learned-from-ai#

Turn transient AI chat output into structured, reviewed, long-lived learning material that is easy for a human to study, remember, and revisit.

Non-negotiable rules#

1. Always handle tasks under this skill through a subagent by default so the main session does not get blocked, unless the user explicitly asks otherwise.
2. Use the preferred subagent settings by default: runtime: subagent, model: openai-codex/gpt-5.4, thinking: medium.
3. Always save outputs in notes/ unless the user explicitly asks for a different location.
4. Always keep the original shared/source link in the main summary note when a link exists, so the source can be traced easily.
5. Before writing, search the notes/ folder for existing related notes by subject/project so you do not overwrite durable knowledge accidentally.
6. For boundary cases on the same project/topic, do not rewrite the existing note by default. Create a new summary and cheat sheet instead.
7. Name new boundary-case files intelligently: use either a more specific sub-subject name or the existing knowledge name plus an incremented suffix.
8. Always generate a cheat sheet based on the reviewed main note.
9. Do not violate the preferred structure unless the user explicitly asks for a different one.
10. Strongly remove AI slop, repetition, weak filler, and hallucinated claims.
11. Cross-check questionable facts, formulas, standards, and numbers when needed.
12. Keep the main note and cheat sheet separate.

---

Preferred structure#

Use this exact structure unless the user explicitly overrides it:

1. Definition
2. Essential ideas / engineering practice
3. Worked examples and calculations
4. Important theoretical derivations
5. Q&A from the discussion
6. Further reading / viewing

Always create a separate cheat sheet file based on the reviewed main note.

---

Workflow#

1. Start by spawning the working subagent
   • For tasks under this skill, start with a subagent by default so the main session stays responsive.
   • When this skill is activated with a slash command and the user appends a chat/share link, immediately spawn the subagent.
   • Use the default settings unless the user explicitly overrides them:
     • runtime: subagent
     • model: openai-codex/gpt-5.4
     • thinking: medium
   • Give the subagent the link or source material and the required output structure.
2. Inspect the source
   • Read the shared link, pasted chat, file, or notes.
   • Extract the real technical content.
   • Ignore UI noise, fluff, and repeated AI phrasing.
3. Pre-search the knowledge base in notes/
   • Before naming or writing files, inspect existing note filenames in notes/ for the same subject, project, or nearby topic.
   • Use this step to avoid overwriting durable notes.
   • If the new source is clearly a new subtopic or a separate chat on the same project, plan a new note instead of rewriting the old one.
4. Identify the subject and output files
   • Pick a short subject-based filename.
   • By default, write a new note rather than overwriting an existing one when the source is a new chat, new link, or new subtopic.
   • Write the main note to notes/<subject>.md.
   • Always write the cheat sheet to notes/<subject>-cheatsheet.md.
   • If needed, use either:
     • a more specific sub-subject name, or
     • the existing knowledge name plus an incremented suffix.
   • If the source came from a shared/public link, record that original link near the top of the main note so the summary can be traced back to its source easily.
5. Review and verify before polishing
   • During review, use strong reasoning and factual discipline.
   • Catch factual errors.
   • Remove hallucinations.
   • Strip AI slop.
   • Cross-check formulas, standards, fit values, and calculations when needed.
   • Distinguish exact statements from approximations.
   • Preserve useful approximations, but label them honestly as approximations, first-pass checks, or worst-case bounds.
6. Write the main note
   • Follow the preferred structure exactly.
   • Do not reorder or silently replace it with a different teaching flow.
   • Make definitions crisp, logic coherent, and examples numerically consistent.
   • The preferred structure must not be violated.
   • Do not overwrite an existing durable note unless the user explicitly asks for revision of that specific file.
7. Write the cheat sheet
   • Base it on the reviewed main note.
   • Keep it separate from the main note.
   • Distill, do not duplicate.
   • The preferred main-note structure must still remain intact and must not be violated.
8. Finalize and organize
   • Ensure files are in notes/.
   • Use short, practical names.
   • Avoid redundant filenames like -study-note unless the user explicitly wants them.

---

Writing standards#

Keep#

• precise definitions
• practical engineering or domain logic
• worked numerical examples
• short derivations that reveal the principle
• explicit assumptions and limitations
• Q&A clearly separated from exposition

Remove#

• AI filler
• repetitive hype
• vague certainty
• unsupported claims
• long padding that does not improve learning

Prefer#

• short sections
• bullets over bloated prose
• equations when they clarify reasoning
• ASCII sketches when a simple drawing helps
• concise filenames

---

Review checklist#

Before finishing, check:

• Are the files in notes/?
• Does the main note keep the original shared/source link when one exists?
• Does the main note follow the preferred structure exactly?
• Is the cheat sheet separate and genuinely distilled?
• Were suspicious claims cross-checked?
• Were hallucinations and AI slop removed?
• Are examples and calculations internally consistent?
• Are approximations labeled clearly?
• Are filenames short and subject-based?

---

Example file layout#

text

notes/
  gdt.md
  gdt-cheatsheet.md

---

Scope#

This skill is for turning AI chat interactions into durable human learning materials.

It is not mainly for:

• writing full textbooks from scratch
• doing exhaustive literature reviews
• dumping raw chat transcripts into files without review

If the source is rough, correct it. If it is verbose, compress it. If it is uncertain, verify it.