Why 72% of Opus 4.7 Code Is “Highly Reusable”
We ran a small evaluation: for each Opus 4.7 repo, feed a structured summary into an LLM and ask it to grade the reuse potential (high / medium / low). Out of 413 repos evaluated so far:
| Reuse potential | Count | Share |
|---|---|---|
| High | 301 | 72.9% |
| Medium | 122 | 29.5% |
| Low | 21 | 5.1% |
Almost three out of four Opus 4.7 repos are “highly reusable” according to the grader. That’s an unusual density. For reference, random GitHub dumps grade in the 20–30% high-reuse range under the same rubric.
Why the number is so high
1. Scaffolds are inherently reusable
A large share of the corpus consists of project scaffolds — an initial prompt spits out a Next.js app, a FastAPI service, a CLI tool. Scaffolds by definition aim to be reusable starter code.
2. A consistent stack amplifies reuse
When the same Opus 4.7 repo uses shadcn + Tailwind + Next.js App Router + Zod + TanStack Query (as most do), any component or hook lifts out cleanly into another Opus 4.7 repo.
3. Small functions lift out easier
Median function size is 9 lines, p90 is 64. Short functions are easier to reuse than monolithic ones. The long tail of 1,000-line mega-functions is where reuse dies, and Opus 4.7 has a thin tail.
4. The grader knows the stack
The grader LLM was trained on similar data. It recognizes the shadcn/Next/Zod/Tailwind pattern and rates familiar code higher. This is partly a rubber-stamp effect — worth noting honestly.
Reference quality distribution
A separate grading axis asks: “If I were to learn from this repo, how high-quality a reference is it?”
| Bucket | Count |
|---|---|
| 90–100 (excellent) | 2 |
| 75–89 (good) | 102 (24%) |
| 60–74 (ok) | 335 (81% when combined with above) |
| 40–59 (weak) | 2 |
| 0–39 (poor) | 3 |
Notice the extreme concentration in the 60–74 band. Opus 4.7 writes “solid ok” code. It rarely peaks at 90+. It rarely drops below 60. That consistency is itself a signal — generative AI averages out to a stable quality profile.
Top patterns the grader identified
The grader extracts key patterns for each repo. Counting across the 413 sample:
| Pattern | Repos mentioning |
|---|---|
| JWT authentication | 55 |
| Responsive design | 54 |
| RESTful API endpoints | 44 |
| RESTful API design | 44 |
| MVC architecture | 36 |
| Static site generation | 33 |
| RESTful API (generic) | 33 |
| Dockerized deployment | 29 |
| Docker containerization | 19 |
| SEO optimization | 18 |
| Markdown content | 17 |
| CI/CD pipeline | 17 |
| ORM-based database access | 12 |
| React component architecture | 11 |
| Component-based architecture | 11 |
| Unit and integration tests | 10 |
What it means for training data
72% high-reuse is a strong endorsement for using this corpus as a positive signal in fine-tuning. You don’t need heavy curation — the population is already biased toward reusable patterns.
The caveat: you also inherit the stack bias (shadcn, Next, Tailwind). If your downstream use case is different, you need to filter or balance.
See also: The Top 30 Libraries Claude Opus 4.7 Actually Imports.