Recommended Free Tools
AI coding tools can help developers complete more tasks in some settings, yet slow experienced maintainers in others—and a small learning trial found weaker immediate comprehension among participants who used AI. Those findings are not contradictory so much as answers to different questions. More generated code or tasks completed is not the same measure as understanding, debugging ability, or long-term engineering skill.
What is the craft-versus-output divide?
“Output” can mean code generated, tasks completed, or repository issues closed. These measures describe activity or throughput; they do not, on their own, establish that the code is correct, maintainable, or valuable over time.
As an Amazon Associate I earn from qualifying purchases.
“Craft” is a useful editorial shorthand for the work behind dependable software: understanding a problem and its constraints, making design choices, testing behavior, tracing failures, and maintaining code someone else can safely change. The studies discussed here measure only parts of this picture. They do not provide a single score for engineering mastery.
What do the studies actually show?
Three recent studies point in different directions because they involved different developers, tasks, tools, and outcomes. Their percentages should not be combined into one estimate of AI’s effect on software productivity.
#1 Best Overall
| Study and participants | Work and setting | Outcome measured | Finding |
|---|---|---|---|
| Microsoft Research, June 2025: 4,867 developers in a pooled analysis of three randomized field experiments at Microsoft, Accenture, and an anonymous Fortune 100 company. | Company deployments of an AI coding assistant. The experiments were noisy; the report does not make this a test of every kind of software work. | Completed task counts. | The analysis reported 26.08% more completed tasks among AI-tool users, with a standard error of 10.3%. This is not a finding about code quality, long-term value, or a universal productivity gain. |
| METR, July 10, 2025: 16 experienced open-source contributors, assigned 246 issues across AI-allowed and AI-disallowed conditions. | Contributors worked in their own mature repositories, which averaged more than 22,000 stars and one million lines of code. The trial used early-2025 tools. | Time to complete issues. | Developers took 19% longer when AI was allowed in this study. Beforehand, they expected a 24% speedup; after the work, they still believed AI had sped them up. METR describes this as a snapshot of one setting, not a representative result for most software work. |
| Anthropic, January 29, 2026: 52 mostly junior software engineers who knew Python but were unfamiliar with Trio. | A randomized task in which participants learned and used the Python library. The report compares AI-assisted work with hand-coding. | Score on an immediate comprehension quiz after the task; time to finish. | The AI group averaged 50% on the quiz versus 67% for the hand-coding group (Cohen’s d=0.738, p=0.01). AI users finished about two minutes faster on average, but that time difference was not statistically significant. The largest quiz-score gap was on debugging questions. |
The results are not direct replications. The Microsoft experiments counted tasks completed across company deployments; METR timed experienced contributors changing large repositories; Anthropic assessed near-term understanding after a learning task. A task-count increase cannot tell us whether someone learned an unfamiliar library, and a slowdown in mature codebases cannot settle how AI affects routine work in a company.
Does AI help developers learn, or get in the way?
Anthropic’s trial found an immediate comprehension difference in one specific situation: junior engineers using AI while learning a library they did not know scored lower on a quiz shortly afterward than those who hand-coded. Debugging questions showed the largest gap. That is a reason to take the learning trade-off seriously, not proof that AI permanently weakens developers’ skills.
Rank #2
The assessment followed a short task. It does not establish whether the quiz difference predicts retention, later debugging performance, code ownership, or professional competence. Anthropic explicitly identifies longer-term skill development as unresolved. The other two studies measured task completion or time, not lasting skill.
Outdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchWindows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstallAnthropic’s report says, “Our results suggest that incorporating AI aggressively into the workplace, particularly with respect to software engineering, comes with trade-offs.” The qualification matters: this was a small, task-specific trial measuring immediate comprehension, not a longitudinal study of engineers’ careers.
Rank #3
Why can AI speed up one kind of work and slow another?
- Familiarity changes the work. Asking for help on a known task is different from relying on a tool while learning an unfamiliar concept. The latter can produce a working result without requiring the same amount of deliberate practice.
- Repositories impose context costs. A change in a mature project may require locating the right code, understanding local conventions, tracing dependencies, and checking edge cases. Generating a patch does not remove those obligations.
- Outcomes answer different questions. Task counts, completion time, and quiz scores are not interchangeable. None alone captures correctness, maintainability, or future ability to solve problems independently.
- Expectations can differ from measured performance. In METR’s trial, participants expected a speedup and continued to feel faster even though measured issue completion took longer. Subjective ease is not a substitute for checking outcomes.
These explanations help make sense of the evidence, but they do not establish which tasks or developers will benefit most. The studies do not identify one universally productive amount of AI use.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What does AI interaction style tell us about preserving craft?
Anthropic observed stronger quiz performance in usage patterns that included requests for explanations and conceptual questions, and weaker performance in patterns involving heavier delegation. The report treats this as qualitative analysis: it does not establish that those interaction styles caused the score differences or prove that a particular prompting method preserves skill.
Rank #4
Still, developers who want to learn as well as finish a task can use these observations as prompts for deliberate practice:
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
- Before asking for a solution, write down what you think the relevant behavior or design constraint is.
- Ask the tool to explain an unfamiliar API, trade-off, or failure mode, then check that explanation against the code and tests.
- Try to predict what a test or change will do before running it; investigate mismatches rather than accepting a plausible-looking answer.
- Keep some debugging and code-review work active: trace the failure, inspect the proposed patch, and explain why it fixes the issue.
- For consequential changes, evaluate the result with the project’s tests and review standards, not by how quickly code appeared.
These are practical ways to keep understanding in the loop, not interventions proven by the trial. The evidence does not yet establish a workflow that reliably preserves mastery across roles, tools, and tasks.
How should teams judge whether AI is improving engineering work?
Teams should define the result they want before treating faster generation as success. If the goal is throughput, track completed work in a consistent setting. If the goal is reliable software, also examine correctness, review findings, regressions, and the effort required to maintain changes. If the goal includes developing engineers, assess whether people can explain, debug, and extend the code later.
Compare like with like: the same kinds of tasks, comparable experience levels, clear AI-use conditions, and outcomes measured over an appropriate period. A short-term completion metric can reveal speed or volume, but it cannot answer whether developers are building durable judgment. The open question is whether near-term comprehension differences translate into lasting changes in debugging skill, code ownership, or professional competence.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
Free tools Windows power users keep installed
One-click scans. No signup required.
