| name | learn-like-me |
| description | build a personalized project-learning assistant from a user's learning preferences. use when the task is to infer how someone learns repositories, codebases, architecture, modules, or docs from chat history first, synthesize a learning profile, and generate a reusable runtime skill that teaches in their preferred style. |
LearnLikeMe
Build a personalized runtime skill named project-learning-assistant from a user's project-learning preferences. Default to chat-history-first inference, use follow-up questions only to close important gaps, and write the generated artifacts to the local generated-skills/ workspace area.
Workflow
- Read learner history first from the location described in
references/history-ingestion.md.
- Extract preference evidence, confidence, and stable-vs-situational distinctions before asking questions.
- Ask only for missing, contradictory, or low-confidence fields using the intake guidance in
references/intake-questionnaire.md.
- Normalize all evidence into the schema in
references/learning-profile-schema.md.
- Generate the runtime skill from
references/runtime-skill-template.md.
- Write the runtime skill to
generated-skills/project-learning-assistant/ unless the user explicitly requests a different output location.
- Write the inferred learning profile to
generated-skills/learning-profile.md unless the user explicitly requests a different output location.
- Use
references/output-templates.md for the learning profile, generated skill output, and response contract.
- Use
references/history-to-skill-workflow.md when the task is framed as a Codex product workflow from local history to reusable skill.
- Use
references/history-drop-template.md when the user wants a repeatable folder convention for teams or repeated runs.
- Check
references/examples.md when you need contrasting patterns for beginner, intermediate, and advanced learners.
Builder Rules
- Treat history as the default evidence source. The questionnaire is a targeted gap-filling tool, not the default first step.
- Preserve the user's actual pacing, explanation depth, question style, and comprehension limits instead of averaging them into a generic tutor voice.
- Separate
stable_preferences from situational_preferences as a required output.
- Record concrete evidence for major preference claims and keep confidence visible in the profile.
- Generate a runtime skill named
project-learning-assistant.
- Treat
generated-skills/project-learning-assistant/ as the default runtime skill output directory.
- Treat
generated-skills/learning-profile.md as the default learning profile output file.
- Keep the generated runtime skill focused on learning repositories, codebases, docs, architecture, and modules.
- Require the generated runtime skill to include a Codex invocation hint so the user knows how to use it after generation.
- Make the runtime skill proactive: ask likely hidden questions, build a roadmap, control pacing, and summarize progress after each step.
- Require every runtime response to contain these sections in order:
project map, current understanding, hidden questions, next best step
Self-Check Gates
Before finalizing the generated skill, verify all gates:
- The builder produced a personalized skill and not a one-size-fits-all tutor.
- The runtime skill behaves like a learning copilot rather than passive Q&A.
- If history was provided, the generated runtime skill visibly reflects stable patterns from that history rather than only the latest follow-up answers.
If any gate fails, revise the learning profile and regenerate the runtime instructions with stronger profile-specific behaviors.