Skip to content

Answer new-hire questions with context files and skills

Build a new-hire helpdesk that always knows a handful of small, stable facts and only reaches for a written procedure when the question actually needs one. The two capabilities behind that are context files and skills, and this page exists because picking the wrong one is the usual reason an agent either ignores what it was told or never opens what it needed. This is a procedure to run, with one recorded run as a reference.

Which one fits

Holds The model sees it
Context file Standing facts, small and stable — payroll timing, the IT channel Always (inject), or on demand (link)
Skill A procedure for one kind of task — how to request access Only when the model decides that task is what is happening
Knowledge collection A corpus too large to read in full — a whole handbook, every policy PDF Only the chunks a search returns

The onboarding guide below is short and always relevant, so it is a context file. "How do I get access to a system" is a procedure with steps and an exception, so it is a skill. If your onboarding material is instead a fifty-page handbook, bind it as a knowledge collection and keep this page's pattern only for the short, standing facts. See skills or knowledge? for the same distinction from the skill side.

What you need

A running installation with a model profile. No sandbox and no embedding model — both files here are small enough to inject or load whole.

Prepare the input

A context file of standing facts:

# Acme Robotics — new-hire quick facts

- Payroll runs on the last business day of the month.
- The standard laptop is a MacBook Pro; loaner laptops are requested from IT, not HR.
- The internal help channel for IT questions is #it-help.
- Health insurance enrollment is open during your first 30 days; after that, only
  during the November open-enrollment window.

A skill for the one procedure new hires ask about most:

# Requesting access

Most access requests go through the #it-help channel, not a person directly.

1. Post in #it-help naming the system and the reason you need it.
2. IT grants standard tools (chat, email, laptop) within one business day.
3. Anything touching customer data (the CRM, production databases) needs your
   manager's written approval first - tag them in the same thread.
4. Access to the payroll system is never granted through chat; email
   payroll@acme-example.com instead.

Acme Robotics is invented. The reference facts: payroll on the last business day, a MacBook Pro by default, and CRM access needing a manager's approval first.

Build the agent

  1. In Context → New, create a file named onboarding-guide, paste the facts above, set Mode to inject, and give it a description a person would recognise later.
  2. In Skills → New, create request-access with the procedure above and a description written for the model: "When somebody asks how to get access to a tool, a repository, a system, or is not sure who grants it." — a linked skill is chosen from its name and this line alone.
  3. Create an agent in Agents → New agent and select your model profile.
  4. In Toolbox, enable Context and bind onboarding-guide. Enable Skills and bind request-access.
  5. Set the instructions below, then Publish.
You are Acme Robotics' new-hire helpdesk assistant.
Answer from the standing facts you were given, and use a bound skill's
procedure when a question is about how to do something.
If you are not sure, say so rather than guessing.

Run it

Ask a standing fact first, then a procedure question:

When does payroll run, and what laptop will I get?

How do I get access to the CRM?

Check the result

Check Reference
Payroll and laptop question Answered directly, no tool call — the context file is already in the prompt
CRM access question Calls load_capability for request-access before answering
The CRM answer Names the manager's written approval as the first step, not just "post in #it-help"
A question the context file does not cover (e.g. "what's the dress code?") Says it does not know, rather than inventing a policy
Editing the context file afterwards The next run reflects the edit with nothing republished on the agent

The first two rows are the check that matters: one answer comes from text sitting in every prompt, the other from a tool call the model chose to make. If either happens the other way around, the wrong capability was reached for.

Recorded on v0.0.504, 25 September 2026

Model: Claude Sonnet 4.6 through OpenRouter. The payroll/laptop question used 914 input tokens with no tool call and answered: "Payroll runs on the last business day of the month... The standard laptop is a MacBook Pro". Cost 0.004 USD.

The CRM question called load_capability with {"id": "request-access"}, got back the skill's full body, and answered: "Since the CRM touches customer data, there's a specific process... Get your manager's written approval first... Post in #it-help. Tag your manager in the same thread". Cost 0.009 USD.

When it goes wrong

  • A standing fact is missing from an answer. Check the context file's Mode. A link file is not in the prompt at all until the model decides to read it — for facts that must never be missed, use inject.
  • The skill never loads. The model chooses it by name and description alone; a description that reads like a title ("Access requests") tells it less than one written as when to reach for this.
  • The agent recites the skill for every question. The skill's description is too broad, or the instructions do not distinguish "standing fact" from "procedure" clearly enough for the model to tell which this question is.
  • An edited context file does not change the answer. Confirm you edited the organization's file and not a copy — context files are bound by id, and there is no per-agent fork of one.
  • A skill proposal appears instead of a direct answer. The agent may try to improve the skill mid-conversation; that is a proposal for a person holding skills:edit to apply or discard, not something a run applies on its own. See skills.

Record the trial

Keep both files' content and ids, the agent version, the two questions and answers, and whether each one used a tool call. A person still writes and edits the standing facts and the procedure — this pattern only decides where each piece of text lives, not who is right about payroll dates.

Next steps

The same agent can be bound to a Slack bot for a team channel instead of the console — see answer a handbook question in Slack, which walks through binding, linking accounts and checking which run belongs to whom. If the onboarding material grows past a page or two, move it into a knowledge collection instead of stretching an injected context file.