# Getting Reliable Answers

Treasure AI Studio knows less on its own than it appears to. The model brings general SQL and general reasoning; almost everything specific to *this* platform — Treasure AI's own SQL functions, the schema a segment definition has to follow, what a journey step may contain — is carried by **skills**, and which of those are enabled differs by account. It knows nothing at all about *your business* until someone tells it: that "active customer" means something specific at your company, that revenue is read from one table and never another, that cancelled rows do not count, that a finding should come back as three bullets and a number.

So an answer rests on two layers you control. The skills decide how well the agent speaks the platform — several are on by default, and the SQL skills are not. Your instructions and your own skills decide whether it knows your business. Left unsupplied, either gap is filled by inference, and reasonable inference is not the same as right.

That is where unreliable answers come from, and better prompting alone does not close it. This page explains what shapes an answer, why the same question can come back differently twice, and where each kind of knowledge belongs so you stop supplying it by hand.

The two questions worth separating
**Is the output right?** — that comes from your request, the skills in play, and the tools the agent ran. **Does the conclusion follow?** — that is the model's interpretation, which is generated text and carries no guarantee, however well the work beneath it went.

## Prerequisites

- Signed in to Treasure AI Studio ([Getting Started](/products/ai-studio/getting-started))
- Familiarity with [Core Concepts](/products/ai-studio/concepts) — work folders, skills, and how a chat is set up


## What Shapes an Answer

Every answer is assembled from layers, and each one can be inspected on its own. When something comes back wrong, the useful question is not "is Studio broken" but "which of these was it".

| Layer  | What it decides  | How to inspect it  |
|  --- | --- | --- |
| **Your request** | Which objects, which period, which records count as real | Re-read what you asked. Anything you left implicit, the agent chose for you |
| **Custom instructions** | The organization-wide baseline applied to every chat | **Settings → Instructions** (administrators) |
| **Work folder instructions** | Standing context for every chat in one folder — default database, conventions, house rules | Open the work folder's instructions |
| **Skills** | The procedures and platform conventions the agent follows — SQL patterns, segment schemas, workflow syntax | **Settings → Marketplace**: which plugins are on. Custom and shared skills count too |
| **The tools it ran** | What was actually done — the query executed, the segment written, the file produced | Expand the tool calls in the chat and read the inputs and outputs |
| **The interpretation** | The summary, the causes, the recommendation | The prose in the reply — compare every figure against the tool output |


The mechanics of how these are assembled into a plan are in [Agent Orchestrator](/products/ai-studio/concepts/orchestrator). This page is about the half you control: what you put into those layers.

## Why the Same Question Can Come Back Differently

Asking twice can produce two different answers, and both can be defensible. Three reasons account for nearly all of it:

- **The model is not deterministic.** Two runs of the same prompt can take different routes to a result. This is a property of the model, not a fault in Studio.
- **Your request under-specified something.** If you did not fix the objects, the period, or the exclusions, each run picks them again.
- **The context differs.** Which skills are on, what the work folder instructs, what was said earlier in the conversation, and what the agent discovered along the way all feed the next step.


Only the second and third are yours to remove — and removing them is what turns a demo into something a team can rely on.

## Where Your Business Knowledge Belongs

Write your specifics down. Where you write them decides how far they reach and who gets them.

| What you have  | Where it belongs  | Why there  |
|  --- | --- | --- |
| A constraint for this one request | The prompt | No reuse expected. Name the objects, the period, the exclusions and move on |
| Term definitions, the tables a project uses, filters that always apply to this work | [Work folder instructions](/products/ai-studio/concepts/work-folders) | Read at the start of every chat in that folder, so nobody restates them. Saved as the folder's `CLAUDE.md` |
| Conventions everyone in the organization shares — a default database, a house output style, terms with a fixed meaning | [Custom instructions](/products/ai-studio/custom-instructions) | Added to every agent's system prompt across the organization. One per account, set by an administrator |
| A procedure worth repeating: a SQL template for something hard to get right, a required output format, how a metric is to be read, the steps of a recurring task | A [skill](/products/ai-studio/skills/skills) | Applied automatically when a request matches, and [shareable with your team](/products/ai-studio/skills/sharing-custom-skills) so everyone works from the same procedure instead of each person's own |


### Instructions State Facts; Skills Carry Procedures

The line between the last two is the one people get wrong. **Instructions state standing facts** — "revenue means `analytics.orders_clean`, excluding `status = 'cancelled'`". **A skill carries a procedure** — the actual SQL template for the quarterly revenue query, the shape of the output, the steps for interpreting it. If you find yourself writing SQL, or a numbered sequence, into an instruction field, that is a skill trying to exist.

Skills have two properties instructions do not: they are matched to a request rather than always-on, and they can be shared, so one person's hard-won procedure becomes the whole team's starting point.

### Two Rules of Thumb

- **Write it the moment you correct the agent twice for the same thing.** A second correction is evidence the convention is real and unwritten.
- **Put it at the widest scope where it is always true.** A rule that holds for one project belongs in that work folder; a rule that holds for the company belongs in the custom instruction. Rules written too wide are worse than none, because they fire where they should not.


## Skills Are the Lever You Control

Much of what looks like "the AI is bad at this" is an agent working without the knowledge it needs. Skills carry that knowledge, and they are yours to manage:

- **Enable the ones that match your work.** The SQL skills for querying, workflow skills for pipelines, tdx skills for CDP objects — several are not on by default. See [Skills & Marketplace](/products/ai-studio/skills/skills).
- **Remember that custom and shared skills also steer the agent.** A skill written for one team's conventions applies to everyone who enables it, and a vague or conflicting instruction inside one shows up as unstable output.
- **Change one thing at a time.** When output quality shifts, toggle a single plugin and repeat the same request; anything else and you cannot tell what moved.


## The Interpretation Is Generated Text

The summary that follows a result is written by a language model reading that result. It can restate a figure incorrectly, and it can offer a cause the data does not support — a plausible sentence carries no more evidence than the work behind it.

Two habits keep this contained: check every figure you plan to act on against the tool output, and ask "what in the result supports that?" when a conclusion arrives unasked. Studio keeps a standing reminder under the composer for the same reason — *AI-generated output may be incorrect. Verify important information before relying on it.* See [AI-Generated Output Is Yours to Review](/products/ai-studio/security#ai-generated-output-is-yours-to-review).

## Next Steps

- [When a Query Answer Looks Wrong](/products/ai-studio/query/answer-quality) — Triage for a specific number or query
- [Work Folders](/products/ai-studio/concepts/work-folders) — Where a project's standing context lives
- [Custom Instructions](/products/ai-studio/custom-instructions) — The organization-wide layer
- [Skills & Marketplace](/products/ai-studio/skills/skills) — Enabling, writing, and sharing procedures
- [Agent Orchestrator](/products/ai-studio/concepts/orchestrator) — How a request becomes a plan and tool calls