10 Rules for Getting Better Results from AI Coding Agents

HackerRank's 2025 Developer Skills Report found that 97% of developers use at least one AI assistant and nearly a third of code is now AI-generated, prompting a set of 10 practical rules for improving results from AI coding agents such as Claude Code, Codex, Cursor, Copilot Agent, and Gemini CLI. The rules emphasize starting with detailed specifications, using repository-level instruction files like AGENTS.md, and keeping instructions concise to avoid context overload.

10 Rules for Getting Better Results from AI Coding Agents Everyone's using AI coding agents. Here's how to make yours actually useful. AI coding agents are no longer just autocomplete tools. They can read your repository, edit multiple files, run commands, create pull requests, and work through multi-step development tasks. Claude Code , , Codex https://openai.com/index/introducing-codex/ , Cursor https://cursor.com/ , Copilot Agent https://github.com/features/copilot/agents , and similar tools are changing how developers build software. But better tools do not automatically mean better code. Gemini CLI https://geminicli.com/ HackerRank's 2025 Developer Skills Report found that 97% of developers use at least one AI assistant, and that nearly a third of code is now AI-generated. It also notes that AI is increasing delivery pressure, not removing the need for strong engineering judgment. The difference between a good AI-assisted developer and a frustrated one often comes down to workflow. AI coding agents perform best when they are given clear goals, project context, validation rules, and a safe way to iterate. Here are 10 practical rules for getting better results . 1. Starting With a Specification, Not a Vague Prompt Bad prompt: Build the dashboard. Better prompt: Build a customer churn dashboard. Goal: Show churn rate, active customers, monthly revenue, and top churn risk factors. Scope: - Add a dashboard page at /dashboard. - Use the existing API client. - Reuse the current chart component. - Do not change the database schema. Acceptance criteria: - Page loads without console errors. - Metrics match the /analytics/churn endpoint. - Add tests for the data transformation function. - Run lint and tests before final response. Coding agents are good at execution, but they need a target. A good specification should include the goal, scope, constraints, files likely to change, acceptance criteria, and test commands. This mirrors how professional developers already work: the task is not "write code," but "make a change that satisfies a clear definition of done." A recent paper on coding-agent bootstrapping makes a similar point from a research angle: for agents, the specification can become the stable artifact of record, while implementations may be regenerated or revised. 2. Using an AGENTS.md , CLAUDE.md , or Copilot Instructions File Do not repeat the same project rules in every prompt. Put persistent instructions in a repository-level agent file. The open AGENTS.md format describes itself as a README for agents: a predictable place to give coding agents setup commands, test commands, coding conventions, and repository-specific instructions. It is already used by more than 60,000 open-source projects. For example: AGENTS.md Setup - Install dependencies with pnpm install . - Start the app with pnpm dev . - Run tests with pnpm test . Code style - Use TypeScript strict mode. - Prefer functional components. - Do not add new dependencies without approval. Before finishing - Run lint. - Run relevant tests. - Summarize changed files and why they changed. Codex reads AGENTS.md before doing work and supports layered guidance from global, project, and directory-specific files. GitHub Copilot also supports repository custom instructions in .github/copilot-instructions.md , which can tell the agent how to build, test, validate, and follow project conventions. 3. Keeping Agent Instructions Short and Useful An agent instruction file is not the place to paste your entire engineering handbook. Anthropic's skill-authoring guidance says good skills should be concise, well-structured, and tested with real usage. It also warns that once instructions are loaded, every token competes with the rest of the task context. A recent paper on found common "configuration smells," including lint leakage, context bloat, skill leakage, and conflicting instructions. In its sample of 100 popular repositories, lint leakage appeared in 62% of files and context bloat in 42%. AGENTS.md and CLAUDE.md filesGood instruction files include: - How to install, build, test, and lint. - Project-specific architecture notes. - Naming and style rules. - Security constraints. - What not to touch. - How to report completion. Bad instruction files include: - Generic coding advice the model already knows. - Long explanations of common frameworks. - Contradictory rules. - Outdated commands. - Too many "always" and "never" instructions. 4. Asking the Agent to Inspect Before Editing For non-trivial tasks, tell the agent to understand the repository before changing it. Example: Before editing, inspect the relevant files and summarize: 1. which files control authentication, 2. where the bug likely lives, 3. what tests already cover this area, 4. the smallest safe change. Do not modify files until after this summary. This prevents the common failure mode where the agent writes a plausible fix in the wrong location. Make the agent locate the system before asking it to change the system. 5. Using Planning for Complex Tasks, but Not Over-Planning Tiny Edits For large changes, a plan helps. For small changes, too much planning slows the loop. GitHub Copilot CLI's best-practices documentation explicitly recommends plan mode for tasks where a structured implementation plan is useful before code is written. Use planning for: - Migrations. - Multi-file refactors. - Auth changes. - Database changes. - Performance work. - Production bug fixes. - Anything touching security or payments. Skip heavy planning for: - Typo fixes. - Small test additions. - Simple CSS changes. - One-function refactors. 6. Making Tests the Contract AI-generated code often looks right before it is right. HackerRank argues that debugging is becoming a central AI-age skill because AI-generated code still needs reliability, security, and integration work. Its guidance recommends practical, multi-file debugging scenarios with failing tests, misleading logs, and integration edge cases. Use tests as the agent's contract: Write failing tests first for this bug. Confirm they fail. Then implement the smallest fix. Do not modify the tests after implementation unless the test itself is wrong. Run the relevant test suite before finishing. This pattern is especially powerful with agents because it gives them a feedback loop. Without tests, the agent optimizes for plausible code. With tests, it optimizes for working code. 7. Giving Examples of the Desired Style Agents follow examples better than abstract taste. Instead of saying: Make it clean and production-ready. Say: Follow the style of src/features/billing/CreateInvoice.tsx . Use the same error-handling pattern as src/lib/apiClient.ts . Use the existing Result<T type instead of throwing raw errors. GitHub's Copilot best-practices guide recommends breaking down complex tasks, being specific, providing examples of inputs and outputs, and following good coding practices when prompting. Examples reduce ambiguity. They also prevent the agent from inventing a new style that conflicts with the existing codebase. 8. Controlling Dependencies and Permissions Agents like to solve problems by installing packages, changing configs, or widening permissions. That may work locally but create long-term maintenance risk. Add rules such as: Dependency policy - Do not add production dependencies without approval. - Prefer existing utilities before adding new packages. - If a new dependency is necessary, explain why and list alternatives. This is especially important because modern coding agents can run commands and interact with development tools. If your tool supports hooks or permission controls, use them. Claude Code hooks can run deterministic commands at specific lifecycle points, which is useful when you need certain checks to happen reliably rather than hoping the model remembers. 9. Reviewing AI Changes Do not review AI-generated code by asking, "Does it look good?" Review it by asking: - Did it solve the requested problem? - Did it change unrelated behavior? - Did it add unnecessary abstraction? - Did it weaken security? - Did it hide errors instead of fixing them? - Did it update tests? - Did it follow project conventions? - Can the diff be smaller? Experienced developers value agents as productivity tools, but still retain control over design and implementation because they care about quality attributes. This is the right mental model. The agent can draft, explore, refactor, and test. The developer still owns architecture, correctness, and maintainability. 10. Iterating on Your Agent Instructions Your first AGENTS.md will not be perfect. When the agent makes a mistake, do not only fix the code. Fix the instruction that allowed the mistake. Example: Agent mistake: It modified generated files directly. Instruction update: Generated files - Do not edit files in src/generated/ . - Update the schema or generator source instead. - If unsure, ask before changing generated files. Agent mistake: It ran the entire slow test suite every time. Instruction update: Test strategy - For frontend component changes, run the affected component tests first. - Run the full test suite only before final completion or when shared utilities change. Final Thoughts AI coding agents reward developers who can write clear specifications, design good feedback loops, and review code carefully. The goal is not to "vibe code" your way through production systems. The goal is to turn the agent into a faster implementation partner inside a controlled engineering workflow. The best results come from giving the agent: - A clear task. - A small, accurate context. - A repository instruction file. - Examples of existing style. - Tests as a contract. - Permission boundaries. - A human reviewer who still owns the outcome. In short: better agent output does not start with a better model. It starts with better engineering discipline. is a machine learning engineer and a technical writer with a profound passion for data science and the intersection of AI with medicine. She co-authored the ebook "Maximizing Productivity with ChatGPT". As a Google Generation Scholar 2022 for APAC, she champions diversity and academic excellence. She's also recognized as a Teradata Diversity in Tech Scholar, Mitacs Globalink Research Scholar, and Harvard WeCode Scholar. Kanwal is an ardent advocate for change, having founded FEMCodes to empower women in STEM fields. Kanwal Mehreen