Home / Workshop, capstone and exam
Placing yourself, from beginner to expert, before choosing your project
Four levels of autonomy, a Project fitted with instructions, a Claude Code session framed by a configuration file, a homemade skill, a linked and consolidated memory folder, show which of the three end of course projects matches your own level.
Three projects wait at the end of the course, but they do not call for the same level of autonomy. Placing yourself on four levels before choosing amounts to measuring what you can already drive on your own, rather than what you believe you understood from reading.
Four levels, four ways of framing Claude
The first level sits in a Project on claude.ai, fitted with instructions written once in its instructions field, distinct from its knowledge base where documents are placed, and then applied to every conversation that depends on it, without needing to copy them out again in each message.
The second level is a Claude Code session guided by a configuration file read at start-up: the rules are set there once, and the session applies them throughout the work without needing to repeat them with every request.
The third level is the homemade skill, and it is filed differently depending on which surface is being used.
Skill personnelle, Claude Code : dossier ~/.claude/skills/ma-skill/
Skill de projet, Claude Code : dossier .claude/skills/ma-skill/, dans le dépôt lui même
Skill sur claude.ai : import d'une archive zip, dans Paramètres puis Fonctionnalités, réservé aux plans Pro, Max, Team et Enterprise avec exécution de code activée
Skill sur l'API Claude : envoi vers le point d'entrée /v1/skills
A point not to overlook before relying on a skill built in Claude Code: it does not automatically sync to claude.ai or to the API, each surface keeps its own skills, filed in the location that belongs to it.
The fourth level is not measured the same way
The linked and consolidated memory folder, unlike the first three levels, is not an object documented by Anthropic under that name: it is a way of working, a set of notes linked together and reread at the start of a session, built by the person who uses it rather than supplied ready made by the product. This level is therefore judged by the actual use made of it, not by a feature switched on with a click.
Placing yourself on these four levels before choosing one of the three capstone projects avoids aiming for a project that assumes a level not yet reached, and also makes it possible to recognise a level long since surpassed.
Four levels of autonomy, from the most framed to the most built
A person sets up a Project on claude.ai with instructions written once, then opens a Claude Code session in a folder that contains a configuration file prepared in advance. She sends a first request in this session and observes the reply applying a rule set in this file.
Write in one sentence what this situation establishes, and in one sentence what it does not establish.
What this establishes: This process establishes that the rule set in the configuration file was applied from the session's very first request, without her needing to repeat it in her message.
What this does not establish: It does not establish that this same person has mastered the next level, that of the homemade skill built and filed correctly according to the surface used.
The three most common miscalibrations
- Too broad This person has crossed all four levels of autonomy described in this lesson.
- Too narrow Nothing can be concluded from this single reply until the session has run for several days.
- Beside the point This situation shows that a Project on claude.ai is quicker to set up than a Claude Code session.
- A Project on claude.ai applies the instructions written once in its instructions field to every conversation that depends on it, without needing to copy them out again in each message.
- A homemade skill is filed differently depending on the surface used: a local folder for Claude Code, an archive import for claude.ai, a submission through a dedicated endpoint for the API.
- A skill built in Claude Code does not automatically turn up on claude.ai or on the API, each surface keeps its own skills.
- The linked and consolidated memory folder is a practice built by the person who uses it, not a named feature supplied by Anthropic.
- Placing yourself on the four levels before choosing a project avoids aiming for a level of autonomy not yet reached.
Open your own Project or your own Claude Code session, check whether an instruction or a configuration rule actually applies from the very first message sent, then note which of the four levels described here you sit on today.
Every datable claim in this lesson links here to the public text behind it. A source that does not open proves nothing.
- Anthropic, How can I create and manage Projects consultée le 2026-09-02
- Anthropic, Agent Skills overview consultée le 2026-09-02