Skills, shared blocks, prompts & instructions
This page brings together four Web UI screens that define reusable blocks for agents: Skills, Shared blocks (presets), Prompts and Instructions. The idea is to describe a rule or a skill once and attach it to many agents.
Skills
The “Skills” screen is the workspace’s shared-skills catalog. A skill has a name, a description (the slug is generated automatically), a body (Markdown), an enabled flag and an order. Curated out-of-the-box skills carry multilingual descriptions. Skills are then linked to agents on the agent screen; an agent can also have its own skills.
Skills are instructions and grant no additional tool access. From Core 1.5.30, execution adapts to the task: simple questions and short actions run immediately; longer work continues with saved progress and a separate result check, without asking the user to switch modes or repeat the request. Earlier ordinary/durable mode preferences do not require a manual switch. Explicit administrative disabling, current permissions, approvals, limits and the user’s stop request remain binding. An execution policy included in an import remains a separate, explicitly selected component.
Task execution: Ordinary work / Long running with verification.
Shared blocks (presets)
The “Shared blocks” screen holds company-wide behavior blocks (shared rules, personas, policies). A block has a kind (from the section set), a name, a body and a sensitivity (safe / sensitive). Editing sensitive blocks goes through the approvals queue. These are the “presets”: link one block to many agents and change the rule in one place.
Prompts
The “Prompts” screen is a template library. A template has name, description
and body; the body uses {{variable}} placeholders (a badge shows the variable
count). The preview button asks for each variable’s value and renders the
result. Prompts apply wherever you need reusable request text.
Instructions
The “Instructions” screen has two parts:
- Custom instructions — tenant-wide text appended to every agent’s system prompt.
- Allowed models — a checkbox grid over the models of configured providers. This restricts which models agents may use at all.
Tips
- Push recurring rules into shared blocks instead of copying them into every system prompt.
- Keep the allowed-models list narrow — it is cheaper and more predictable.