The typical way to get acquainted with Claude Cowork is this: open a folder, write a request, get a mediocre result, and decide it's just the same chat but with file access. Usually the issue isn't the tool or how you phrase the request. Cowork enters a folder without context: it doesn't know your tone, your requirements, or what it can't guess at. So it improvises.

This is solved with a small system of files. No code in it.

The main idea: context lives in files

In a chatbot, answer quality depends on the request. Cowork works with everything in the folder available to it: project materials, rules, drafts, examples of good results. If you describe once in files how you work, you won't have to repeat it in every request.

Step 1. Installation and global instructions

Cowork is part of the Claude desktop application. Download it from the official page, sign into your account, and select Cowork in the sidebar.

Then open Cowork settings and find the section with global instructions. They apply to every new session. You should record the basic behavior there:

Ты работаешь вместе со мной в этой папке.

До начала любой задачи прочитай:
1. context/HOW-I-WORK.md
2. system/SYSTEM-RULES.md
3. папку нужного проекта в projects/

Если не ясны цель, аудитория или формат результата, сначала спроси.
Не придумывай факты, цифры, даты и договорённости.
Когда не уверен, скажи об этом прямо.

Главный источник данных — файлы в этой папке.
Готовые результаты сохраняй в outputs/.

Step 2. Four folders

workspace/
├── system/
├── context/
├── projects/
└── outputs/
  • system/ — Cowork rules. Change rarely.
  • context/ — information about you and your work. Change rarely as well.
  • projects/ — current tasks, one folder per task.
  • outputs/ — everything Cowork creates.

Don't start with ten folders and a large template. This structure is enough to get going, and it expands without clutter.

Step 3. Two files

context/HOW-I-WORK.md describes you:

  • what you do and for whom;
  • how you approach tasks;
  • how you write, what sounds like you and what doesn't;
  • what irritates you in others' work;
  • which sources you trust when there are contradictions;
  • examples of your texts or results.

system/SYSTEM-RULES.md describes Cowork's behavior:

  • when to ask questions before starting work;
  • how to proceed with incomplete or contradictory data;
  • what counts as a strong result and what counts as weak;
  • hard prohibitions;
  • verification before submission.

Step 4. Don't write these files yourself

When describing yourself from scratch, a person describes who they'd like to be. Cowork gets an embellished portrait that later clashes with reality. An interview is more reliable: Cowork asks questions one by one, you answer, and it assembles both files from your answers.

Here's the request for such an interview:

Проведи со мной интервью, чтобы создать два файла:
1. context/HOW-I-WORK.md
2. system/SYSTEM-RULES.md

Три блока вопросов:

A. Работа. Чем я занимаюсь, что создаю, для кого,
   что ценит моя аудитория, каким источникам я доверяю,
   в чём эта папка должна мне помогать.

B. Стиль и мышление. Как я пишу в лучшей форме и на автопилоте,
   какие слова для меня естественны, что я не выношу в текстах,
   как я разбираю сложные задачи. Попроси реальные отрывки моих текстов.

C. Правила для тебя. Когда спрашивать до начала работы,
   что нельзя додумывать, как поступать при противоречиях,
   какая помощь мне полезна, а какая мешает.

Как вести интервью:
- Один вопрос за раз, жди ответа.
- Не больше 50 вопросов. Остановись, когда информации достаточно.
- На расплывчатый ответ проси конкретный пример.
- Если я противоречу себе, укажи на это.
- Если я описываю желаемое, а не действительное, попроси подтверждение.
- Не хвали ответы. Просто задавай следующий вопрос.

В конце выдай оба файла целиком.
Темы, по которым данных мало, перечисли отдельным списком.
Не создавай файлы до конца интервью.

Начни с блока A.

The quality of the files depends on the detail of your answers. Set aside about an hour for the interview.

How any task is structured

The formula is simple: a project folder, two basic files, and a short request. From task to task, only the contents of projects/ change.

projects/
└── market-review/
    ├── brief.md
    ├── notes.md
    ├── references/
    └── drafts/
  • brief.md — goal, audience, what matters, what should not be.
  • notes.md — raw observations and open questions.
  • references/ — materials to base your work on.
  • drafts/ — intermediate versions.

Four examples

  • A decision memo. The notes contain a month's worth of observations about competitors and risks. The request asks for a document you can make a decision by, not a summary. The output is options, risks, and a recommendation.
  • A quarterly report. The brief specifies the audience and constraints: brief, assumptions separate, focus on consequences. The references/ folder contains metrics and past reports.
  • A presentation outline. The brief states what the presentation should explain and what it shouldn't include. The request asks for the logic of transitions and the task of each slide.
  • Project analysis before investment. The folder contains tokenomics, team messages, third-party opinions, and your past analyses. The request sets the structure: essence, team activity, token distribution, comparison with similar projects, red flags. Cowork relies on the materials you've gathered and doesn't invent partnerships. The investment decision remains yours.

How to phrase requests

Questions first

Прочитай папку проекта.
Задай уточняющие вопросы до начала работы.

Use this technique when the goal isn't fully clear, a specific audience will see the result, or a single wrong assumption will undermine all the work.

A short request

Работай по папке проекта с учётом HOW-I-WORK и SYSTEM-RULES.
Черновик сохрани в outputs/.

Such a request would fail in an empty chat. Here it works because the context already lives in the files. There's a simple indicator: if you end up writing long requests, something's missing from the files and needs to be added there.

How to expand the system

There are two mistakes: staying on the minimum for years when tasks have outgrown it, or immediately building dozens of files for every scenario. Add a file when you see a pattern repeat, not in advance.

  • system/ — when Cowork repeats the same mistake or a rule appears that applies almost always.
  • context/ — when the description of audience or style no longer fits in one file.
  • projects/ — when a typical task has developed a consistent set of files.

The system grows correctly if each task requires fewer explanations, fewer revisions, and fewer restarts. If that's not happening, remove the files nobody uses.