Claude Harness to Avoid Getting Trapped in Junior Developer Purgatory

A free, open-source harness for Claude Code aims to stop developers from outsourcing the learning that keeps their skills alive. Piet Hanegraaf built the tool — documented at piethanegraaf.github.io's report — after a university diagnostic test exposed a gap between his AI-assisted practice scores and what he could actually do alone.

This article covers how the harness works, who it targets, and what the underlying research says about AI's effect on skill development. It is relevant to students using Claude Code for coursework, junior developers working alongside AI agents, and educators designing programming courses where AI access is permitted.

01Guard hook blocks Claude at learning levels L1 through L3

The harness introduces a skill-level system running from L0 (demonstrate) to L4 (automated). At levels L1 through L3, a hook actively blocks Claude Code from writing, deleting, building, or testing the developer's code. Help arrives in graduated rungs: a question first, then a pointer to where to look, then a rough sketch — each step released only after the learner makes a visible attempt.

A task is not counted complete until the learner explains, in their own words, why the solution works. Any override a developer grants to Claude during a blocked session is logged automatically. Everything stays local: data is stored in a .learning/ folder inside the project directory and the hooks make no network connections.

02Skill delegation and spaced maintenance keep growth continuous

Once a skill reaches L4, Claude can handle it immediately and the developer supervises rather than codes. The freed attention goes to whatever the learner is still building.

That is not the end: at growing intervals, the next time an L4 skill appears in real work, the developer does that part themselves. If it goes wrong, the skill drops a level and the cycle restarts.

The harness ships with two included skill maps: an Object-Oriented Design map in C# aimed at Dutch HBO-ICT students, and a junior developer example for C#/.NET. For C#/.NET, task logging runs automatically from dotnet test; for other test runners, developers log tasks manually. Any other language or course can plug in a custom skill map stored under harness/maps/<id>/.

03Bastani 2025 and Shen 2026 trials quantify the dependency risk

Hanegraaf built the tool after noticing a pattern that two published trials later confirmed. In a high-school mathematics study by Bastani et al. (2025), unlimited AI access raised practice scores by 48% but cut exam scores by 17% once the AI was removed. A separate short randomised trial by Shen and Tamkin (2026) found that engineers who learned a new programming library with AI scored 50% on a follow-up quiz, against 67% for those who learned without it.

The harness documentation frames the junior developer version of the same risk as "junior developer purgatory": an agent that completes work before the developer can do it themselves gets the task shipped but leaves debugging, code reading, design judgement, and AI-output verification undeveloped. Section 2.12 of the linked research covers this dynamic directly. Whether the harness itself reverses the effect has not yet been tested in a randomised trial; Hanegraaf explicitly invites researchers to run one.

04Students, junior developers, and educators each get a separate entry point

For students, the harness prevents Claude Code from taking over practice sessions on coursework. The included C# OOP skill map covers classes, inheritance, UML, and design patterns as defined by the HBO-ICT curriculum in the Netherlands. Whether using AI is permitted at all remains a course-policy question, not one the harness answers; every exception a student grants to Claude is logged for transparency.

Junior developers in a team repository can configure the harness so that code belonging to no tracked skill is freely written by the AI, and only the skills still under development are protected. Educators and researchers get dedicated documentation sections — For educators and For researchers respectively — with the research track inviting formal experimental testing. A live dashboard showing progress per microskill, with real data and a demo mode, is also available.

05Installation, licensing, and Claude plan requirement

The harness installs into any project folder with a single command and keeps all learning data out of git via the .learning/ directory. If a developer chooses to publish a dashboard, a filtered export runs through a whitelist of fields and strips names, paths, and email addresses before anything leaves the folder.

The code is open source under AGPL-3.0. Documentation, texts, and data are licensed under CC BY-NC 4.0.

Using the harness requires access to Claude Code itself, either through an Anthropic Claude plan or via the API — there is no bundled access. Hanegraaf notes that the harness and its documentation were built with help from Claude, and that all learning data and dashboard reflections are the learner's own.