Using role skills

How it fits together

Your sources

Notion

Slack

GitHub

…

Context layer

everything your company knows

Your agent · a role skill

What you get

A team member, from the first runbriefs · reviews · prep — cited

Four pieces, each owned by the right party:

• The context layer holds what the company knows — curated from your own sources, cited, kept current. It is the understanding a senior hire takes years to build. What the layer is →
• The agent is whichever one you already use — Claude, Cursor, Codex, anything that speaks MCP. It brings its own memory and its own tools; Pensieve gives it the company.
• The skill is the role: what a Chief of Staff holds in their head, how they judge, the shifts they run. Pensieve writes and hosts the skills. Update the plugin to receive improvements.
• The schedule is the shift: one line naming the skill and where output lands, plus a cadence. For the bundled skills, set the destination and cadence in the task so the same skill can serve different teams. In your own skills, you can also define the output and handoff as part of the role.

What a skill contains

Every role skill has the same shape:

• Who you are — the role, and Pensieve as its understanding of the company. Live systems the company also uses (a calendar, a CRM, a channel) are where the newest facts are, and the agent uses them the way the role would.
• What you hold in your head — the parts of the context layer the role must have read before it has an opinion.
• How you judge — the handful of principles that make the role good at its job: a Chief of Staff surfaces decisions over status; a Head of Product weighs evidence over volume; a Head of Growth names mechanisms concretely and knows what has already been tried.
• Shifts you run — the recurring pieces of work, each with the shape its output takes. The task names which; the skill also covers anything else the role would reasonably be asked.
• Writing back — when the agent should add to the context layer, and when it should not. The rule is judgement, never routine: save_data for something the company should keep and does not have, edit_page for a Page the agent has verified is wrong. The agent's own briefs and reviews are not saved unless you ask.

The agent has latitude. The skill says what matters and how to judge; it does not script which Pages to open.

Installing a skill

The Pensieve plugin includes every role and the connection to your company. Install it once, sign in, and start a new conversation.

Choose your client and install Pensieve →

Keep marketplace updates enabled where supported, or update the plugin manually. Receiving new skills depends on those update settings.

Only need a skill file? Each role's skill page lets you read, copy or download it. Filesystem clients can also use npx skills add pensieve-io/plugins; this installs the skills without the MCP connection or context hooks. A copied file stays as it was until you update it.

Using MCP prompts? Clients with a prompt picker can load a role from the Pensieve server's prompt list. This supplies the instructions for that conversation; it does not install a plugin.

Run a role

Name the installed skill and the work you want, for example: “Use the Chief of Staff skill to prepare a brief for tomorrow’s leadership meeting.” Start with a use-case guide for its prerequisites and a copyable prompt.

Creating a scheduled agent

A scheduled task runs the skill regularly and tells it where to put the results. You can set the destination and requested output in the task instructions without changing the skill itself.

In Claude, open Scheduled → New task → Set up manually. Name the task, add its Instructions, then choose a Frequency and time. Connect any tools it needs to write the result, review Permissions, and save. Claude's scheduling guide explains the available options.

Each role guide includes a copyable example, a setup screenshot and a possible handoff:

• Head of Product: suggestions in Notion, then coding agents and pull requests.
• Chief of Staff: commitments in Notion, then decision briefs and project plans.
• Head of Growth: experiments in Notion, then research and campaign drafts.

The schedule repeats the review. Handoffs to other agents need their own automation. To change how a role works, adapt a copy of its skill or create your own.