Developer Builds Personalized Go Tutor and 24/7 AI Assistant Using Claude Code

The system, dubbed Algotutor, leverages Claude Code’s session‑memory feature to turn a local development directory into a persistent teaching agent. When a user launches Claude Code inside the cloned repository and types the command `train`, the model ingests the entire folder hierarchy—including a `progress.json` file that records skill levels across 32 algorithmic concepts, the source files for prior solutions, and any review cards the user has generated. Because Claude Code retains the full context of the session, it can reference these artifacts on the fly, selecting the next problem that targets the learner’s weakest area rather than serving a random challenge (gentic.news, Apr 15).

The curriculum is encoded as a structured directory tree, with each node representing a specific topic such as arrays, strings, dynamic programming, or system design. Claude Code parses the `progress.json` to compute a proficiency score for each node, then uses a simple heuristic to prioritize concepts whose scores fall below a configurable threshold. When a problem is presented, the model also generates a set of spaced‑repetition flashcards based on the user’s solution and any mistakes recorded during the attempt. These cards are stored alongside the problem files and later resurfaced by Claude Code in subsequent sessions, creating a closed feedback loop that mirrors proven educational techniques (gentic.news).

The underlying architecture mirrors the design of the “Friday” AI assistant, another Claude Code‑only project that runs continuously on a personal machine. Both tools rely on a lightweight Flask server that exposes a memory API backed by SQLite; this API stores conversation history, embeddings, and the `progress.json` data. In the Friday showcase, the memory server also hosts self‑evolving subsystems—skill acquisition, daily reflection, and preference learning—that run as scheduled cron jobs and feed insights back into Claude Code without any model fine‑tuning (missingus3r GitHub). Algotutor adopts the same pattern, using the memory layer to persist the learner’s state across restarts, thereby eliminating the need to re‑seed the model with a summary of past performance each time the tutor is invoked.

Claude Code itself is accessed via Anthropic’s CLI on the $100‑per‑month Max Plan, which provides a single‑model endpoint capable of both code generation and natural‑language reasoning. The Algotutor workflow requires only three steps: clone the repository, start Claude Code in the directory, and issue the `train` command. No additional cloud infrastructure, fine‑tuned models, or external orchestration frameworks are needed, making the setup comparable to the “Friday” assistant’s claim of running entirely on a standard Windows, Linux, or macOS host (missingus3r GitHub). This minimal dependency stack underscores the feasibility of building sophisticated, stateful AI applications using only a persistent LLM session and a modest local compute environment.

From a technical standpoint, the key innovation lies in treating Claude Code as a mutable agent rather than a stateless query engine. By embedding the learner’s curriculum, progress metrics, and review artifacts directly within the file system, the system exploits Claude’s ability to maintain a coherent mental model across an entire interactive session. This approach sidesteps the typical “prompt‑engineering” overhead associated with large‑language‑model tutoring, where each new query must re‑establish context. Instead, the model continuously refines its internal representation of the user’s abilities, enabling real‑time adaptation and personalized problem sequencing—a capability that, according to the gentic.news report, distinguishes Algotutor from conventional coding‑challenge platforms.