Learning Mode plugin makes Claude refuse to write code, guides learner instead

Muhammad El Sayed has created Learning Mode, an agent-agnostic guided-learning skill available as a plugin for Claude Code and Codex. The skill fundamentally changes how these assistants respond to code-learning requests: instead of writing code, they guide learners through a structured problem-solving cycle while learners own the load-bearing reasoning.

The plugin's workflow begins by determining intent—quick answer, guided learning, or direct delivery. In guided learning mode, the assistant defines one clear learning outcome, checks the smallest relevant prerequisite, then enters a cycle of predict, attempt, feedback, and retry. When the learner's attempt fails, the assistant diagnoses and revises. When it passes, the assistant validates whether the learner can explain the mechanism and handle a transfer question (a changed example or counterexample). If not, the assistant climbs a four-rung hint ladder; if the learner still lacks the underlying schema, it shows one fully worked analogous problem, then returns to the original target. Only when the learner can explain and transfer the concept does Learning Mode mark progress—optionally saving it to a journal at ~/learning-journals/<date>-<topic>/journal.md.

The boundary between learner and assistant ownership is explicit: the learner owns the target skill—state models, configuration, architectural decisions, test logic. The assistant may provide unrelated scaffolding (empty harnesses, project setup) to remove friction, but the load-bearing work belongs to the learner. For visual topics (data structures, flows, hierarchies, state diagrams), Learning Mode prefers diagrams, tables, and interactive models, asking the learner to predict, sketch, trace, or annotate rather than passively read explanations. For each substantial concept, the learner performs one authentic application (implementation, design, prediction, diagnosis, or simulation), compares one plausible alternative, and explains what the concept improves, what it costs, and when to avoid it.

The skill distinguishes validation from verification. Validation checks whether the learner's work satisfies the outcome in the tested conditions—restating success criteria, having the learner propose important test cases, running a small test or probe, and comparing results. Verification checks whether technical claims are accurate and current: version-specific APIs, service limits, protocol requirements, and production practices are verified against authoritative sources (documentation, specifications, machine-readable contracts). The skill marks claims as verified, inferred, or unconfirmed when provenance matters, but skips live lookups for stable foundational knowledge and incidental facts that cannot change the outcome.

Installation is straightforward. For Claude Code, users run /plugin marketplace add melsayedx/learning-mode and then /plugin install learning-mode@learning-mode-marketplace. For Codex, they add the repository as a marketplace and install via codex plugin add learning-mode@learning-mode-marketplace. Journaling is opt-in; the assistant asks once whether to enable it for all Learning Mode sessions or only the current project, then stores journals using durable memory, project instructions, or a preference under the journal root. The learner can pause, disable, or change the journal location at any time. If the learner requests direct code delivery during guided learning, the assistant confirms before switching: "You want me to leave Learning Mode and deliver the solution directly. Should I switch?" However, guided struggle pauses immediately if delay could cause security risk, data loss, financial harm, privacy exposure, or an active production incident; the assistant then gives protective guidance and asks whether to switch to direct remediation. The plugin is packaged with a behavioral core (SKILL.md), agent-specific metadata for Codex and ChatGPT, optional documentation-verification wrappers, and a journal template.

Learning Mode addresses a fundamental tension in AI-assisted learning: LLMs (AI systems that understand and generate text) are so capable at writing code that learners can obtain working solutions without understanding them. El Sayed's skill flips the default by making the assistant refuse to write code in learning contexts and instead guide the learner through a diagnostic cycle. The approach is grounded in established learning science—the skill explicitly combines diagnostic checkpoints, guided inquiry, graduated hints (a four-rung ladder climbed when the learner is stuck), self-explanation, and transfer checks (asking the learner to handle a changed example or counterexample). The boundary between assistance and ownership is precise: the learner owns the target skill, and the assistant removes incidental friction. This prevents the assistant from doing the work while maintaining pedagogical scaffolding.

The skill's architecture adapts to intent. It first asks whether the learner wants a quick answer, guided learning, or direct delivery, then activates only for explicit active-learning requests ("Teach me," "Guide me," "I want to practice"). Routine questions and delivery requests bypass Learning Mode entirely. For visual topics—data structures, state machines, architectural patterns—Learning Mode prefers diagrams, tables, and flows, asking the learner to predict, sketch, or trace rather than passively consume. Verification is built in: the skill checks version-specific APIs, current service limits, security requirements, and surprising claims against authoritative sources, marking them as verified, inferred, or unconfirmed. Journaling is opt-in and preserves progress across sessions if enabled.

No comments yet. Be the first to share your thoughts!