Detailed Analysis
A developer has built and shared a third-party tool called Produck that analyzes local Claude Code session history and generates a personality-style scorecard for how a given user works with the AI coding assistant. The tool reads the session logs that Claude Code already writes to disk, then scores the user across six measurable axes—planning, specification, steering, orchestration, taste, and velocity. The two highest-scoring axes determine which of 16 "archetypes" the user is assigned, complete with a rarity tier, gamifying what is otherwise a fairly dry log-analysis exercise. The creator frames this less as a novelty and more as a diagnostic instrument: the real utility, they argue, comes from the output afterward, where the tool generates a personalized set of rules intended to be pasted into a CLAUDE.md file, each rule tied back to a specific statistical finding from the user's own session data rather than generic best-practice advice.
This reflects a broader and increasingly visible trend in the Claude Code ecosystem: developers treating their interaction logs with the coding agent as a rich, underexploited dataset for self-improvement. Because Claude Code persists full session transcripts locally, power users have started building tooling around those files—not just for debugging or auditing, but for behavioral analytics about how humans collaborate with an AI agent. The example given in the post, where "rapid short feedback turns" correlated with a 53% correction rate versus 39% for longer turns, illustrates the kind of granular, person-specific insight this approach can surface: it's not telling users to write shorter prompts in general, it's showing them that their own iteration cadence produces measurably different error-correction rates. The creator's anecdote about testing it on a cofounder, whose "Standard-Bearer" archetype nudged him toward writing plans upfront and reportedly reduced rewrites, is offered as a small case study in the idea that CLAUDE.md—Anthropic's mechanism for persistent, project-level instructions to the model—can be populated not just with codebase conventions but with individualized workflow corrections derived from empirical usage patterns.
The privacy framing is notable and speaks to a growing sensitivity among developers who share coding-agent tooling: the post emphasizes that redaction happens locally before any data is uploaded and that reports default to private. This matters because Claude Code session logs can contain proprietary source code, internal file paths, API keys, and other sensitive material, so any third-party tool that ingests those logs invites scrutiny about what leaves the machine. Positioning the redaction step as local-first is a deliberate trust signal aimed at a technically sophisticated audience that is otherwise likely to be wary of uploading raw agent transcripts to an unfamiliar web service.
More broadly, this kind of project is emblematic of the ecosystem of grassroots, community-built tooling that has sprung up around Claude Code specifically because Anthropic exposed session data and the CLAUDE.md customization mechanism as accessible, hackable surfaces. Rather than waiting for Anthropic to ship built-in analytics or coaching features, independent developers are building meta-tools that treat the agent's own logs as a feedback loop for improving human-AI collaboration—effectively applying quantified-self and gamification techniques (archetypes, rarity tiers, percentile scores) to what is fundamentally a workflow-optimization problem. It's a small but telling data point in the larger story of AI coding assistants: as these tools become embedded in daily development work, the interesting frontier is shifting from "how good is the model" to "how do individual developers work with the model," and self-quantification tools like this one are an early, informal attempt to answer that question systematically.
Read original article →