Skip to main content

Prompt Engineering

Prompt engineering is the practice of writing inputs that reliably get the result you want from a large language model. It's the craft that underlies everything you build in Revelry Ai: the one-off prompts you type into Chat, the instructions inside a Skill, the persona instructions that define an Agent, and the recommended prompts you attach to an Agent are all just text you write for a model to act on.

So this page is the foundation for the rest of the system: you engineer prompts, you build Skills out of well-engineered prompts, and you compose Agents out of Skills, instructions, and recommended prompts. Get the writing right and every layer above it gets better.

The throughline of every technique: reduce ambiguity. A model can only act on what you actually say, so the clearer and more structured your instruction, the better and more consistent the output.

Start With the Outcome​

Before writing the prompt, decide what a good answer looks like. Define an output contract: the format, length, tone, and any sections or fields the response must include. Knowing the target up front makes the prompt easier to write, and makes it easy to tell whether the result is right.

Core Techniques​

These apply to any text you write for a model, on any surface.

1. Be Specific and Provide Context​

Vague prompts produce vague results. Spell out the task, the audience, the purpose, and any constraints. Give the model the background it needs rather than assuming it knows.

❌ "Write about our product."

✅ "Write a 150-word product description for a B2B audience evaluating our scheduling app, emphasizing ease of use and cross-platform support."

2. Assign a Role​

Telling the model who to be focuses its tone, vocabulary, and expertise. A single sentence is often enough to measurably improve domain-specific output. (This is exactly what an Agent's instructions do for a whole conversation.)

"You are a senior technical writer. Draft release notes for the feature below for a non-technical audience."

3. Show Examples (Few-Shot Prompting)​

If you want a particular style or structure, demonstrate it. One or two examples of the input-and-ideal-output pattern teach the model far more than description alone.

"Rewrite each headline in our brand voice. Examples: Input: 'New dashboard released' → Output: 'Meet your new dashboard' Input: 'Bug fixes shipped' → Output: 'Smoother, faster, fewer snags'"

4. Specify the Output Format​

State exactly how the answer should be structured: a table, JSON, a numbered list, a fixed set of headings. Explicit formats are easier to read and easier to use programmatically.

"Return the result as a Markdown table with columns: Risk, Likelihood (Low/Med/High), Mitigation."

5. Use Delimiters and Structure​

When a prompt mixes instructions, context, and data, separate them clearly so the model never confuses one for another. Delimiters (XML-style tags, triple quotes, or Markdown headings) make the boundaries unambiguous.

<instructions>
Summarize the document below in three bullet points.
</instructions>

<document>
...source text...
</document>

For more structured prompts, XML-style tags (e.g. <context>, <task>, <example>) are a particularly robust way to delineate sections and tend to work well across models.

6. Ask for Step-by-Step Reasoning​

For analysis, math, or multi-step logic, invite the model to work through the problem before answering. This "chain-of-thought" approach noticeably improves accuracy on complex tasks.

"Think through this step by step: first identify the constraints, then evaluate each option against them, then recommend one."

7. Phrase Instructions as Positive Directives​

Tell the model what to do rather than what to avoid. Positive instructions are easier to follow than prohibitions.

Prefer "Keep this to at most 150 words" over "Don't go over 150 words."

8. Guide the Start of the Response​

You can steer format and direction by beginning the answer yourself ("prefilling" or starter text). Opening with { nudges the model toward JSON; a leading phrase sets the tone.

"The three biggest risks to this launch are: 1."

9. Order the Prompt Deliberately​

Put role and context first, then the task, then detailed instructions and the output format last. Establishing the framing before the ask helps the model apply the right lens to everything that follows.

Keep It Tight​

More words is not more control. Model reasoning degrades as prompts bloat with redundant detail, so favor structure over length: a focused, well-organized prompt beats a long, rambling one. Cut anything that doesn't change the output.

Writing for each surface​

The core techniques above are universal, but you'll apply them differently depending on where the text lives. Revelry Ai gives you four surfaces to write for, each with its own emphasis.

One-off prompts (Chat)​

The text you type into Chat in the moment. Optimize for speed and iteration: write enough to reduce ambiguity, run it, and tighten as you go. When a prompt earns its keep through repeated use, promote it to a Skill.

Skill instructions​

The base prompt text inside a Skill. Because a Skill runs repeatedly across different inputs, invest in structure and reusability: lean on delimiters, a clear output contract, and user inputs for the parts that change each run. Project context is injected automatically, so don't hard-code it. See Building a Skill.

Agent instructions​

The persona prompt that defines an Agent. These apply to every message in a conversation, so focus on durable behavior, not a single task: who the Agent is, its tone and expertise, the kinds of work it's for, and the rules it must always follow. Describe how it should use its attached Skills, but leave the per-task specifics to the conversation. Think of this as technique #2 (Assign a Role) scaled up to a whole assistant.

A recommended prompt is a saved snippet of prompt text attached to an Agent, not a Skill. It doesn't execute anything; clicking its tagline pre-fills the chat input so the user can edit and send it. Recommended prompts serve two purposes:

  • Inspiration: they show a new user what the Agent is good at instead of leaving them at a blank box. Write taglines that read like the jobs people actually want done.
  • Starting points for known Skills: since an Agent's attached Skills are the only ones available in its conversations, you can phrase a recommended prompt so that sending it naturally leads the Agent to invoke the right Skill.

Tips for writing recommended prompts that reliably trigger the Skill you intend:

  • Mirror the Skill's job in the prompt text. If the Agent owns a "Sprint Summary" Skill, a recommended prompt like "Draft a sprint kickoff summary for this project" points straight at it. Use the language the Skill is built around.
  • Keep the tagline short and outcome-focused; let the prompt carry the detail. The tagline is the button label ("Summarize last sprint"); the prompt is the fuller instruction that gets pre-filled.
  • Write prompts in the Agent's voice. They're the Agent's suggestions to the user, so they should match the persona you set in the Agent's instructions.
  • Cover the Agent's range. A handful of prompts that each map to a different attached Skill is the fastest way to teach users everything the Agent can do.
  • Leave room to edit. Users can change the pre-filled text before sending, so a recommended prompt can be a strong default starting point rather than a rigid, fill-in-every-blank command.

Iterate and Evaluate​

Treat any prompt as a living artifact; this is just as true for an Agent's instructions and recommended prompts as for a one-off message. Run it, compare the result against your output contract, and tighten the wording. Small clarifications often yield large gains. When a prompt earns its keep through repeated use, save it as a Skill, and when a set of Skills belongs together for a recurring job, compose them into an Agent so your whole team gets the same quality.

Before: "Draft an email for customer service scenarios."

After: "Write a polite, empathetic email template for customers who received a faulty product, including clear steps to exchange or return it."

Further Reading​

These vendor guides go deeper on the techniques above:

Next Steps​