On this page
Developers use AI coding tools to understand unfamiliar code, write tests, review pull requests, write documentation, debug errors and learn new frameworks. The best results come when the developer stays in charge: reading every line, running the tests and keeping secrets and personal data out of prompts. Treat AI output as a draft from a fast but careless colleague.
This guide is for working developers, from juniors in their first job to team leads setting rules. If you are not a developer and want to know what agent tools do, read our plain guides to Claude Code and OpenAI Codex.
What the evidence says
AI coding tools are now normal, and trust in them is low. The 2025 Stack Overflow Developer Survey found:
- 84% of respondents use or plan to use AI tools in development, up from 76% the year before.
- 51% of professional developers use them daily.
- More developers distrust the accuracy of AI output (46%) than trust it (33%). Only about 3% trust it highly.
- The top frustration, named by 66%, is answers that are "almost right, but not quite". The second, at 45%, is that debugging AI-generated code takes longer.
Speed gains are not automatic. In a 2025 randomised trial by the research group METR, 16 experienced open-source developers worked on 246 tasks in projects they knew well. With early-2025 AI tools allowed, they took 19% longer. Yet afterwards they believed AI had cut their time by about 20%. That is one small study of expert developers on familiar code, but it is a good reason to measure your own results rather than trust the feeling.
These are global figures. We found no reliable survey of how Kenyan developers use AI.
Three kinds of tools
- Chat assistants (ChatGPT, Claude, Gemini, Copilot chat): you paste code or a question and get an answer.
- In-editor assistants (for example GitHub Copilot in VS Code): suggestions and chat inside your editor, with your files as context.
- Agents (for example Claude Code, OpenAI Codex, Copilot agents): they read the repository, edit files and run commands, then show you what changed.
Plans and prices change often, so check each vendor's own page before you choose.
Six tasks where AI helps, with prompts
At FuKazee we teach the four Cs: Clarity (what exactly you want), Context (what the tool cannot guess), Constraints (limits and what to avoid) and Customisation (who will use the result and where).
1. Understanding code you did not write
Explain what this module does, in plain English, for a developer new to the codebase. Context: it is part of a [Django] app that handles [order payments]. List: its inputs and outputs, side effects (database writes, network calls), and anything that looks risky or unusual. Do not suggest changes yet. Code: [paste]
Check: trace one real request through the code yourself. Explanations can sound right and miss a side effect.
2. Writing tests
Write unit tests for the function below using [pytest]. Cover: normal cases, edge cases (empty input, zero, very large values, wrong types) and the error paths. Constraints: do not change the function; no network calls; one assertion idea per test. Then list any behaviour that seems undefined, as questions for me.
Check: run them, and make at least one fail on purpose. A test that can never fail tests nothing.
3. A second reviewer on a pull request
Review this diff as a careful senior developer. Look for: logic errors, missing error handling, security problems (injection, missing authorisation checks, secrets in code), performance issues and unclear names. Rank findings by severity and quote the lines. Do not comment on style unless it hides a bug. Diff: [paste]
Check: each finding against the code. Keep a human reviewer for every merge.
4. Documentation
Write a README section for this service: what it does, how to run it locally, environment variables needed (names only, never values), and how to run the tests. The reader is a new team member. Use only what is in the files I paste; mark anything you cannot confirm with [CHECK].
Check: follow your own README on a clean machine. If it does not work, fix it.
5. Debugging from an error
Here is an error and the code around it. Give the three most likely causes, ranked, and for each, how I can confirm or rule it out. Do not rewrite the whole file. Error and code: [paste, with keys, tokens, customer data and internal URLs removed]
Check: confirm the cause before you apply any fix.
6. Learning a new language or framework
I know [JavaScript and Express] well. Teach me [Go] by comparing each idea to what I already know. Start with project structure, error handling and concurrency. After each part, give me one small exercise and wait for my answer before moving on.
Check: the official documentation for anything you will ship. Tutorials from AI can be out of date.
Security, licences and data: the rules that matter
Keep secrets out. API keys, database passwords, payment API credentials such as M-PESA keys, and .env files never go into a prompt. Strip them before you paste anything.
Keep personal data out. Customer names, phone numbers and ID numbers fall under Kenya's Data Protection Act, 2019, regulated by the Office of the Data Protection Commissioner (ODPC). Use fake or masked data in prompts and test fixtures. In September 2026 the ODPC published a draft notice proposing extra categories of sensitive personal data. It was still a draft as of November 2026, so watch for updates. Our post on the Data Protection Act and AI explains more.
Check every package name. AI tools sometimes suggest packages that do not exist. A study for the 2025 USENIX Security Symposium tested 16 models on 576,000 code samples. On average, at least 5.2% of suggested packages from commercial models, and 21.7% from open-source models, did not exist. Attackers can publish harmful packages under such invented names. Look the package up in the official registry before you install it.
Watch licences. GitHub's documentation says Copilot can generate code that matches public code, and that matches typically occur in under 1% of inline suggestions. Depending on your settings, matching suggestions can be blocked or shown with the source and any licence found. Read that information and attribute, keep or remove the code as the licence requires.
If you build AI features, know the risks. The OWASP Top 10 for LLM Applications (2025 edition) lists prompt injection first, followed by sensitive information disclosure. "Excessive agency", giving an AI system more power than it needs, is also on the list. Read it before you connect an AI model to your users or your data.
Limit what agents can touch. Run agents on a branch or a copy, never with production credentials, and review every change before it merges.
Follow your company's code policy
Before you paste company code anywhere, ask:
- Which AI tools are approved, and on which plan?
- May source code leave our systems? Which repositories are off-limits?
- Is training on our data switched off?
- Who reviews AI-written code before it merges?
- Do our client contracts say anything about AI tools?
If you freelance, read your client's contract first. If your team has no policy, our AI acceptable use policy template is a starting point. For code that must never leave your machines, see our guide to running open-source models locally.
Questions people ask
Will AI replace software developers in Kenya?
The evidence above shows developers using AI heavily while still checking its work, because it is often close but wrong. The job is shifting towards reviewing, testing and deciding. Developers who read code well are likely to become more valuable.
Do AI coding tools make developers faster?
Not automatically. In the METR trial above, experienced developers on familiar code were slower with AI, even though they felt faster. Time a few real tasks with and without it before you decide.
Should junior developers use AI tools?
Yes, carefully. Use AI to explain and to quiz you, and write code yourself while you learn. If you cannot explain a line, do not commit it.
Can I use AI on my employer's code?
Only with tools your employer has approved. Pasting company code into a personal account can break your contract and leak secrets.
Which is better: Claude Code, Codex or Copilot?
It depends on your stack, your editor and your company's approved tools. Run the same real task in each and compare. Our guides to Claude Code and Codex explain the differences in plain words.
Sources
All sources last checked 14 November 2026.
- Stack Overflow, 2025 Developer Survey: AI: usage, trust and frustration figures.
- Becker and others (METR), Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity: 19% longer completion time with AI allowed.
- Spracklen and others, We Have a Package for You! (USENIX Security 2025): package hallucination rates.
- GitHub Docs, GitHub Copilot in IDEs: references to matching public code: matching public code and licence information.
- OWASP Gen AI Security Project, Top 10 for LLM Applications 2025: prompt injection, sensitive information disclosure, excessive agency.
- Office of the Data Protection Commissioner: mandate under the Data Protection Act, 2019; draft notice on additional sensitive data categories (11 September 2026).
- FuKazee, AI training for startups in Kenya: developer sessions on code and security.
Where FuKazee fits in
FuKazee trains tech teams and developers by arrangement. As our startups page explains, developers can have a separate session on coding assistants, testing, documentation and security habits, and we adapt after finding out what your developers already do. Our advice there is to treat AI-written code like a pull request from a new intern. For teams choosing Anthropic's tools, our Claude training page explains honestly when Claude Code is worth it.
Tell us what your developers build. Email lead@fukazee.com
Want guided support instead?
Not sure which programme fits? Tell us who you are and we will point you to the right path: corporate AI training, the ScholarShift Accelerator for postgraduates, or KCSE study plans and tuition.





