Concept Explainer

concept-explainer — Kind, Rigorous, Visualization-Heavy Explanations

Produce a single explanation document that walks a named audience through a specific concept without logical leaps and without sacrificing mathematical rigor. Equations carry their weight, every symbol is defined, every step in a derivation names the rule it applies, and visualizations appear wherever they shorten the gap between a formula and an intuition.

The skill writes:

<out-dir>/                   # <YYYYMMDD>_<concept-slug>, e.g. 20260728_lagrangian-mechanics
├── explanation.md           # the document (language follows user's request)
├── plots/                   # matplotlib scripts + their rendered PNGs
│   ├── *.py
│   └── *.png
└── schematics/              # OPTIONAL — only when conceptual diagrams help
    ├── *_prompt.md          # copy-pasteable wide-slide-illustrator prompts
    └── *.png                # rendered images (user fills these in, or codex-image)

<out-dir> is <YYYYMMDD>_<concept-slug>: the day the directory is created (date +%Y%m%d) followed by the kebab-case concept slug. Alphabetical order then equals chronological order in any folder that collects explanations.

After the document is complete, the skill invokes md2pdf-typora to produce <out-dir>/explanation.pdf, then mirrors the explanation (markdown, PDF, plots, schematics) into ~/Dropbox/ConceptExplainer/<Topic>/<out-dir>/, organised by topic. The archive folder keeps the same name as the local directory; the date prefix carries the ordering, so there is no separate index. This archive step runs for every explanation regardless of language.

Mandatory invariants (non-negotiable)

These hold across every explanation produced by this skill.

Rigor invariants

1. Define before use. Every symbol that appears in an equation is defined before its first non-definitional use, on the same line or in the line above. Quantities with units carry the unit on first appearance.
2. Name the rule. Every step in a derivation states the rule it applies ("by the chain rule", "integrating both sides", "using assumption A1", "Taylor-expanding around x = 0 to second order"). Never write "it can be shown that" without either a one-line sketch or a citation.
3. Mark approximations. Use = for exact equalities, ≈ for approximations (state the order: "to leading order in ε"), ∼ for asymptotic behavior, ∝ for proportionality. Mixing these silently is a logical leap and is forbidden.
4. State assumptions up front. A boxed ## Assumptions block lists the regime / smoothness / boundary conditions you depend on. When you later drop or generalize one, say so explicitly.
5. No forward references. A concept cannot appear in prose before its definition. If the reader needs X to understand Y, X is in an earlier section.
6. Cite the non-derivable. Empirical facts, named theorems used without proof, and any numerical constant beyond textbook precision need a citation. Invoke the reference-search skill to find sources rather than guessing.
7. Forbidden hedges. "Obviously", "clearly", "trivially", "it is well-known that" — strip them. If a step is genuinely one line, write the line. If it isn't, write the steps.

Visualization invariants

8. scienceplots ["science", "nature"], always. Every matplotlib script uses import scienceplots (registered by side-effect, never referenced), wraps all plotting in with plt.style.context(["science", "nature"]):, and saves with dpi=300, bbox_inches='tight'. The no-latex style is forbidden. Plots must render LaTeX through the system TeX install.
9. Raw-string LaTeX labels. Every axis label, title, legend entry, and text annotation uses r'...'. Non-raw strings silently break the moment a backslash appears.
10. pparam dict + ax.set(**pparam) + ax.autoscale(tight=True). Same structural invariants as scienceplot-py. Do not inline ax.set_xlabel(...) calls when pparam would do.
11. Caption discipline. Every figure in explanation.md carries a caption that says what to look at, not just what the plot contains. The caption interprets the figure for the named audience.
12. Schematic vs data plot — pick the right tool. Conceptual diagrams (boxes, arrows, "pipeline overview", "what each component does") are composed via wide-slide-illustrator Friendly Whiteboard style and live in schematics/. Mathematical visualizations (functions, parameter sweeps, fields, phase portraits, convergence plots) are matplotlib scripts in plots/. Do not draw boxes-and-arrows with matplotlib; do not plot f(x) curves through an image-gen model.

Output invariants

13. Language follows user's request. If the user wrote in Korean, the document is in Korean. If in English, English. Equations are language-independent and stay verbatim. Section headings, prose, figure captions all match the request language.
14. One concept per run. This skill produces one focused explanation. For a multi-concept curriculum, run the skill once per concept and let the user thread them.
15. No co-author tags in any generated file. Per the user's global rule.

Workflow

1. Gather inputs

Ask only for what is missing. Required:

• Concept: the thing to explain. Be specific — "the Lagrangian formulation of classical mechanics" rather than "Lagrangians".
• Audience: a free-form one-liner describing the reader. Examples: "학부 2학년 물리, 라그랑지안 모름, 미적분은 됨" or "ML PhD student new to measure-theoretic probability".
• Concept slug: kebab-case directory name. Default: derived from the concept (e.g. lagrangian-mechanics).
• Output directory: default <cwd>/<YYYYMMDD>_<concept-slug>/, referred to as <out-dir> below. Take the date from date +%Y%m%d.

Optional:

• Length target: short (800 words), standard (2000 words), or long (~4000 words). Default: standard.
• Schematic figures: yes / no / "only if it helps". Default: "only if it helps" — produce schematics only for genuinely conceptual structure (architectures, pipelines, taxonomy), never for purely mathematical content.
• Reference search: yes / no. Default: yes for non-derivable claims.

2. Calibrate to the audience

Translate the free-form audience description into a calibration table. See references/audience-calibration.md for the full protocol. The output is an internal dict (do not put this in the document):

- prerequisite_knowledge:  [list of concepts treated as known]
- needs_definition:        [list of concepts that must be defined inline]
- notation_conventions:    [e.g. "physicist's index notation", "ML-style E[·]"]
- math_density:            low | moderate | high
- worked_example_density:  low | moderate | high
- approximation_tolerance: how casually the audience treats "≈"

If the audience description is too vague to calibrate (e.g. just "학생"), ask one targeted clarifying question. Do not guess.

3. Plan the section structure

Use the skeleton in references/structure-template.md. The default 10 sections (translate headings to match the requested language):

1. Why this matters — the motivating question. 1–2 paragraphs.
2. Setup — the simplest non-trivial case that still exhibits the phenomenon.
3. Definitions and assumptions — every symbol, every regime constraint.
4. Intuition — verbal picture + (optional) schematic figure.
5. Step-by-step derivation — each step names its rule; show the intermediate equations.
6. Worked example — concrete numbers, with a function plot or sweep.
7. Visualizations — parameter dependence, limits, edge cases.
8. Generalization — how the result extends; what assumptions can be relaxed.
9. Limitations and common misconceptions — failure modes, pitfalls, numbered "Misconception N → 오개념 N" callouts.
10. Where to go next — references, follow-up concepts, exercises.

Drop or merge sections only when the concept genuinely doesn't need them (e.g. a definition-only note may not need §8). Adding sections is fine.

4. Draft the document

Write <out-dir>/explanation.md with the structure from step 3. While drafting, the rigor invariants from above are load-bearing — read references/rigor-checklist.md and apply each item to every paragraph.

Equation typesetting in markdown:

Inline math: $\alpha = 0.05$, $\hat{\theta} \in \mathbb{R}^d$.

Display math on its own line:

$$
\mathcal{L}(q, \dot{q}, t) = T(\dot{q}) - V(q)
$$

with $T$ the kinetic energy, $V$ the potential, $q$ the generalized
coordinate, and $\dot{q} = \mathrm{d}q/\mathrm{d}t$.

The line after a display equation defines every symbol introduced in it. This is the define-before-use rule operationalized.

5. Plan visualizations

For each section that benefits from a figure, decide schematic vs plot using references/visualization-playbook.md. Common patterns:

For more advanced data sources (parquet/CSV/npy/npz) or plot variants (scatter/errorbar, generic subplots), defer to the scienceplot-py skill templates — do not duplicate them here.

6. Render the plots

For each matplotlib plot:

1. Adapt the matching skeleton from references/plot_skeletons/ into a concrete .py file under <out-dir>/plots/<plot_name>.py. Keep every invariant from scienceplot-py intact.

2. Execute it: uv run <out-dir>/plots/<plot_name>.py. This skill does execute its own plot scripts (unlike scienceplot-py, which does not). The PDF assembly later depends on the PNGs existing.

3. Verify the PNG exists and is non-empty: ls -la <out-dir>/plots/<plot_name>.png.

4. Reference the PNG in explanation.md with a caption that says what to look at:

   ![Action vs path length for three trajectories. The straight-line path
   minimizes the action; small wiggles raise it; large wiggles raise it
   much more. The minimum here corresponds to the classical trajectory.](
   plots/action_vs_path.png)
   

If a plot script fails to render, fix the script — do not silently remove the figure from the document. Common causes: missing TeX system package, wrong column name, math expression that needs raw-string.

7. Compose schematic prompts (only if planned in step 5)

For each schematic figure:

1. Invoke the wide-slide-illustrator skill with the Friendly Whiteboard style. See references/schematic_friendly.md for the brief — it gives the panel-count guidance, the on-canvas-text-is-English rule, and the per-panel content fields needed.
2. Save the composed prompt as <out-dir>/schematics/<schematic_name>_prompt.md.
3. The user feeds this prompt to ChatGPT Image / DALL-E / Sora / Midjourney and saves the result as schematics/<schematic_name>.png. (Alternative: the codex-image skill can run this automatically — mention this option to the user but do not auto-invoke unless they ask.)
4. Reference the image in explanation.md with a caption. If the image does not yet exist, write the markdown reference anyway and tell the user explicitly which files they need to generate before the PDF step.

8. Reference search (when needed)

For any non-derivable claim flagged in step 4, invoke the reference-search skill to find supporting references. Inline citations use a short parenthetical like (Goldstein 2002, §1.4) or a numeric [1] style with the source list in §10. Do not fabricate citations.

9. Validate the draft

Before assembling the PDF, run the checklist in references/rigor-checklist.md:

• Every display equation followed by a symbol-definition line.
• No forbidden hedges ("obviously", "clearly", etc.) — grep for them.
• Every figure referenced from the markdown exists in plots/ or schematics/.
• Every figure has a caption that interprets the figure.
• Every cited reference appears in §10.
• No forward references.

If any item fails, fix the document — do not paper over it.

10. Assemble the PDF and archive to Dropbox (automatic)

Every finished explanation is exported to PDF and mirrored into the user's ~/Dropbox/ConceptExplainer archive, organised by topic. Do this automatically for every explanation regardless of language — do not wait to be asked.

1. Export the PDF with the md2pdf-typora skill on <out-dir>/explanation.md, producing <out-dir>/explanation.pdf. The Whitey theme handles Korean and English both.
2. Pick the topic folder. The archive is organised as ~/Dropbox/ConceptExplainer/<Topic>/<out-dir>/, reusing the same <YYYYMMDD>_<concept-slug> folder name used locally, so the topic listing sorts chronologically without a separate index.
   • List the existing topics (ls ~/Dropbox/ConceptExplainer) and choose the one that fits the concept. If none fits, ask the user which topic to use or whether to create a new one (reuse the JournalClub topic vocabulary — InverseProblem, NeuralOperators, PBH, SMEFT, Unfolding, ... — when sensible); create it only after they confirm the name. Do not silently invent a topic.
   • If the local directory predates the convention and has no date prefix, prepend one for the archive copy only, taken from the explanation.md mtime. Leave the local directory alone.
   • If the concept is already archived under an older name, update that folder in place rather than creating a second, renamed copy.
3. Copy the artifacts into ~/Dropbox/ConceptExplainer/<Topic>/<out-dir>/: explanation.md, explanation.pdf, the plots/ directory (scripts + rendered PNGs), and the schematics/ directory if it exists. Match the existing layout in sibling explanation folders.
4. Confirm the destination path and file sizes to the user.

11. Report and stop

Report to the user:

• The output directory.
• The list of generated plot scripts and which produced PNGs.
• The list of schematic prompts (and which images still need to be generated by the user).
• The PDF path.
• The Dropbox archive path.

Do not push, commit, or otherwise touch git state. The commit-triage skill handles that when the user is ready.

Audience calibration

See references/audience-calibration.md for the full protocol. Short version: convert a free-form one-liner into a calibration table covering prerequisite knowledge, vocabulary that needs inline definition, notation conventions, and density dials. When in doubt, ask one targeted question rather than guessing — getting the audience wrong wastes the rest of the workflow.

Rigor checklist

See references/rigor-checklist.md. Apply every item to every paragraph before declaring the draft complete. The checklist exists because every item on it is a failure mode I have seen in unaudited "kind explanations": silent assumption-dropping, ambiguous =/≈, hidden forward references, "obviously" used as a load-bearing inference, fabricated citations, mismatched units.

Visualization playbook

See references/visualization-playbook.md. Picks the right tool per visualization need, lists which plot skeleton to start from, and gives the "what to look at" caption pattern.

Plot skeletons

Self-contained, runnable matplotlib templates that satisfy every visualization invariant.

For other variants (scatter/errorbar, data files from parquet/CSV/npy), generate the script through the scienceplot-py skill and place the output under plots/.

What this skill does NOT do

• It does not produce slide decks. For talks, hand the markdown to a slide-generation pipeline (claude-typst-slides:slides-present, frontend-slides).
• It does not run hyperparameter sweeps or fit models. If the explanation needs experimental data, run that out of band and pass the resulting parquet/CSV through scienceplot-py.
• It does not turn an entire textbook chapter into a single document. One concept per run. Threading is the user's job.
• It does not invent citations. If reference-search returns nothing usable, leave the claim un-cited and flag it to the user, or rewrite the claim so it no longer needs a citation.
• It does not skip the auto-PDF step on success. Every explanation always reaches ~/Dropbox/ConceptExplainer/<Topic>/<YYYYMMDD>_<concept-slug>/.
• It does not commit anything. commit-triage does that, when the user asks.