Mastering AI Coding: Token Mechanics, Tool Calls, and Best‑Practice Prompt Design

Introduction

The article systematically shares practical knowledge for using AI‑powered coding assistants efficiently. It covers the underlying token model, tool‑calling workflow, code‑base indexing, prompt engineering, and concrete best‑practice recommendations, targeting developers of any experience level.

1. Token Calculation in AI Coding

Each interaction with the model consists of an initial input (system prompt, user query, rules, and dialogue history) and tool‑call results . The total token count is the sum of both parts.

Example:

Initial input = SystemPrompt (500) + UserQuestion (200) + Rules (800) + DialogueHistory (300) = 1800 tokens
Tool results = FileContent (2000) + SearchResult (1500) + LintReport (300) = 3800 tokens
Total tokens = 1800 + 3800 = 5600 tokens

Understanding this budget helps keep conversations within model limits and avoid costly over‑runs.

2. Tool Calls and Workflow

AI assistants use a set of predefined tools to retrieve information, edit files, run commands, or search the web. The article lists the most common tools and their JSON schemas.

{
  "name": "read_file",
  "description": "Read a file (or a range of lines).",
  "parameters": {
    "type": "object",
    "properties": {
      "target_file": {"type": "string"},
      "start_line_one_indexed": {"type": "integer"},
      "end_line_one_indexed_inclusive": {"type": "integer"},
      "should_read_entire_file": {"type": "boolean"},
      "explanation": {"type": "string"}
    },
    "required": ["target_file", "should_read_entire_file", "start_line_one_indexed", "end_line_one_indexed_inclusive"]
  }
}

Other tools include codebase_search, edit_file, run_terminal_cmd, list_dir, and web_search. The assistant must explain why a tool is invoked, call it only when necessary, and limit the number of calls per turn.

3. Codebase Indexing and Retrieval

Cursor builds a Codebase Index by chunking each source file into logical fragments (functions, classes, or code blocks) and converting them into vector embeddings. These vectors are stored in a high‑performance vector database (e.g., Turbopuffer) and searched with semantic similarity.

Indexing steps:

Synchronize the workspace to the server.

Split files into meaningful chunks.

Compute a vector for each chunk using the LLM.

Store vectors in the database for fast nearest‑neighbor lookup.

During a query, the user’s natural‑language request is also embedded, and the most similar code fragments are returned with file paths and line numbers.

4. Prompt Design for Claude Code / Cursor

The assistant’s system prompt combines role description, proactive behavior rules, task management, and security guidelines. A trimmed example is shown below.

You are Claude Code, Anthropic’s official CLI for Claude. You assist users with software‑engineering tasks. Follow these rules:
- Never generate undocumented code unless asked.
- Always explain why you call a tool.
- Keep responses under four lines unless the user requests detail.
- Use the provided <code>TodoWrite</code> and <code>TodoRead</code> tools to track tasks.
- Respect security best practices (no secret leakage).

Project‑specific and user‑specific Rule files can be attached automatically, allowing the model to reuse conventions such as “always respond in Chinese” or “avoid generating test files”.

5. Practical Best Practices

Use clear, concise problem statements that include function names, file names, or module paths to help the model retrieve the right code.

Control context length by trimming irrelevant history; start a new conversation when the topic diverges.

Leverage Revert or frequent commits for CLI tools to roll back to a clean state after a failed step.

Provide diverse information – code snippets, images, web links – so the model can reason with richer context.

Adopt progressive development (pair‑programming style) instead of feeding the whole specification at once.

6. Real‑World Applications

Examples include:

Project familiarization : ask the assistant to describe each module, draw dependency graphs, or locate the best place for new code.

Diagram generation : feed UI screenshots and request PlantUML/Mermaid diagrams; the model returns the textual diagram definition.

Debugging assistance : provide error screenshots, let the model search the repository, and suggest precise code fixes.

Web context augmentation : use @Web to fetch and summarize external articles, then ask follow‑up questions.

7. Recommended Documentation & Rules

Maintain a set of markdown files at the repository root: README.md – project overview, quick‑start, and key features. CHANGELOG.md – version history, breaking changes, and migration notes. ARCHITECTURE.md – high‑level architecture, module breakdown, and core flowcharts. RULES.md (generated by /init) – coding style, naming conventions, and AI‑assistant preferences.

These files serve both human readers and LLMs; the latter can ingest the concise rule sections directly into the prompt.

Conclusion

By understanding token budgeting, mastering tool‑call schemas, leveraging vector‑based code retrieval, and following the outlined best practices, developers can turn AI coding assistants into reliable co‑programmers that boost productivity while preserving code quality and security.