AI Tools

AI Coding Assistants Explained: What They Do and How to Use Them Well

From autocomplete to agents that edit whole projects, AI coding tools have changed how software gets written. Here is a map of the options and the habits that keep code quality high.

By Updated 3 min read
Computer code screen
Photo: Rawpixel (CC0)

Software development was one of the first professions transformed by generative AI, and the tools have evolved quickly. What started as smarter autocomplete now includes assistants that can read an entire codebase, make changes across many files, run tests and propose fixes.

Updated September 2026: we added research on invented software packages, the 2025 Stack Overflow developer survey and a randomized study of developer productivity.

Three kinds of coding assistant

1. Inline completion. As you type, the assistant suggests the next line or block of code. GitHub Copilot popularized this, and most editors now offer something similar. It is excellent for boilerplate and repetitive patterns.

2. Chat in the editor. You highlight code and ask questions: explain this function, find the bug, write tests. Tools such as Copilot Chat, Cursor and assistants in JetBrains IDEs work this way, with awareness of your open files.

3. Agentic coding tools. You describe a task, and the assistant plans it, edits multiple files, runs commands and tests, and iterates. Examples include Claude Code, OpenAI’s Codex, Cursor’s agent features and GitHub Copilot’s coding agent. These can complete meaningful chunks of work, but they need clear instructions and careful review.

What they are good at

  • Writing boilerplate, tests and documentation
  • Explaining unfamiliar code or languages
  • Translating code between languages and frameworks
  • Debugging with error messages and stack traces
  • Prototyping ideas quickly
  • Routine refactoring and dependency upgrades

Where to be careful

  • Plausible but wrong code. AI can produce code that looks right and compiles but mishandles edge cases.
  • Security. Generated code can include insecure patterns, such as missing input validation. Review it as you would a new colleague’s pull request.
  • Invented dependencies. Models sometimes suggest packages that do not exist. In a study presented at USENIX Security 2025, 19.7% of 2.23 million code samples from 16 models referenced at least one nonexistent package, with open models (21.7%) far more prone to it than commercial ones (5.2%). Attackers can register those names with malicious code, a tactic called slopsquatting. Verify every package before installing.
  • Secrets. Never paste API keys or credentials into prompts.
  • Licensing and policy. Follow your organization’s rules on which tools can see proprietary code.

A habit that pays off Ask the assistant to write or update tests along with any change, then run them yourself. Tests turn “looks right” into “demonstrably works.”

Getting better results

  1. Give context. Point the assistant at relevant files, conventions and constraints. Many tools support a project instructions file that is read automatically.
  2. Work in small steps. Ask for one change at a time and review each diff.
  3. Ask for a plan first on bigger tasks, and correct it before any code is written.
  4. Keep humans responsible. The person who merges the code owns it.

Our general prompting guide applies here too: clear goals and good examples produce better code.

Is it making developers faster?

Adoption is widespread: 84% of respondents to the 2025 Stack Overflow Developer Survey said they use or plan to use AI tools, though more developers distrusted the accuracy of AI output (46%) than trusted it (33%). Rigorous studies show mixed results. In a randomized trial by METR in early 2025, 16 experienced open-source developers took 19% longer on real tasks when allowed to use AI tools, even though they believed the tools had sped them up by 20%. Results depend heavily on the task, the developer’s experience and how the time spent reviewing AI output is counted. The honest answer is that the benefit depends on how the tools are used.

A tool for people who can judge the output

AI coding assistants are powerful accelerators for people who already understand what good code looks like. Use them to go faster on the parts you understand, and to learn the parts you do not, but keep review, testing and security in human hands.

Sources

  1. We Have a Package for You! A Comprehensive Analysis of Package Hallucinations by Code Generating LLMs, USENIX Security 2025 (arXiv)
  2. AI section, 2025 Developer Survey, Stack Overflow
  3. Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity, METR, July 2025
Share X LinkedIn Facebook Email

Token & Tell Staff

The Token & Tell editorial desk covers artificial intelligence for everyday users and professionals: the tools, the research and the policy questions behind them. Every piece is researched, edited and checked for accuracy before publication.

Read our editorial standards