DocsAgentsSkills & jobs

Skills, policies & jobs.

Domain specialization for agents — and the pre-built jobs that ship out of the box. Skills are how an agent learns the shape of a task; jobs are how a one-line ask becomes a typed deliverable.

§ 01Skills

A skill is a markdown file in the skill library — a piece of domain guidance the agent can pull into context on demand. Where the system prompt tells the agent what kind of agent it is, a skill tells the agent how to do this specific kind of task: the shape of the input, the structure of the output, the gotchas, the worked example.

How skills are loaded

Skills are not all loaded up front. The agent calls View_Skills at the start of a task to see what's available, picks the relevant ones, and calls Read_Skill(name) to pull each one's markdown into its working context. This keeps the active context small and lets the skill library grow without ballooning every prompt.

flow

1. agent receives prompt
2. agent calls View_Skills() → list of available skill names
3. agent picks relevant skills
4. agent calls Read_Skill("contract_summary")
5. skill markdown is now in context for the rest of the run

Editing skills

Users can edit, delete, or reset skills at runtime from the workspace UI. Edits apply to subsequent runs in the same project; resets restore the platform default. Skills are scoped per user — your edits to a skill don't change anyone else's copy.

TipThe right time to edit a skill is when you've done the same task three times and the agent keeps making the same small mistake. A two-line addition to the relevant skill is usually enough.

§ 02Policies

A policy is a guardrail bundle — a small set of do's, don'ts, output specifications, and forbidden actions attached to an agent type. Policies are kept deliberately minimal: the goal is to encode invariants that hold regardless of the prompt, not to micromanage the agent's behavior.

A typical policy contains:

Do. Things this agent should always do — cite sources, attach a totals reconciliation, include a confidence note.
Don't. Things this agent should never do — fabricate quotes, edit user files in place, run destructive shell commands without confirmation.
Output spec. The required structure of the deliverable — sections, fields, file format.
Forbidden actions. Tools or surfaces the agent must not use for this task type.

Policies are combined with the system prompt at run time and surface as part of the agent's tool-use validation loop. A policy violation produces a typed error rather than a silent slip.

§ 03Jobs

A job turns a one-line ask into a typed deliverable. Each job is a pre-built bundle of prompt, skills, policy, and output schema, designed to produce the same shape of output every time on the same shape of input.

Ten jobs ship with the platform today.

Job	What it does
Accountant	Categorizes a period's transactions and produces a monthly close pack.
Invoice Extractor	Pulls vendor, dates, line items, totals, and tax from a folder of invoices into a CSV.
AP Manager	Reconciles invoices against POs and contracts; flags exceptions and produces an approval queue.
Meeting Notes	Turns a transcript or recording into a structured summary with action items and owners.
Contract Summary	Extracts parties, term, renewal, termination, and payment from any contract into a one-page memo.
Executive Brief	Synthesizes a topic, deal, or company into a board-ready briefing with citations.
Architect	Produces a concept package — design intent, key decisions, options matrix, references.
RFP Responder	Drafts an RFP response and a compliance matrix mapping each requirement to a section of the response.
SOW Builder	Generates a `sow.md` and a `milestones.csv` from a deal brief and an engagement scope.
Underwriting Brief	Builds a credit/risk underwriting brief from financials, references, and disclosures.

Each job is invokable from the workspace UI. You fill in the fields the job asks for — input files, parameters — and the agent runs the job's playbook end to end, producing the typed deliverable as one or more artifacts on the run.

§ 04Authoring your own jobs

Custom jobs are currently invite-only. Design partners get access to the authoring surface, where you can:

Compose a new job from a skill bundle, a policy bundle, and an output schema.
Test it against representative inputs.
Publish it into your workspace so your team can run it.

The authoring surface itself is documented separately. If you'd like to be part of the design partner cohort, see Requesting access.

NoteThe pre-built jobs are not the ceiling — they are the ground floor. The model behind them is general-purpose and the same agent loop runs your free-form chats. Jobs are a structured wrapper over the same engine.

§ 05Where to go next

How files flow into jobs and out as artifacts → Drive & artifacts.
How the runtime invokes skills and policies → Runtime & subagents.
How to use the agents API end to end → API overview.