Prompting

A prompt is everything the model reads before it answers: the system message, the conversation, any documents and tool results. Most quality problems in model-based features are prompt problems, and most of them have well-understood fixes. The pages in this section each take one technique, show a weak prompt next to a stronger one, and explain why the change helps.

Every example runs against POST https://app.syntic.ai/v1/chat/completions. For a one-page checklist, see Prompting tips.

Before you change a prompt

Have a way to tell whether a change helped. Ten to fifty real inputs with the answer you expect, and a script that runs them all, turns prompt writing from guesswork into measurement. Test your prompts walks through the setup; do it first.

Techniques

Page	Use it when
Write clear instructions	Replies are vague, too long, or miss part of the task.
Use the system message	You need rules that hold for a whole conversation.
Use examples	The output format or judgement is easier to show than describe.
Ask for step-by-step reasoning	Answers to multi-step problems are wrong or inconsistent.
Work with long documents	You send contracts, reports or many files in one request.
Format the output	Code reads the reply, or the reply must follow a fixed layout.
Reduce hallucinations	The model states things that are not in your material.
Keep a persona	A branded assistant drifts out of character or tone.
Chain prompts	One prompt is trying to do too many things at once.
Test your prompts	You change prompts or models and need to know nothing broke.
Prompting each Syntic model	You move a feature between models or pick one for a new task.

Guardrails and quality

Page	Use it when
Increase consistency	The same input gives answers that vary more than you can accept.
Reduce latency	Replies take too long to start or to finish.
Resist misuse and prompt injection	Users or documents try to change what the assistant does.
Keep your prompt private	Users ask the assistant to reveal its instructions.
Handle interrupted and declined replies	Streams stop early or the model declines a request.

Things that work differently on Syntic

• Sampling settings are fixed. temperature, top_p and the penalty fields are accepted and then ignored, so they cannot make output more or less varied. Shape behaviour through the prompt instead. Per-model sampling control is planned; see Sampling.
• A short identity instruction comes first. The API places a brief Syntic identity line ahead of your system message. Your instructions follow it unchanged and take effect as written. See What the API adds to your request.
• Message content is text. Put documents, tables and data into the prompt as text. Image and file inputs are in preview.
• Long prompts are affordable when they repeat. Identical prompt beginnings are billed at the cached-input rate, so a long, stable system message is cheaper than it looks. See Pricing.