# Agent recipes

> Patterns for getting reliable, high-quality work out of the agent — from a one-shot dataset description to a full draft report. Each recipe shows the prompt, what the agent does, and how to verify the output.

*Source:* https://recense.ai/docs/agent-recipes

## Before you start

- A dataset loaded.
- An active AI provider (BYOK or built-in) or a connected MCP client.

## How to prompt the agent

The agent is good at survey-shaped work: describing data, building tables, applying filters and weights, summarising results, and drafting narrative. It is less reliable when the prompt is vague or asks it to make a strategic call without enough context.

Three habits make the difference. Be specific about the variables and segments. Name the measure you want. State the format of the answer.

> **ℹ The single biggest lever**  
> Write good dataset instructions. The agent reads them on every prompt — they are the difference between "analyse the data" working badly and working well.

## Recipe — describe the dataset

Use this as a first prompt on a new dataset to confirm the agent has loaded it and understands the structure.

- What you should see: a short structural summary, named splits, and a flagged anomaly if one exists.
- Verify by spot-checking the sample size and at least one demographic split against the Variables view.

*Prompt:*

```text
Describe this dataset: sample size, panel, key demographic splits, and any question groups that look stackable. Highlight anything that looks unusual.
```

## Recipe — build a cross-tabulation

Specify variables, segments, measure, and weighting explicitly.

- What you should see: a new table on the canvas matching the spec.
- Verify the weight is applied (Weight pill shows the variable name) and the unweighted base looks plausible.

*Prompt:*

```text
Cross-tabulate Q5 (brand awareness) by AgeGroup with column percentages. Apply the design weight. Show significance letters at 95%.
```

## Recipe — segment comparison

Useful for comparing two cohorts on a battery of metrics.

- What you should see: a side-by-side table with a Difference column and significance markers.
- Verify segment definitions match what you intended — the agent should state them back.

*Prompt:*

```text
For respondents in the High-engagement segment vs Low-engagement segment, compare mean scores on the Q12 satisfaction grid. Show the difference and flag any significant gaps.
```

## Recipe — narrative summary

Use after you have built and reviewed the supporting tables. The agent reads what is on the canvas.

- What you should see: three paragraphs in a workspace note, anchored near the source tables.
- Verify every statistic in the narrative against the source table — agents sometimes round inconsistently.

*Prompt:*

```text
Draft a three-paragraph executive summary from the four tables tagged "headline". Lead with the strongest finding, qualify with sample size, and end with the one thing a stakeholder should do next.
```

## Recipe — text coding

Use during text coding to consolidate or refine themes.

- What you should see: a revised theme list with explicit merges and splits explained.
- Verify by sampling responses for any merged or split themes before publishing the codebook.

*Prompt:*

```text
Review the proposed themes for Q15. Merge anything semantically duplicate, split themes that are doing two jobs, and propose at most one new theme if you see a gap.
```

## Verifying agent output

Treat the agent like a fast junior analyst. The output is usually right, sometimes wrong, and always worth a quick check before it leaves your workspace.

- For tables: check the weight, base, and at least one cell against the variables view.
- For narratives: check every number in the narrative against its source table.
- For coded data: spot-check 10 responses across high- and low-confidence themes.
- For statistical claims: confirm the test the agent named matches what the table is configured to run.

## Known limitations

- The agent does not invent statistical methods Recense doesn't support. If you ask for a method outside the methodology page, it will tell you and propose the nearest equivalent.
- The agent will not delete tables or notes without explicit instruction.
- For very long conversations, summarise progress occasionally — context windows have limits.
- Built-in mode and BYOK use the same tools and the same prompt; differences in output usually reflect model capability, not Recense behaviour.

## Next steps

- **[Write dataset instructions](/docs/dataset-instructions)** — Persistent context is the highest-leverage way to improve agent output.
- **[Methodology](/docs/methodology)** — Know which test the agent will reach for in each scenario.
- **[Connect MCP clients](/docs/mcp-setup)** — Drive the same agent from Claude Desktop, Claude Code, or ChatGPT.