Library of skills and roles
Settings → Library — a catalog of 29 items: 21 skills and 8 ready-made agent roles. A skill is a manual for an agent (SKILL.md): when to use, how to act, what not to do. Role is a ready-made prompt for the agent variant.
How the directory is structured
Each card shows:
- name and one-liner — what is it and when will it be useful;
- description — 2–3 sentences;
- characteristics on six axes: analysis, bugs, tests, security, refactoring, documentation (0–5). This is a hint of what a position is strong for, not a quality rating;
- category (AI and LLM, architecture, documentation, git, performance, code quality, review, security, testing);
- action: “Set” for the skill, “Create option” for the role.
Filters: search by name and description, type (skills / agents), category. The order of the cards is in groups by category or according to AI service (see below).
The page is divided into three tabs: Catalog (ready-made skills and roles), My skills (written by you) and UI Catalog (ready-made interface blocks for layout).
Skill installation
Button "Install" copies the file to ~/.claude/skills/<id>/SKILL.md. After this skill is available:
- in composer via
/- like a regular team; - in the agent variant editor - in the list of connected skills.
If such a skill already exists and is different, the previous version is saved next to it as SKILL.md.bak.
Descriptions of other people's skills are often in English - Pulse translates them into Russian in the background and remembers the translation, so you see Russian tips in the “/” list.
For an installed skill, two more buttons appear on the card:
- "Remove" - deletes a file from
~/.claude/skills; the skill remains in the library and is re-installed with one click; - "Update" — appears when the library version differs from the installed one: it applies it, and saves the current one next to it as
SKILL.md.bak.
Your skills
Tab "My skills" — editor SKILL.md right in Pulse: name, one-liner for menu /
and the body of the skill. This skill is placed in ~/.claude/skills just like a library one, and just as visible to agents; its id always starts with my-, so as not to collide with the library one.
Each has its own skill version history: edits are saved, the previous text can be viewed and returned - the same mechanism as the versions of the agent instructions.
Quality assessment
Button "Rate" on the card, sends the skill to the AI service, and it evaluates it along the same six axes that are shown in the characteristics. The rating is saved and visible on both tabs, and the order of cards can be switched to “by rating” - unrated ones go to the end of the list.
The rating is the model's opinion of the skill text, not a measurement of the result. What exactly do agent work inspectors check? Quality inspectors.
Comparison
The “Compare” checkbox on cards (up to four) opens a table: the characteristics are side by side, the best value in each axis is highlighted, the leading position has a mark. There is also a warning if the selected positions conflict in meaning (for example, two opposing review methods).
Creating an agent variant from a role
The roles have a button "Create Variant": select a base agent - Pulse creates a variant with a name, character and a full role prompt. Then you can edit it as usual: model, mode, effort, skills, auto mode. Cm. Agents and models.
Some roles in frontmatter have a list tools (for example, reviewers have read-only: Read, Grep, Glob). It moves to the option and is shown as chips in the editor (uncheck - “No restrictions”). The limitation applies to built-in agents (Local model, GigaChat): tools outside the list are not available to them. External agents (Claude, Codex, Kimi, OpenCode) have tools managed by their own CLI - for them the list is just a hint.
Field model in role files indicates the recommended local H200 model:
qwen3-32b for planning and architecture, qwen2.5-coder-32b-instruct for other roles. When creating a variant, the frontmatter is cut off; field model does not set the session model. The actual model is determined by the selected base agent and its settings. For a swarm with the “H200 only” rule, select a local agent connected to this server and check the model in the variant settings before launching.
How do skills get to an agent?
The skills connected to the option go as a line on the first turn: “Connected skills: /name1, /name2 - use them when appropriate for the task.” This is a recommendation, not a limitation: the agent still sees all of his skills.
When a person starts a request with /name,Pulse passes the original request and content to the built-in agent (Local Model or GigaChat) .claude/skills/name/SKILL.md from a working copy of the task. If there is no such skill, take it ~/.claude/skills/name/SKILL.md. The file is transmitted as file text, not as a system instruction; lines similar to secrets are masked. For the built-in agent, the menu shows only suitable SKILL.md: global ones or those that are already in the current repository commit. An uncommitted skill will not be included in the new working copy of the task. Invalid frontmatter, symlink and file larger than 40 KB are hidden from this menu; a manually entered slash command remains as written. Teams from
.claude/commands are available in the menu of external ACP agents and are processed by their CLI.
UI Catalog
Third tab - UI Catalog: ready-made interface blocks (application framework, tables, forms, states, landing sections) with live preview and rules. The agent takes the block as a basis instead of a template blank - through tools ui_catalog and
ui_template.
Exchange of swarm compositions and agent variants by file
The assembled composition of the swarm and the configured version of the agent can be sent outside as a file
.pulsepreset - block "File Sharing" at the bottom of the pages Settings → Swarm and
Settings → Agents and models.
- Export to file... — system save dialog. Squads are leaving the “Swarm” page, and agent options are leaving the “Agents” page.
- Import from file... — the file is first shown, but not applied: what is in it, what is missing from it, and what it argues with by name.
The file is regular JSON, you can put it in the repository next to the code and send it to a colleague.
What's in the file: names and descriptions, roles of participants and presenter, base agents, selected models, auto mode, own role instructions, prompts and connected option skills.
What is not and will not be in the file: API keys, tokens, paths to binaries and server addresses. They lie in the device key ring and do not cross the file border - just like when transferring settings between account machines.
The person decides whether the names match
Someone else's file is external data. He understands the scheme: a position with an unknown role or an unknown agent does not reach, and each such find is described in a line. A cast that arrives without a designated leader is given the first participant as leader—this is also said out loud, not silently.
If the name is already taken, Pulse doesn't rewrite - it asks:
| Answer | What's going on |
|---|---|
| skip | the visiting position is not applied (Pulse does the same if there is no response) |
| add a copy | “Name (from file)” appears next to it, this one remains untouched |
| replace the local one | the content is replaced, but the name and internal id remain local - they are referenced by the instruction history |
Nothing changes until pressed "Apply".
What's planned
- their roles as agents next to their skills: in the editor so far there are only skills;
- your blocks in the UI catalog: now blocks come with the Pulse update.