AI at work: from the first request to your own agent

A chat answers. An agent gets it done

The same ChatGPT or Claude, but with your inbox, calendar and spreadsheets in reach, and your stop switch before anything leaves the room. First you learn to get what you need on the first request. Then you build your own agent and break it against your own tests until it holds. 93 lessons, 65 lessons with a trainer, no code anywhere.

The course is not open to everyone yet: accounts are set up one at a time.

93lessons
65lessons with a trainer
13modules
29hours
StartAny day. Four questions decide your route.
LengthAbout 29 hours, at your own pace
FormatRead a little, run the trainer, send it for review
LevelNo code is needed in any lesson
// WHOSE_ROUTINE_IS_THIS

The routine the course takes apart, step by step

01
Incoming from people you do not know

Mail, messages, comments. The morning goes on reading all of them before you can answer anyone.

02
Shoots, copy, publishing

The same request rewritten a fifth time until something usable comes back. And a calendar that will not fill itself.

03
Listings, stock and orders

The same fields in dozens of listings. Photos, stock and delivery terms that need to match what the customer sees.

04
Documents, checks and rules

An agreement, an invoice and a register that need to agree. Find the source of each number, flag differences and leave the decision to a person.

None of these yours? The core of the course is the same for everyone and runs on the studio's own tasks. A track is added on top, and you pick it on the map once you know which one is yours.

// WHAT_IS_INSIDE

13 modules: from the first request to a team of agents

Everything happens at Sáng Coffee: a small coffee business with a brand book, a calendar, customers and papers that never quite agree. Every module adds one more piece of the work, and by the end the agent walks it through on its own, under your eye. The coffee shop is made up. Your tasks are not.

00

Lift-off

9 lessons · 148 min

01

Anatomy of an agent

7 lessons · 120 min

02

Real agents in ChatGPT and Claude

9 lessons · 177 min

03

Build your own agent

9 lessons · 155 min

0412

and 9 more modules

The whole programme, lesson by lesson, is below.

Open the programme
// WHO_DOES_WHAT

What the agent does, and what stays yours

Coral is the agent. The rest is you. Most of the course is about drawing that line on purpose instead of finding out where it fell.

Reads the source and the rulesagent
Writes the first draftsagentyou
Checks facts and numbersagentyou
Decides what reaches a customeryou
Says the work is finishedyou
// HOW_IT_RUNS

Read little, do it straight away

Practice, not slides

Most lessons end in something you run yourself: you write the instruction, watch the steps go by, and find the one where it went wrong. There are trainers in 65 of the 93 lessons, and they run inside the course.

A person reads your run

Send any run for review. A person opens the same replay you saw, step for step, and answers in your Reviews. Before that the course gives you its own reading of the run, so you are never stuck waiting.

// WHAT_YOU_WALK_OUT_WITH

Four things you keep

01

An agent you built

In module 9 you design it on one page, assemble it, then break it on purpose against eight tests you wrote for your own tasks. Two export files leave with you.

02

A certificate with a code

Issued from what you actually passed, never on request. Anyone can check the code on a public page, including a client.

03

Your own rules and tests

The judging rules, the retry rule, the test set, the four boxes of a design page. You leave with them filled in for your own tasks, not for the coffee shop.

04

A one-page cheat sheet

The glossary, the seven ways a hire breaks, the seven stages and the seven rules, on one page you can print.

// WHERE_THIS_SITS

Why this course now

Before
You write a request and fix the answer by hand

You write a request, read what comes back, and fix it yourself. Everything happens in one window and nothing happens without you.

You are here
Hand over the routine, keep the stop switch

The agent, the new hire at the desk beside you, reads the brief, opens the apps and acts, and stops before anything reaches a person. You read the replay and say what to change.

After
A team of agents on one process

Several hires on the sources that do not wait for each other, one that puts the answers together, and a table that says whether the team was worth it.

// WHO_WROTE_IT
Danya

Written by someone who does this work

The course comes out of everyday work with these tools, not out of theory. Every run you send is read by a person, and the answer comes back in your Reviews.

AccountsSet up on request
ReviewsEvery run you send gets a person's answer
LanguagesEnglish, Vietnamese, Russian
CertificateOnly when the required lessons are done
// THE_PROGRAMME

Every lesson, in the open

Nothing here is hidden until you are inside. Open a module and read what it holds.

13 modules · 93 lessons · 65 lessons with a trainer
00Lift-off9 lessons · 148 min

  • 0.1Chat answers, agents actrequired10 min
  • 0.2The loop, seen oncerequired17 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
  • 0.7Make it work before it answersrequired18 min
  • 0.8Ask for rows, not paragraphsrequired17 min
  • 0.9Only from this documentrequired18 min
01Anatomy of an agent7 lessons · 120 min

  • 1.1Think, act, observe: the loop from the insiderequired17 min
  • 1.2Apps and actions: how the hire touches the studiorequired17 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 sub-tasks: a five-post series18 min
  • 1.7When apps fail: errors, retries, dead endsrequired16 min
02Real agents in ChatGPT and Claude9 lessons · 177 min

  • 2.1The autonomy dial for content workrequired17 min
  • 2.2Briefing the hire: five lines that replace a guessrequired17 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 runs free, publishing waits for you22 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 hire will inventrequired18 min
03Build your own agent9 lessons · 155 min

  • 3.1Scope and boundaries: the Content Producer's jobrequired17 min
  • 3.2The instructions: voice, rules, facts and stopsrequired17 min
  • 3.3Instruction traps and conflicts: which rule wins17 min
  • 3.4Choosing the apps: boundaries, side effects, safe to repeatrequired17 min
  • 3.5Documents in, actions out17 min
  • 3.6The notebook: what the hire 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-production5 lessons · 88 min

  • 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 pipelines7 lessons · 127 min

  • 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 agents7 lessons · 124 min

  • 6.1The research agent: findings with sourcesrequired18 min
  • 6.2Copy at scale: one caption hire, eight testsrequired18 min
  • 6.3Publishing and scheduling: the last metre18 min
  • 6.4The community inbox: triage, tone, escalationrequired18 min
  • 6.5KOC outreach: shortlist, message, reply tracked17 min
  • 6.6Analytics and reporting: every number has a sourcerequired18 min
  • 6.7Experiments: two versions and an honest read17 min
07Teams and long campaigns5 lessons · 89 min

  • 7.1When one hire is not enough18 min
  • 7.2Side by side, then merged17 min
  • 7.3Three ways a team goes wrong18 min
  • 7.4Long campaigns: summarise, hand over, come back18 min
  • 7.5Team against solo, measured18 min
08Reliability, safety, effort8 lessons · 147 min

  • 8.1Seven ways a hire breaksrequired18 min
  • 8.2Reading a replay like a detective20 min
  • 8.3Poisoned content: what the attack looks like16 min
  • 8.4Poisoned content: the rules that stop itrequired20 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 agent3 lessons · 60 min

  • 9.1The design page: one page before any buildingrequired20 min
  • 9.2Build it, then break it on purposerequired22 min
  • 9.3Three readings: the reviewer, the teacher, your peerrequired18 min
10Requests, meetings and agreements8 lessons · 163 min

  • 10.1The week that runs on other people's requests18 min
  • 10.2Two letters, one question20 min
  • 10.3Answer from the rules, and name the clause20 min
  • 10.4One hour, two people20 min
  • 10.5The date you do not have20 min
  • 10.6The request that is not an answer18 min
  • 10.7One morning, end to end25 min
  • 10.8Your own week of requests22 min
11Listings and catalogues8 lessons · 163 min

  • 11.1The desk where things get described18 min
  • 11.2Every field has an owner20 min
  • 11.3The gap in the row20 min
  • 11.4Two codes, almost one name20 min
  • 11.5The sentence you do not repeat20 min
  • 11.6The picture of this thing18 min
  • 11.7One shelf, end to end25 min
  • 11.8Your own catalogue22 min
12Papers that have to agree8 lessons · 161 min

  • 12.1The room where papers have to agree18 min
  • 12.2Which differences count20 min
  • 12.3The paper with no date18 min
  • 12.4The line you do not sign20 min
  • 12.5The correction you do not make20 min
  • 12.6The register of what came in18 min
  • 12.7One folder, end to end25 min
  • 12.8Your own records room22 min
// FOUR_QUESTIONS_THEN_YOUR_ROUTE

Four questions, then your route

The same four the course asks at the door. The answer is a count: how many lessons are on your route, how long they take, and which ones a short test can skip.

Question 1 of 4

Which routine eats the most of your time?

Pick the work, not the job title. One answer covers several professions with the same routine.

// HOW_TO_GET_IN

What an account gives you

The whole course
  • Every lesson and every trainer in the course
  • The cheat sheet and every word the course uses
  • Every run you send is read by a person, with an answer in your Reviews
  • A certificate with a code anyone can check
  • English, Vietnamese and Russian, switchable mid-lesson
  • Works on a phone, and remembers where you stopped
Request access
// WHAT_PEOPLE_ASK

Asked before you ask

Do I need to know how to program?

No. Nothing in the course asks you to write code. The words developers use appear only inside the glossary, so that you recognise them when someone else says them.

How long does it take?

About 29 hours of reading and practice, spread however suits you. Nothing expires and nothing is scheduled.

What do I need to have?

A chat product you already use, and an hour a week. The trainers run inside the course itself, so you can practise before you touch anything of your own.

Is this only for the Vietnamese market?

No. The studio in the story is Vietnamese because a particular place makes the work concrete. The lessons are about the work, and the work is the same anywhere.

What is the certificate?

A page with your name, what you finished, and a code. Anyone you hand the code to can open the page and see that it is genuine. It is issued only when every required lesson is passed, and nobody can hand it out on request.

Can I open an account myself?

Not yet: accounts are set up one at a time. Answer the four questions above, leave an address, and the account opens on your route.

I already ask a chat for text every day. Is there anything here for me?

Modules 0 and 1 have a way round: five questions at the start can mark module 0 as known, and five more module 1, five out of five each time. Past them the course begins where the chat window ends. The agent opens the apps and acts, and the work is deciding what it may do without you.

The first lesson is ten minutes, with nothing to prepare

By lesson 0.4 you are deciding which of your own tasks should be an agent at all, and which are faster to keep doing yourself.

Request access