Project context for the AI

The AI learns about a project from four sources: the index, the Project Hints, the projectdocs and specs, and skills. This topic covers what you maintain yourself.

Project Hints

The Project Hints are a Markdown text that every session of the exploration receives as part of its instructions, with every provider. Use them for rules that cannot be read from the code, for example:

• "The UI is in English, even when we chat in German."
• "Run cargo clippy -- -D warnings before every commit."
• "Do not add dependencies without asking."

To edit them:

1. Choose File › Edit Project Hints…, or open Analysis › Project Settings… and the AI Hints tab.
2. Edit the text. On first open, the file is created from a template.
3. Save.

The text is stored in .codeexplorer/project_hints.md of the exploration and shared with the exploration. New sessions use the new text.

Keep the hints short and stable. A rule for one task belongs in the chat. The reason behind a piece of code belongs in a projectdocs entry anchored to it.

The Project Hints are the only project-wide instruction channel. Tracemoor does not read AGENTS.md. It reads a CLAUDE.md of the project only when the isolation setting below allows it.

projectdocs and specs

The AI searches projectdocs by keyword and by meaning, and finds the entries and specs anchored to a symbol. The Tracemoor guide tells it to read them before it changes a symbol and to say so when a change contradicts a decision or a spec. See projectdocs and the AI.

The more of the reasoning is written down and anchored, the less you need to repeat in the chat.

Facts about the project and personal memory

Some providers keep their own memory, for example the auto memory of Claude Code under ~/.claude. That memory belongs to your user account on this machine. It does not travel with the exploration and other providers do not see it.

Put facts about the project into projectdocs and rules into the Project Hints. Keep personal preferences in the memory of the provider.

Project files of the AI: isolation

Many repositories contain their own instructions for AI tools: a CLAUDE.md, a .claude/ folder with settings, hooks, agents and skills. By default Tracemoor does not load them. A session receives the Tracemoor guide and the Project Hints and nothing else from the project. Instructions written for another workflow, for example many parallel agents, would otherwise override the guide.

AI › AI Settings…, tab Isolation:

The settings apply to Claude (CLI). If the AI suddenly ignores the Tracemoor tools or anchors, check these settings first.

Skills and agents

A skill is an instruction for one kind of task, for example a review. An agent is a helper that the session can delegate work to. See Longer tasks for the agents.

AI › Browse Skills & Agents… lists them with their triggers and shows the prompt of the selected entry.

Built-in skills:

A skill is activated by its trigger at the start of a message or by choosing it from the slash command list. The active skill shows as a chip in the input bar. AI › Detach Active Skill or a click on the chip removes it.

Tracemoor reads skills and agents from these places:

On equal names, a project entry replaces your own, and your own replaces a built-in one. A skill is a Markdown file with name, description and triggers in the frontmatter and the instruction as body. If a skill from ~/.claude has the name of a built-in skill, the built-in one stays available as CE_<name>.