Skill library

The skill library is a set of task guides that the model loads only when a task needs them. The prompt stays small, and the model still gets detailed advice for common work.

The short version: the system prompt lists each guide's name and description. The model reads a full guide with the skills tool when a task matches it.

Progressive disclosure

Otto follows the Agent Skills integration guide: metadata at session start, instructions when a skill applies and references when needed.

flowchart LR
  start([Session starts]):::muted --> catalog[Names and<br/>descriptions in prompt]
  catalog -- task matches --> guide[skills read<br/>SKILL.md]:::accent
  guide -- condition applies --> ref[skills read<br/>a reference]

Level	What it holds	When the model gets it
Catalog	Name and description of each task guide	In the system prompt, from the start
Guide	The body of `SKILL.md`	When it calls `skills` with `action: "read"`
Reference	A file in the guide's `references` folder	Only when the guide says its condition applies

Four core guides skip this and are always in the static system prompt: connected apps, schedules and watches, artifacts (creating files) and video links.

The task guides

The library has 25 task guides in src/runtime/skills.

Home and life	Work	Mail, files and research
`events-and-leisure`	`business-finance`	`calendar`
`family-admin`	`customer-support`	`data-and-spreadsheets`
`food-and-groceries`	`ecommerce-operations`	`documents-and-forms`
`health-admin`	`hiring-and-career`	`files-and-knowledge`
`home-services`	`project-operations`	`inbox`
`learning-and-study`	`sales-and-crm`	`research`
`personal-finance`	`software-and-websites`	`writing-and-content`
`privacy-and-accounts`
`shopping`
`subscriptions-and-bills`
`travel`

Seven guides have a reference for one specific case: recurring meetings, data imports, government forms, catalog synchronization, corpus research, purchase claims and existing travel bookings.

How the model reads a guide

The skills tool has two actions:

{action: "list"} returns names and descriptions.
{action: "read", name} loads SKILL.md. Add file: "references/<name>.md" to load one reference.

The system prompt tells the model to pick a guide by the task's purpose, the user's constraints and the required result. A compound task can load more than one guide. After compaction, the model can read a guide again.

Read-only turns can load guides. Activity checks can't, because they're limited to app reads.

Add or change a guide

Create src/runtime/skills/<name>/SKILL.md, using a lowercase name with hyphens.
Add front matter with name, which must match the folder, and description, which says when the model must use the guide.
In the body, write the task decisions, the failure cases and the evidence needed to check the result.
For a long procedure that applies only in one case, add references/<name>.md. Link it from the guide and state the condition.
Keep each file at or below 64 KiB.

---
name: example-task
description: Plan an example task and check the result. Use for example requests. Use research for source investigations.
---

The decisions this task needs, its failure cases and the checks that prove the result.

Restart after you change the catalog or any metadata. Task guide bodies and references are read from disk on each call, while core guides load once at startup. Invalid metadata or a missing guide stops Otto from starting. A guide with disable-model-invocation: true can't be listed or read by the model.

File and permission limits

The library reads only its own folder in the server source tree. Pi's default skill discovery is off, so it never searches the workspace, the home folder, .pi folders or connected accounts.

Hidden folders and folder symlinks are ignored.
Each read must stay inside the selected skill folder after the real path is resolved.
Only SKILL.md and named Markdown files directly inside references can be read.
Each file must be a regular file of 64 KiB or less.

Skill content is trusted server guidance. Reading it doesn't mark the thread as exposed to outside content.

What the evaluation showed

A saved evaluation of an earlier source version compared 300 attempts without task guides to 300 attempts with them.

Primary judge	Without guides	With guides
Task completion	71.0%	76.7%
Unsafe attempts	2	4

An independent judge found no gain in the full-case pass rate. The recorded skill reads covered 16 of the 25 task guides.