← Volver a las guías

System Prompt Troubleshooting: Fix Drift and Conflicts

What is a system prompt?

A system prompt defines the standing instructions an AI assistant should follow: its role, boundaries, workflow, and response format. Unlike a user message, which usually describes the current task, a system prompt establishes behavior across tasks within the supplied conversation.

For example, a support assistant might receive: You troubleshoot billing issues. Ask for the error message before suggesting fixes. Never request passwords. If account access is required, direct the user to authenticated support. This is more actionable than “Be a helpful billing expert.”

Problem: The model ignores your instructions

First, inspect the outgoing request rather than assuming your application sent the prompt correctly. Logging a redacted request payload often reveals misplaced instructions, missing history, or middleware that overwrites messages.

• Check the instruction field. In OpenAI Chat Completions, system instructions typically appear in a message with role system; some newer models use developer messages instead. Anthropic’s Messages API uses a top-level system parameter, not a system entry inside messages. Follow the endpoint and model documentation.
• Remove contradictions. “Always answer immediately” conflicts with “Ask a clarifying question when requirements are missing.” State which behavior takes precedence.
• Replace vague goals with observable rules. Instead of “Be concise,” specify “Use no more than five bullets unless the user requests detail.”
• Check conversation handling. Include the instructions on each independent request unless the API explicitly manages persistent conversation state. Verify that truncation or summarization does not discard them.

Problem: The prompt is long but results are inconsistent

Organize instructions into role, task, constraints, and output format. Keep each rule in one place. Repeating slightly different versions creates ambiguity rather than authority.

A useful starting prompt is: You are a developer support assistant. Identify the likely cause using the supplied evidence. If essential information is missing, ask one targeted question. Otherwise provide numbered troubleshooting steps and a verification step. Do not invent logs, prices, or API capabilities.

Add one short example only when a recurring edge case needs clarification. Then test whether it improves unseen cases; examples can accidentally teach unwanted wording or assumptions.

Problem: Retrieved content overrides your rules

This is often prompt injection: a document, webpage, or tool result contains instructions such as “ignore previous rules.” Tell the model to treat external content as evidence, not authority, and clearly separate that content from trusted instructions.

However, a system prompt is not a security boundary. Keep credentials out of prompts, validate tool arguments, restrict permissions, and require approval for sensitive actions. Enforce access control in application code, even when the model says a request is authorized.

Problem: JSON output breaks your application

“Return JSON” is not sufficient for a production parser. Define required fields, value types, and the behavior for missing information. Use schema-constrained structured output when the selected model and endpoint support it. Parse and validate the response server-side, and apply a bounded retry or repair strategy when validation fails.

FAQ: How should I test a system prompt?

Build a small regression set covering normal requests, missing inputs, conflicting demands, malicious retrieved text, and strict formatting. Record the model identifier and prompt version. Change one variable at a time and measure concrete outcomes, such as valid JSON rate or correct escalation decisions. Repeat important cases because generation can vary.

FAQ: Can I reuse one prompt across Claude and GPT?

Reuse the intent, but test each model separately. Instruction handling, supported parameters, and output features differ. A relay simplifies connectivity; it does not make every model behave identically.

59API provides low-cost, pay-as-you-go access to Claude and GPT models, with native official-quality models and no downgrades. It supports Claude Code, Codex, and OpenAI SDK workflows. Configure its API base URL, https://api.59api.com, according to your client’s URL conventions, checking whether it expects a /v1 suffix. Affordable requests make repeated prompt testing more practical.

FAQ: When should I shorten the prompt?

Shorten it when rules duplicate each other, irrelevant background consumes tokens, or tests show no benefit from extra examples. Keep necessary safeguards. If you want an economical place to compare prompt revisions, consider signing up for 59API; its referral rebate offers an additional potential saving.

Precios · Docs