Coding

An incident.io engineer asked Claude Code to diagnose slow API generation and implement improvements incrementally. The author reports an 18% runtime reduction, equivalent to 30 seconds, for approximately five minutes of personal work and $8 in Claude usage. Commands and workflow examples are included.

Microsoft's .NET team assigned simpler refactoring tasks to Copilot in dotnet/android. Over 28 days, 82.4% of 17 agent pull requests were merged. Harder dotnet/maui tasks had a much lower merge rate, showing the importance of task selection.

The Original Tamale Co. used ChatGPT to build and publish a farmers-market tracker in one afternoon. The operator had no coding experience. No comparative development cost is provided.

Rakuten engineers use Claude Code for tests, components, bug fixes and documentation. Rakuten reports average feature delivery falling from 24 working days to five. A separate seven-hour vLLM implementation reached 99.9% numerical agreement with its reference, rather than general code correctness.

Cisco engineers use Codex to review both human-written and AI-generated code. OpenAI reports review-time reductions of up to 50%. The customer example does not publish the sample size or measurement procedure.

Doctolib runs Claude Code agents through GitHub Actions for implementation, review, documentation and production checks. Its updated case reports 25% shorter change lead time between January and September 2026. This is an internal operational measurement; causal attribution to Claude is the company's interpretation.

After Git training, Grab designers used Cursor to implement interface corrections directly in code. Design managers reviewed hundreds of merged fixes, often on the same day. Faster delivery is reported, but no controlled time-saving comparison is provided.

Jiaming Shang used Copilot to generate a deployed typing practice website with English, Chinese Pinyin and 300 Tang poems. Source code and demonstrations are linked. The account documents repeated corrections and limitations rather than measured productivity gains.

Navan used self-hosted Cursor Cloud Agents to migrate eight applications and 15 libraries to Spring Boot 4 in about one week. A comparable earlier migration took over a month. These are company-reported project durations, not a controlled experiment.

Port engineers delegate small refactorings, bug fixes and TODO items to Copilot while developing other features. Their retrospective reports better self-review and fewer minor review comments, without quantified savings.

A READYFOR engineering survey documents Claude Code use for coding, Git operations and pull requests. Among 12 respondents, 66% reported saving one to two hours daily, while 42% reported more fatigue. These are self-reported savings rather than controlled productivity measurements.

Vanta uses Claude to turn failed compliance tests into tailored remediation instructions, Terraform templates or cloud commands. Its internal evaluation reported roughly 15% better performance than alternatives on Terraform remediation. Customer time savings are described qualitatively.

Aldo used Copilot's coding agent to add optional header images to an existing Markdown blog. The post links the agent session and describes compatibility and testing requirements. No task-specific time saving is measured.

GitHub assigns Copilot issues to remove obsolete feature flags, conditional branches and associated tests across its platform codebase. Engineers review the proposed pull requests. The article describes reduced maintenance work but gives no savings measurement for this task.

During a healthcare software project, Microsoft ISE used Copilot to follow development guides, create required files and identify missing setup steps or tests. The team reports easier onboarding and time savings but provides no controlled measurement for this workflow.

Okta engineers used Copilot for routine coding and testing. A study involving 74 survey participants found improved motivation and perceived skills, but no significant change in pull requests, code volume or cycle time. Later results for Claude Code are separate.

Simon Willison used Claude to build a Python utility converting Claude Code sessions into navigable HTML. He publishes source code, example output and development transcripts. Codex handled a separate API reverse-engineering step. No quantified development-time saving is reported.

IMDEX used Cursor to rebuild fragmented geological data applications into an integrated platform delivered in eight months. Its comparison with two to three years and 20 additional engineers is a counterfactual estimate, not a measured alternative implementation.

National Australia Bank's merchant services team used Cursor to ship a hardware-independent payment application in three weeks. The original scope was four months. The story reports this delivery against a planning estimate, without a parallel comparison project.

Australian developer Geoffrey Huntley ran Claude Code in repeated loops for three months to build Cursed, a programming language using Gen Z vocabulary. The compiler and demonstrations were published. This was a learning experiment with incomplete components, without measured labor savings.

Two Nokia engineers used Cursor to analyze more than 50 million lines of code in two weeks and prepare an architectural decomposition plan. The comparison with a dozen specialists working for months is an estimate, rather than an observed control project.