Set up
Connect your tools
Claude is more powerful when it works directly with your systems. You control permissions and access. Learn about tool access(opens in new tab).
Add a connector(opens in new tab) for each tool you want Claude to use.
Set your working folder
Drag the files that anchor the agent (your service catalog export, the team directory, the glossary, the "where to find things" onboarding doc) into one folder and give Claude access to it(opens in new tab) in the Claude desktop app. These become the routing table the agent consults before searching anywhere else. Create a project(opens in new tab) from that folder so the catalog and ground rules stay attached as the agent evolves.
The prompt
Copy this into Claude
Build an internal "ask anything" agent for engineers. When someone asks "how do I get a staging API key," "what owns the orders table," or "why did deploys slow down last month," find the answer across our systems, cite the source, and tell them who to ask if you're not sure. Get one good answer first, then turn it into a shared skill.
Why this works
Give an example for each question type. A how-to, an ownership lookup, a metric why: covers the question types engineers actually ask.
Cite the source, every time. Engineers trust the link, not the prose; citations make the answer verifiable.
Name an owner when unsure. Asking for "the owning team or channel" gives Claude something to fall back on when it's not confident. You get a pointer to follow up on instead of a guess to verify.
Put reference docs in the working folder. Service catalog and team directory tell Claude where to look and who to point to.
Get a better draft
Add an example to match. Drop an example you like into the folder and Claude matches your structure and voice.
Ask it to flag uncertainty. Add "flag anything you're not confident about" so you know where to look first when you review the draft.
Make Claude work for you
Once Claude answers your three questions well, tell it what to call the skill. Layer in which spaces and repos to search first, the "I'm not sure" threshold, the tone, and the systems that are off-limits. A few minutes of conversation and the skill speaks your stack from then on.
Make what we've done in this task so far into a skill called /ask-the-company.
Review the skill, and save it if Claude hasn't already.
Make it repeatable
Share it as an artifact
Once the agent answers your test questions well, publish it as an artifact so anyone in the org has one link to ask.
Log every question it couldn't answer
The agent gets better when you know where it falls short. Ask Claude to set up a scheduled task or open Scheduled in the sidebar, and a weekly run writes the unanswered questions to your folder so the platform team can fill the gaps.
Set up a scheduled task for every Friday at 4pm: list the questions /ask-the-company couldn't confidently answer this week, group them by topic, and write them to Platform/ask-company/gaps-<week>.md with a suggested owner for each.
Weekly, lists the questions /ask-the-company couldn't answer, groups them by topic, and writes a gap report with suggested owners.
To use a folder on your computer during a scheduled task, keep your computer on and awake with the Claude desktop app open so Claude can access it.
Share with your teammates
Your /ask-the-company skill now carries your sources, your routing table, and your confidence rules. Share it workspace-wide so every engineer asks the company the same way, and the answer to "how do I…" stops depending on who's in the channel.