Catalyst Agent Skills

An agent skill is a set of instructions that helps an AI app use Catalyst consistently. It tells the app how to select an organization, find the right TrialStack operation, and stop for your approval before a change.

You do not need to edit the skill. When a TrialStack package includes it, the package and Catalyst connection provide the instructions to the app.

What the skill does for you

The trialstack-catalyst skill tells the AI app to:

• confirm which organization it is using;
• start with read-only work;
• use record identifiers returned by TrialStack;
• treat an empty result as data, not a connection failure;
• show missing information instead of making assumptions;
• preview every proposed change; and
• wait for a separate approval message before applying a change.

These instructions make conversations more predictable across Codex, Claude, and Microsoft 365 Copilot.

What the skill cannot do

The skill cannot grant access, change your role, or bypass TrialStack permissions. Catalyst checks your identity, organization membership, and permissions whenever the app uses a tool.

The skill also cannot approve work for you. TrialStack remains the source of truth, and you keep responsibility for clinical judgment and formal approval.

Starter requests

Find information

Find the TrialStack information related to [trial or question]. Work read-only,
tell me which organization and records you use, and do not make assumptions
about missing information.

Review a trial

Review TrialStack trial [trial ID] with a focus on [topic]. Show missing or
uncertain information. Do not make changes.

Prepare a change

Prepare this TrialStack change: [requested change]. Show the organization,
record, proposed values, and expiry time. Do not apply it.

For administrators and developers

Catalyst publishes the skill through its MCP skills interface and as a verifiable resource. Supported platform packages include the same canonical instructions.

Some clients also show reusable prompts:

Prompt	What it helps a person do
discover-trialstack-data	Find relevant information without changes.
review-trial	Review one trial and identify missing data.
prepare-governed-change	Create a change preview for human approval.

The published skill includes a SHA-256 digest. Integrators can use the digest to confirm that a client loaded the expected instructions.

See Model Context Protocol (MCP) for the connection, access model, approval flow, and technical tool reference.