Tabnine's Prompting Guide

Defining a Prompting Strategy for Code Projects

Every project is going to require a different strategy based on your priorities. Effective prompting starts with structured thinking.

In Tabnine, you’re not just prompting a text model, but an Agent that has access to your codebase, open files, and custom commands.

A good prompt doesn’t just describe what you want but instructs the Agent how to operate in your project: what it can look at, what rules to follow, and what to leave alone.

---

A. Specify the Key Elements

When designing a prompt, clarify the following:

1. The Task:

   • What are you trying to achieve? What’s the ‘goal state’? What should the result be?

   • Example: "Apply the extract method for refactoring to this code."

2. Context:

   • What is the source material you’re working from, and where should the agent look (open file, specific path, remote repo)?

   • Example: "@related_class" or "See the implementation in utils/helpers.py."

3. Constraints: What boundaries or conditions must be respected?

   • Example: "Keep current behavior. Preserve function signatures."

Note: It is best to avoid negative prompting. A negative version of the above example would be "Do not change behavior. Preserve function signatures."

1. Process: What are the instructions, or how should the agent approach the work step-by-step?

   • Example: "Start by identifying duplicate logic, then suggest reusable methods."

2. Validation: What signals indicate that the prompt was successful?

   • Example: "Refactored code should pass all existing tests."

3. Format: How should the output be structured?

   • Example: "Respond with a JSON object containing method_name, start_line, and new_code."

4. (Re-)Iteration Plan: What are the next steps if the first output isn’t perfect? Treat this like directing a junior engineer: tell the agent when to pause, ask questions, or request approval before applying large changes.

   • Example: "After generating the refactor, ask: 'Do you want to apply this to other classes?'"

5. Ask for Feedback (optional):

   • Example: "If this approach looks flawed, explain why and propose an alternative."

---

Template Example

Goal: Apply extract method refactoring to the selected function
Source: @related_class
Constraints:
  - Do not change behavior
  - Maintain interface compatibility
Process:
  - Identify duplicate code blocks
  - Suggest method names based on action
Validation: Must retain all unit test coverage
Format: List of changes in JSON:
  [
    {
      "method_name": "...",
      "start_line": ...,
      "new_code": "..."
    }
  ]
Next Step: Ask if similar changes should be applied elsewhere
Feedback: Flag any logic that may not refactor cleanly

---

When using Tabnine as an Agent over your codebase, be explicit about scope and permissions:

• Scope: “Only modify this file and test files in /tests/user/.”

• Permissions: “Do not create new services; only change existing methods.”

• Safety checks: “Before proposing changes, summarize what you understood about the current behavior.”

Why This Works

A prompting strategy aligns the LLM with your development mindset. Instead of fishing for the right phrasing, you define a clear lane for the model to operate in—resulting in better, faster, and more predictable completions.