Skills & Connectors

Two ways to extend what the assistant can do in Data Insight: skills teach it how you want an analysis run, and connectors let it reach sources beyond the dataset — published literature, clinical trials, external references — when you ask for them.

Skills

A skill is a saved playbook: a short Markdown document describing how a particular analysis should be carried out. When your request matches a skill, the assistant loads the playbook and follows its method using its normal dataset tools. You get the same analysis, done the same way, every time — without re-explaining it.

What a skill can do

• Standardise a recurring analysis. Survival by mutation status, a cohort attrition breakdown, a standard demographics report — write the steps once and reuse them across datasets and chats.
• Encode your conventions. Which time unit to use, how to define the cohort, which chart to end with, what to caveat. The playbook carries these preferences so you don’t have to repeat them.
• Guide, not override. A skill shapes the method. It cannot widen what the assistant is allowed to do: dataset boundaries, disclosure rules and safety checks still apply exactly as without it.

Official and personal skills

If you upload a personal skill with the same name as an official one, yours overrides it in your chats. Both pages show a notice explaining which version the assistant uses.

Using a skill in a chat

There are three ways to bring a skill into a conversation:

1. Type / in the composer. A palette lists your available skills; keep typing to filter by name or keyword. Selecting one adds a skill chip to your message and inserts the phrase Use the "<skill name>" skill. — send the message and the assistant applies that playbook.
2. Name it in plain text. Writing “use my survival-by-mutation skill” has the same effect. When you name a skill, the assistant uses exactly that one; it never swaps in a different skill that merely looks similar.
3. Let the assistant find it. For a request that looks like a repeatable analysis, the assistant searches your skills first and applies a match when there is one. If nothing fits, it proceeds with its own approach — a missing skill never blocks an answer.

Messages that used a skill are marked with the skill’s name, so you can see which playbook shaped a reply.

Browsing skills

Open Skills in the sidebar. You can search by name, description or keywords, and filter to Official or Personal. Each skill page shows its description and the full SKILL.md playbook the assistant loads. You can delete your own skills; official skills are read-only.

Creating your own skill

Click Add skill and upload a SKILL.md file. Requirements:

• A YAML frontmatter block with a name and a description, then the playbook in Markdown below it.
• The name is lowercase words separated by hyphens. It is the skill’s identity: uploading a file with the same name replaces that skill, which is how you publish a new version.
• The description says when the skill applies — it is what the assistant matches your request against, so make it specific.
• The playbook body can be up to 20,000 characters. Supporting files such as scripts are not used yet.
• The file must be 64 KB or smaller and contain plain text only.

---
name: survival-by-mutation-status
description: >
  Use when the user asks how survival differs between patients with and
  without a mutation in a given gene.
---

## Workflow

1. Inspect the dataset so you use real column names.
2. Count the cohort split by mutation status, then chart both curves.
3. Report the median survival per group and note the censoring rate.

Write the playbook as instructions to an analyst: the order of steps, the definitions to use, and how to present the result. Uploads are screened, and a skill that tries to change the assistant’s safety behaviour or extract internal details is rejected.

Good candidates for a skill

• An analysis you have asked for more than twice.
• A report your team expects in a fixed shape.
• A definition that is easy to get subtly wrong — a cohort inclusion rule, a survival endpoint, a stage grouping.

Connectors

A connector is an external source the assistant can search from inside a chat — for example a literature index or a clinical-trials registry. Connectors are configured by PAICON; when any are available you see them under the + menu in the composer as Connected sources, and under Settings → Connectors.

What a connector can do

• Add context the dataset does not have. Ask for the published evidence on a gene, the trials recruiting for an indication, or a drug target’s references, and the assistant queries the connected source and cites it — with links to the original record where the source provides them.
• Combine with your data. A typical flow: explore a cohort in the dataset, then ask “what does the literature say about this mutation’s prognosis?” in the same chat.

How it behaves

• Opt-in only. The assistant reaches for a connector only when you ask for outside information or name the source. Ordinary questions about a dataset never leave the platform.
• No identifying data leaves the platform. Only non-identifying scientific terms — gene symbols, disease names, topics — are sent to a connector. Patient or record identifiers and raw dataset rows are never included, and anything a connector returns is screened for personal data before the assistant sees it.
• External text is treated as untrusted. The assistant uses connector results to inform its answer and names the source, but does not follow instructions embedded in returned content.
• Sources are cited. When a result carries a source link, such as a PubMed record, the reply includes it as a link you can open.
• There is a per-chat limit on external calls. If you reach it, start a new chat to continue.

Availability

Skills and connectors are rolled out gradually. If you don’t see Skills in the sidebar or Connected sources in the composer menu, they are not yet enabled for your account — see Help & FAQ for how to get in touch.