Introducing VidByte Learning Skills

VidByte Learning Skills are prompt-level tools for people who use AI harnesses every day and still want to get sharper. They run inside tools like Codex, Claude Code, Cursor, Gemini CLI, Hermes Agent, and OpenCode, then add deliberate learning pressure at the exact moment the user is working.

The goal is not to slow the workflow down. The goal is to stop the quiet slide into autopilot. When an agent answers every question, chooses every tradeoff, and cleans up every mistake, the user can finish more tasks while understanding less of the work. Learning Skills push the user back into the loop.

AI makes it easy to outsource the parts of work that used to force understanding: framing the problem, naming the assumption, comparing alternatives, and explaining why a decision is right. Over time, this creates over-reliance. The user stops reaching for their own judgment because the model is faster, more confident, and always available. The skill of thinking through a problem atrophies not because the person is less capable, but because the environment never demands it.

Dependency is the natural outcome of frictionless AI. When every question gets an immediate answer, the user learns to wait rather than wrestle. They delegate decisions they used to own, accept reasoning they did not inspect, and accumulate output they cannot defend. VidByte Learning Skills push against this by inserting small, deliberate moments of resistance — asking the user to name the assumption, compare the alternative, or retrieve the concept before the model supplies it. The goal is not to slow the work down. It is to keep the user thinking so the work stays theirs.

The first set focuses on moments where the user is most likely to drift. A reasoner skill responds to a question with a small set of clear questions instead of a direct answer. A why skill occasionally asks why the user is choosing a path, especially when the prompt hides a weak assumption. A do-not-repeat skill notices repeated confusion and turns it into one or two reinforcement questions.

The same pattern extends to /question for deeper explanations, /retain for a longer retention exercise, /anti-passive for detecting excessive model decision-making, /no-assumptions for forcing missing context into the open, /explain-away-others for comparing alternatives, and /mental-model for connecting new work to what the user already knows.

A Learning Skill is a portable instruction package that tells an AI harness when to intervene, how much friction to add, and what the user should do next. Some skills are explicit commands. Others run in the background and keep a tiny local memory of recurring mistakes, weak explanations, or topics that deserve retrieval practice.

The important detail is dosage. A skill should not nag every time the user asks for help. It should look for signals: repeated requests, vague delegation, unexplained decisions, curiosity about a concept, or a point where a question before the answer would be more useful than an answer after the fact.

End-to-end, using Learning Skills takes three steps. First, install them with one command: npx vidbyte-skills installs every available skill into your target harness directories. You can scope the install to a single harness with --platform: npx vidbyte-skills --platform claude-code,codex installs into Claude Code and Codex only. Run npx vidbyte-skills --dry-run first to preview exactly which files will land where before writing anything. Second, restart your harness or open a new session. The harness discovers the SKILL.md files in its skills directory and activates them based on the frontmatter description. Third, invoke a skill explicitly with a slash command, like /reasoner to get questions instead of answers, or /retain to generate a reinforcement exercise. Background skills like anti-passive run automatically during your session without needing to be called.

Learning Skills are not arbitrary friction. They are designed around well-studied cognitive principles that explain why people learn more when they work harder at the right moments. Each skill targets a specific mechanism that decades of research have shown produces durable, transferable knowledge.

Retrieval practice strengthens long-term memory more than re-reading. When a skill asks you to recall a concept before showing you the answer, it forces the brain to reconstruct the memory pathway — a process called the testing effect that makes the memory stronger and more accessible later. Every /retain exercise and do-not-repeat question is built on this.

Desirable difficulty is the sweet spot where learning happens. Tasks that are too easy slip past without leaving a trace. Tasks that are too hard cause frustration without progress. Learning Skills calibrate the challenge by asking you to compare, explain, or predict — activities that sit right on the edge of your current ability and create the kind of cognitive effort that strengthens neural connections.

Spaced repetition exploits the brain's forgetting curve. Memories decay over time, but each retrieval attempt resets the curve at a higher plateau. Background Learning Skills track what you struggled with across sessions and reintroduce those concepts at widening intervals, converting fragile short-term recall into long-term retention that survives weeks and months.

The generation effect shows that self-produced answers are remembered better than provided ones. When you name the assumption, frame the alternative, or build the mental model yourself, the act of generating the answer encodes it more deeply than receiving it from the model. That is why /reasoner gives you questions instead of conclusions, and why /explain-away-others asks you to compare before telling you the tradeoff.

Metacognition — knowing what you know and what you do not — is one of the strongest predictors of learning success. AI makes it easy to skip this step entirely by delivering confident answers that mask gaps in understanding. Skills like /anti-passive and /no-assumptions surface those gaps by asking you to state your confidence, name what is missing, and recognize when the model is making decisions you should be making yourself.

VidByte Learning CLI v1 starts with the core anti-autopilot skills and the installation path across common harnesses. More background skills are coming next: coverage for mapping concept gaps, practice for testing one idea from many angles, question-builder for finding the next question, struggle for tracking weak spots, and transfer for catching missed connections.

The direction is straightforward: make AI-assisted work teach the user while it helps the user ship. The better the harness gets, the more important it becomes to keep the human curious, skeptical, and capable of reaching the conclusion themselves.