What 1,394 Git Commits on One Project Teach You About AI-Assisted Development

The Setup

Most conversations about AI-assisted development focus on the wrong metric. They fixate on whether AI can write a function, generate boilerplate, or autocomplete a line of code. What they rarely examine is what happens over the lifetime of a real production build — thousands of commits, hundreds of files, months of sustained execution. The single-task demo is not the hard part. The hard part is whether the workflow holds together across 194,954 lines of code.

The conventional approach to evaluating AI coding tools is benchmarks: pass rates on HumanEval, accuracy on LeetCode problems, speed-to-first-output comparisons. These tell you almost nothing about production software development. A developer who can solve a coding puzzle in 30 seconds may still ship a broken database migration. A tool that generates perfect function signatures may still produce architectural drift that takes weeks to unwind.

What actually matters is the commit record — the full, unedited trail of decisions, corrections, and output over time. Commits do not lie. They show what was built, what was reworked, what was reverted, and at what rate the system stabilized. The problem is that almost nobody publishes this data at the project level, because most production repositories are private and most teams do not instrument their own workflow for analysis.

What the Data Shows

GitHub's 2024 Octoverse report documents that AI-assisted development is accelerating across the platform, with Copilot-assisted developers completing tasks measurably faster than unassisted peers. LinearB's engineering benchmarks place median cycle time for a pull request at approximately 6.3 days across the industry, with coding time — the period a developer is actively writing — averaging around 2.5 days per PR. Haystack Analytics reports that high-performing engineering teams sustain roughly 60-80 PRs per developer per quarter, which translates to a handful of meaningful commits per working day for most engineers. Research from Sieber & Partners, analyzing 3.5 million commits across 47,000 developers, places the median developer at approximately 2 commits per day, with leading contributors reaching 6-14 per day.

PRJ-01 provides a controlled dataset against this backdrop: 1,394 commits across 74 active development days, producing 194,954 lines of custom code (PHP, Blade, SCSS, JavaScript, CSS) in a Laravel 10 application. The project spanned 115 calendar days from October 8, 2025, to January 31, 2026, with a single primary operator (Michael George Keating) responsible for 86.8% of commits, supported by CON-01 at 10.7% and AI-assisted scaffolding at 2.5%. The average across all active days was 18.8 commits per day — roughly 9x the industry median documented by Sieber & Partners.

But the average masks the trajectory. PRJ-01's velocity phases tell a more useful story:

Phase	Dates	Commits	Active Days	Commits/Day
Phase 1 — Foundation	Oct 8–31	73	16	4.6
Phase 2 — Iterative	Nov 1–27	108	17	6.4
Phase 3 — Acceleration	Dec 21–31	241	10	24.1
Phase 4 — Peak Sprint	Jan 1–6	369	6	61.5
Phase 5 — Feature Expansion	Jan 7–13	213	7	30.4
Phase 6 — Dashboard & Funnels	Jan 14–20	231	7	33.0
Phase 7 — Polish & Launch	Jan 21–31	159	11	14.5

That is a 13.4x output multiplier from Phase 1 to Phase 4. The peak single day hit 89 commits on January 1, 2026. The peak sustained sprint: 369 commits in 6 days. These are not micro-commits or auto-generated noise — each commit maps to a tracked feature change, bug fix, or integration step within a system that ultimately processed 616,543 leads, 75,125 transactions, and 958,937 contact points in production.

The rework data is equally instructive. Overall rework rate was 31.3% (436 of 1,394 commits), but the trajectory moved from 45.2% in Phase 3 down to 27.0% by Phase 7 — a 40% reduction in rework as patterns consolidated. Total reverts across the entire build: 8, or 0.6% of all commits. The product bug rate was 18.3% (255 of 1,394 commits), while integration friction — the cost of wiring to external systems like Stripe, Shopify, and Konnektive — accounted for just 21 commits (1.5%).

For context on scale: SQLite is approximately 155,800 lines of code. WordPress is approximately 160,636. PRJ-01, at 194,954 lines, sits between these two — built in 74 active days by a team that was functionally one operator with AI tooling and a single contractor handling 10.7% of the work.

How It Works

The commit pattern is not random, and it is not simply "type faster." What the phase data reveals is a compounding workflow. Early phases are slow because the operator is establishing foundational patterns — authentication, database schemas, service architecture, the design system. Every subsequent feature builds on top of those patterns rather than inventing new ones.

AI tooling accelerates this compounding effect. Once a pattern is established (for example, how a new controller, service, and Blade template interact in this specific codebase), AI can reproduce that pattern at near-zero marginal cost for each subsequent feature. The operator's role shifts from writing every line to directing the pattern application and handling the judgment calls — which integration to use, how to structure the data model, where the business logic belongs. This is why the rework rate drops over time: the patterns stabilize, the AI applies them more reliably, and the operator catches deviations faster because the codebase is familiar.

The 64 scheduled console commands, 135 database tables, 112 Eloquent models, and 20 external integrations in PRJ-01 were not built in a single sprint. They accumulated across 7 velocity phases, each building on the infrastructure laid by the previous one. The peak sprint of 61.5 commits per day in Phase 4 was only possible because Phases 1 through 3 had established the architectural scaffolding. This is the mechanism that commit-level data exposes and that demo-based evaluations miss entirely.

What This Means for Engineering Leaders

If you are evaluating AI-assisted development for your team, commit-frequency benchmarks from tools like Copilot, Cursor, or Claude are only useful when measured across a full build lifecycle — not in isolation. A 40% speed improvement on a single function means nothing if rework doubles. A 2x commit rate means nothing if revert rates climb alongside it.

The PRJ-01 dataset demonstrates that sustained AI-assisted development at 18.8 commits per day (with peaks above 60) is achievable on a production-grade application with a 0.6% revert rate and a downward-trending rework curve. The actual cost of the build was $16,800 in operator sweep cost plus $3,184 in AI tooling — against a market replacement value of $780,000 to $1,560,000 benchmarked against US mid-market rates (FullStack 2025 Price Guide, Keyhole Software 2026 Benchmarks). That is a 46x to 93x cost multiple. The data exists. It is git-verified. It is the kind of evidence this conversation has been missing.

---

Related: C1_S17, C1_S18

References

1. GitHub (2024). "Octoverse Report." AI-assisted development acceleration and Copilot-assisted developer productivity metrics.
2. LinearB (2024). "Engineering Benchmarks." Median cycle time of approximately 6.3 days per pull request.
3. Haystack Analytics (2024). "Developer Productivity Report." High-performing teams sustaining 60-80 PRs per developer per quarter.
4. Sieber & Partners (2024). "Commit Velocity Analysis." 3.5 million commits across 47,000 developers (median: 2 commits/day, leading contributors: 6-14/day).
5. FullStack (2025). Software development price guide. Mid-market US rate benchmarks.
6. Keyhole Software (2026). Software development cost benchmarks.