voice, rules, facts, and when to stoprequired17 min
- 3.3Instruction traps and conflicts: which rule wins17 min
- 3.4Choose the apps: permissions, side effects, and safe retriesrequired17 min
- 3.5Documents in, actions out17 min
- 3.6The notebook: what the assistant remembers between sessions18 min
- 3.7Approval points: nothing goes live without yourequired18 min
- 3.8Assemble the Content Producerrequired17 min
- 3.9For the curious: export and run it for real17 min
04Shooting and post-production
- 4.1Before the shoot: brief to shot list18 min
- 4.2After the shoot: cull and tag18 min
- 4.3The cut: TikTok, Reels, Shopee20 min
- 4.4Versions: subtitles, sizes, covers16 min
- 4.5Livestream selling16 min
05Generation pipelines
- 5.1Generation services are apps like any other18 min
- 5.2The request you write is an input18 min
- 5.3The Creative Director loop19 min
- 5.4The asset library17 min
- 5.5AI content rules: four gates before anything goes outrequired19 min
- 5.6Effort and retries16 min
- 5.7The pipeline: seven stages, one gaterequired20 min
06Marketing ops agents
- 6.1The research agent: findings with sourcesrequired18 min
- 6.2Copy at scale: one caption assistant, eight testsrequired18 min
- 6.3Publishing and scheduling: the final step18 min
- 6.4The community inbox: triage, tone, escalationrequired18 min
- 6.5Creator outreach: shortlist, draft, and track replies17 min
- 6.6Analytics and reporting: every number has a sourcerequired18 min
- 6.7Experiments: two versions and an honest read17 min
07Teams and long campaigns
- 7.1When one assistant is not enough18 min
- 7.2Work in parallel, then combine the results17 min
- 7.3Three ways a team goes wrong18 min
- 7.4Long campaigns: summarize, hand off, and resume18 min
- 7.5Compare a team with a solo assistant18 min
08Reliability, safety, effort
- 8.1Seven common ways an assistant can failrequired18 min
- 8.2Reading a replay like a detective20 min
- 8.3Poisoned content: what the attack looks like16 min
- 8.4Poisoned content: rules and permissions that limit the riskrequired20 min
- 8.5Client data, secrets, brand safety, platform rulesrequired20 min
- 8.6The test set for copy and creative18 min
- 8.7Effort: what makes a run heavyrequired18 min
- 8.8See what it did: the log, the limits, the stop button17 min
09Your agent
- 9.1The design page: define the job before buildingrequired20 min
- 9.2Build it, then break it on purposerequired22 min
- 9.3Three reviews: automated checks, a teacher, and a peerrequired18 min
10Requests, meetings and agreements
- 10.1Handle incoming requests without losing track18 min
- 10.2Different messages, the same question20 min
- 10.3Answer from the rules, and name the clause20 min
- 10.4Two meeting requests, one available slot20 min
- 10.5Ask for a delivery date you do not have20 min
- 10.6Turn an unresolved request into a task18 min
- 10.7One morning, end to end25 min
- 10.8Your own week of requests22 min
11Product listings and catalogs
- 11.1Product listings: facts first, copy second18 min
- 11.2Every field has an owner20 min
- 11.3Leave missing product data blank20 min
- 11.4Two codes, almost one name20 min
- 11.5The sentence you do not repeat20 min
- 11.6Choose a photo of the actual product18 min
- 11.7One shelf, end to end25 min
- 11.8Apply the method to your own catalog22 min
12Checking business documents
- 12.1Compare the right documents18 min
- 12.2Which differences count20 min
- 12.3Do not invent a missing document date18 min
- 12.4Escalate a document you cannot approve20 min
- 12.5Request a correction instead of changing the source20 min
- 12.6Log each incoming document once18 min
- 12.7One folder, end to end25 min
- 12.8Apply the method to your own records22 min
Four questions, then your route
Choose your tasks, goal, experience and chat app. See how many lessons are on your route, their estimated length and which modules you can skip with a knowledge check.
1 / 4
Which tasks take up most of your time?
Choose the work you want to practice. Your job title does not matter.
What an account gives you
- Every lesson and every trainer in the course
- The cheat sheet and every word the course uses
- Saved feedback requests, submitted work and replies
- A certificate with a code anyone can check
- English, Russian and Vietnamese, with a language switch
- Works on a phone, and remembers where you stopped


Common questions
Do I need to know how to program?
No coding is required. Optional technical views and code exports are available if you want to explore further. The glossary explains unfamiliar terms.
How long does it take?
About 29 hours of reading and practice, at your own pace. There are no scheduled classes.
What do I need to have?
A browser is enough for the recorded exercises. To try the optional tasks in ChatGPT, Claude or another service, use your own account; features and costs depend on that service.
Will the practicum fit my work?
The practicum covers content, online retail, messages and work processes. The studio case gives you concrete examples; you can apply the approach to your own work in any country.
What is the certificate?
It records completion of the required learning material. The public verification page shows your name, completion details and issue date. Practical results are recorded separately.
Can I open an account myself?
Try the first practice task and lesson as a guest. Before the next lesson, create a free account to keep learning with your saved work. The first steps take about 15 minutes.
I already ask a chat for text every day. Is there anything here for me?
Yes. Short knowledge checks can mark modules 0 and 1 as already known. Later lessons focus on tool use, approval points, reliable instructions and testing an agent’s work.