Engineer software with AI using composable development skills

A stateful teaching skill (mattpocock/skills, MIT) that treats the current directory as a persistent teaching workspace across multiple sessions, structured around seven file types: MISSION.md capturing why the user wants to learn the topic and grounding all teaching; ./reference/*.html, compressed cheat-sheet-style reference materials; RESOURCES.md, a list of high-trust external resources; ./learning-records/*.md, numbered records of non-obvious lessons learned that drive future sessions and calculate the zone of proximal development; ./lessons/*.html, the primary teaching unit - single, self-contained, beautifully-typeset HTML files each teaching one tightly-scoped thing tied to the mission; ./assets/*, reusable components like stylesheets, quiz widgets, and simulators shared across lessons; and NOTES.md, a scratchpad for user preferences. Its philosophy holds that deep learning needs three things - knowledge from high-quality trusted resources, never the model's own parametric knowledge, skills acquired through interactive lessons, and wisdom from interacting with other learners and practitioners - with the mix shifting by topic, since theoretical physics skews knowledge-heavy while yoga skews skills-heavy. It distinguishes fluency strength, in-the-moment retrieval that can create an illusory sense of mastery, from storage strength, long-term retention and the real goal, and designs lessons around desirable difficulty via retrieval practice, spaced practice, and interleaving for skills. Each lesson must be short and completable quickly since working memory is small, give one tangible win building toward the mission, sit in the user's zone of proximal development, recommend a primary high-trust source, remind the user to ask the agent follow-up questions, and reuse existing ./assets/ components rather than inlining code a future lesson would duplicate - a shared stylesheet is described as the first component every workspace earns. If the mission is unclear or MISSION.md is unpopulated, the skill's first job is to question the user on why they want to learn the topic rather than proceeding ungrounded; missions may change over time but only with an updated MISSION.md, a new learning record, and user confirmation. For skill acquisition specifically, quizzes must give answers of exactly the same word and character count across options so formatting doesn't leak the answer, and wisdom-seeking questions default to an honest attempt plus delegation to a high-reputation real-world or online community, respecting a user preference not to join one.

Use it when the user asks to be taught a new skill or concept within a workspace, intending to learn it over multiple sessions. Do not use it for a one-off explanation with no persistent learning state - the skill is explicitly stateful, building MISSION.md, reference materials, learning records, and lesson files that accumulate and get revisited across sessions.

Input is a topic the user wants to learn, plus their evolving learning-record history. Output is a populated teaching workspace: a MISSION.md, RESOURCES.md, numbered learning-records/000N-<dash-case-name>.md files, numbered lessons/000N-<dash-case-name>.html files each teaching one scoped thing, shared assets/ components, and reference/*.html cheat sheets, all cross-linked via HTML anchors.

Self-contained around a directory-based teaching workspace with defined file formats referenced from companion docs MISSION-FORMAT.md, RESOURCES-FORMAT.md, and LEARNING-RECORD-FORMAT.md; opens generated lesson files for the user via a CLI command when possible.

Users learning a new skill or subject over multiple sessions who want a persistent, mission-grounded curriculum of short, cited, interactive lessons rather than one-off explanations.