The home page of this site can show thousands of skills, thousands of plugins, and a long model list. Installing the top of each list in the first week is how people end up with an agent that is slow, contradictory, and full of prompts they never read. This is a smaller week. It assumes you write or review code and you already have one agent open, whether that is Cursor, Claude, DeepSeek Harness, or something else.
Day one: one model, one repository
Pick a single model and stay with it for the week. If your agent already chose one, keep it. If you are choosing, open models and apply the cuts in how to read a model listing: license, context window big enough for one package of your repo, and a price or a GPU you can tolerate. Do not compare five models today. You do not yet have a prompt worth comparing.
Clone a repository you understand, on a branch. Ask the agent to explain one function you can already explain yourself. If it cannot, the problem is the setup, not the lack of a skill. Fix the setup. A skill will not repair a missing index or a model that never saw the file.
Day two: one skill that matches the work
Choose one skill with the method in how to choose an agent skill. If you are writing a backend, look at Development. If you are updating docs beside the code, look at Knowledge Management. Install it with the procedure in install, update, and remove, on the branch, and ask for a change you can review in five minutes.
Remove it the same day if it touches files outside the task. That is the lesson of day two, and it is more useful than a second install. The stop list tells you which cards to close before you paste.
Day three: read a diff like a reviewer
Stop asking for features. Ask the agent to review a diff you wrote, or a diff it wrote yesterday, and to list risks only. A good review names a line. A bad review says the code "could be more robust." If you want a standing check inside DeepSeek Harness, that is a workflow plugin, and it can wait until day five. For now the review is a prompt you type yourself, so you learn what you actually want checked.
Day four: a second task, not a second stack
Use the same model and the same skill on a different task in the same repository. You are testing whether the skill generalizes or whether yesterday was luck. If it only works when you repeat a long preamble, the preamble is the skill and the installed one is not helping. Delete it and keep the preamble in your own notes.
This is also the day to notice secrets. If the agent asks for a token, stop. Day four is early enough to learn that habit and late enough that you have a real task to protect.
Day five: at most one plugin, if you use Harness
If you run DeepSeek Harness and the UI is the bottleneck, add one client plugin using the plugin guide. If the bottleneck is that the model cannot see a screenshot, look at model inference instead. Do not add both. Do not add a marketplace "so you can browse." You already have this website for browsing.
If you do not run Harness, skip plugins entirely. Spend the day writing down the three prompts you actually reused. Those sentences are your personal skill, and they are allowed to be boring.
Day six and seven: write down the keepers
Keep a short note with four lines: the model name you pinned, the skill you kept or the reason you removed it, the plugin you kept or the fact you need none, and one task the agent is not allowed to do. That last line might be "do not commit," "do not touch migrations," or "do not send mail." Put it where the agent will see it, in a project rule you wrote, not in a third-party skill you have not read.
On day seven, resist the directory. You will have seen categories you never opened: content creation, office, education. They are real shelves for other jobs. They are not homework. Open one later, when the job exists, with the checklist beside you.
A first week that ends with one working loop is a success. A first week that ends with fifteen installs is a cleanup project.
What the second week is allowed to add
Only add something in week two that failed in a specific way during week one. If the agent forgot a decision you wrote down, consider a project note before a memory plugin. If the agent edited a file you had named as forbidden, put that name in the rule you already started, before you shop for a workflow plugin. If the model ran out of window on one package, then you have a reason to open models and change the pin, using how to read a listing. A vague feeling that "power users install more" is not a failure report. It is advertising you are doing to yourself.
Keep the uninstall path from week one working. The skill you kept should still be one you can name and remove, as in the install guide. If you cannot find it in settings anymore, you already lost the thread, and week two starts by cleaning that up rather than by opening Content Creation or Office Efficiency out of curiosity.
A week that is not about code
If your actual job is a lesson plan, a quote, or a document, swap day two's category and keep the shape of the week. Education, Life Service, and Office Efficiency replace Development. The rule does not change: one skill, one fake input whose answer you know, and permission to remove it the same day. Do not import the coding-agent plugin advice if you do not run Harness. Plugins are not a general productivity layer. They are extensions for one environment, which the plugin guide is honest about.
At the end of the week you should be able to answer four questions without opening the site. Which model name is pinned. Which skill is installed, or why none is. Which plugin is installed, or why none is. Which action the agent is not allowed to take. If you need the directory in front of you to answer, the week was browsing. Repeat day two before you add anything else. If the four answers feel too small compared with the size of this site, that feeling is the directory doing what directories do. The about page explains that most cards are an index. Your week is finished when the loop is reliable, not when you have sampled every category. Put the four answers where the next person on the project will see them. A pinned model that lives only in your chat history will be "whatever the default is" by Friday of week two.