Siftly Wants to Train Human Judgment in AI-Assisted Code Review

Most AI-powered code review tools function as an automatic filter: the model flags issues, the developer accepts or dismisses them. The workflow makes sense, but it has a side effect rarely discussed: over time, the reviewer's own judgment atrophies. Siftly has published an announcement that points in exactly the opposite direction.

According to their announcement page, referenced this week on Hacker News, Siftly's proposal is not that AI tells you what's wrong with your code. It's that AI helps you develop the instinct to spot it yourself.

What Siftly proposes exactly

The technical detail available in the announcement is still limited—the tool appears to be in very early stages—but the conceptual approach is clear: create a deliberate practice environment where the developer faces code snippets with real or simulated problems, attempts to identify them, and then receives feedback on their reasoning, not just whether they got it right or wrong.

It's a model closer to clinical skills training or case-based learning than the typical LLM-powered linter. The AI acts as a tutor that explains the why behind each review decision, not as an oracle issuing verdicts.

Why this approach makes sense now

Context matters. In 2026, with tools like Claude Code integrating sub-agents and MCP servers capable of auditing entire repositories, the temptation to completely delegate code review is higher than ever. Small teams in particular are starting to skip comprehensive human review rounds because the opportunity cost seems too high.

The problem is that code review isn't just spot quality control. It's one of the main mechanisms by which technical knowledge transfers between developers on a team. If that process is fully automated, the implicit learning that comes with it is lost too.

Siftly seems to be betting there's a market—probably junior developers and teams wanting to maintain that collective muscle—for whom it makes sense to invest time in actively training that judgment.

Who it's useful for

The proposal makes more sense for some profiles than others:

• Junior developers who still lack the pattern recognition needed to spot subtle architecture, security, or performance issues.
• Teams that have adopted AI heavily and want to ensure their human reviewers can still question what the model proposes.
• Educators and bootcamps seeking ways to make code review more pedagogical and less mechanical.

For a senior with ten years reviewing production code, the immediate value is less obvious. Though even there, a deliberate practice environment in specific areas—cryptography, concurrency, security patterns—could have occasional utility.

What we still don't know

The announcement doesn't detail which model or models Siftly uses under the hood, whether it offers integration with existing workflows like GitHub or GitLab pull requests, or what its business model is. The signal on Hacker News was modest—one point, no comments at publication time—suggesting the tool hasn't yet had significant exposure.

It's also unclear whether the exercise corpus is synthetically generated, hand-curated, or extracted from real repositories. That decision has important implications for training quality: the most interesting bugs typically live in real contexts, not textbook examples.

---

Siftly's approach is refreshing precisely because nothing in its proposal depends on AI being smarter than the developer. It depends on the developer wanting to be better. That's a more modest bet and, probably for that reason, more sustainable. We'll see if the product lives up to the idea.