Skills are reusable instructions you write once and make available to every AI tool connected through the MCP. A skill captures how your team does something, such as your screener conventions, how you pick incentive amounts, or your rules for grounding insights in transcript evidence, so agents follow your process instead of guessing at it.
Admins manage skills under Settings → AI & MCP, the same page where you connect the MCP.
How skills reach your agents
Every skill you enable is advertised to connected AI tools automatically. When an agent connects, it sees a list of your enabled skills, each with its name and description, and loads the full instructions for whichever ones are relevant to the task at hand. You don’t have to tell people which skill to use; the agent picks based on the descriptions you write, which makes a clear, specific description the most important part of a skill.
Creating a skill
- Go to Settings → AI & MCP and find the Skills section.
- Click New skill.
- On the Details tab, fill in:
- Name: a short, human-readable label (e.g. Screener creation).
- Slug: the identifier agents use to load the skill. It’s derived from the name automatically; you can override it. Slugs are normalized to lowercase and hyphens.
- Description: one or two sentences on when an agent should reach for this skill. This is what the agent reads to decide whether the skill applies, so be specific about the situation it covers.
- Instructions: the skill body, written in the rich editor (headings, bold, lists). This is the actual guidance the agent follows.
- Click save. The skill is created disabled by default.
- Toggle it on in the list when you’re ready for agents to see it.
Only enabled skills are advertised to agents. Disable a skill at any time to pull it back without deleting it.
By default a skill is advisory: agents see it and use it when relevant. You can go further and require a skill, so that an agent must load it before it’s allowed to run certain MCP tools. This is how you enforce a process rather than just suggesting one.
On a skill’s Settings tab:
- Turn on Require before tool use.
- Choose the scope:
- Every MCP tool: the skill must be loaded before any tool runs.
- Only the tools I pick: select the specific tools the skill gates (e.g. require your Screener creation skill only before screener-editing tools).
Required skills show a Required pill in the Skills list.
When a required skill hasn’t been loaded, the gated tool call is softly declined and the agent is told exactly which skill to load first, so a well-behaved agent loads it and continues on its own, without the person driving the session having to intervene. Editing a required skill’s instructions re-arms the gate, so agents pick up your latest guidance rather than a stale version.
Start skills as advisory and only require the ones that protect a real standard, like fraud checks in screeners or how incentives are set. Requiring everything makes agents slower without adding much governance value.
What makes a good skill
- Write a precise description. The agent chooses skills by description alone. “How we write screener questions to filter out fraud” beats “screener help.”
- Keep instructions actionable. Concrete rules, examples, and do/don’t lists outperform background prose.
- One skill, one job. Several focused skills are easier for an agent to select from than one long catch-all.
Troubleshooting
Don’t see Skills under Settings → AI & MCP? Skills roll out alongside MCP access. Confirm with your Great Question account team that the feature is enabled for your account.
A tool call keeps getting declined? Check whether a required skill is gating it. The agent is told which skill to load, so a capable AI tool should load it and retry automatically. If it doesn’t, ask it to load that skill by name (or slug) and try again.
Agents aren’t using a skill you created? Confirm the skill is toggled on — new skills are disabled by default. If it’s enabled but still not picked up, tighten the description: agents choose skills by description alone, so a vague one gets skipped.
