# Lambda Draw > Diagram production for λambda — ANY teaching figure OR data chart, any subject. Graphical models (BNs, Markov chains, HMMs, DBNs, jointrees) AND quantitative charts (time-series, line/bar plots, distributions, option-payoff diagrams). Use for "/lambda-draw " or whenever a session's teaching step needs a figure or a chart. Picks the medium (mermaid / mermaid xychart / TikZ / DOT / ASCII), enforces the compile→render→LOOK→fix loop for TikZ and a trace-every-point check for data charts; nothing unviewed or unverified is ever embedded. - Skill: `abaj8494/lambda-draw` (Agent Skill) - Install (CLI): `npx skillmds@latest add abaj8494/lambda-draw` - Raw SKILL.md: https://api.skillmd.com/api/skills/abaj8494/lambda-draw/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Coding & Dev Tools - Author: abaj8494 (https://skillmd.com/u/abaj8494) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/abaj8494/lambda-draw --- # λambda draw protocol Diagrams for three destinations: live session files, vault course notes, and Anki cards. The failure mode this skill exists to kill: TikZ written blind and shipped unverified. **You have vision — every figure gets rendered and looked at before it is embedded, no exceptions.** ## Medium decision table | Situation | Medium | Why | |---|---|---| | Plan DAGs, quick structure inside a live session | **mermaid** | renders instantly in Obsidian/GitHub; zero toolchain; fine for arrows-and-boxes | | Course figures and Anki cards: BNs with observed/latent shading, HMM/DBN time slices, plates, message passing | **TikZ** (default) | course-faithful notation; Anki's builtin LaTeX already renders it | | Layout is the hard part: moral graphs, jointrees, >12 nodes, elimination orderings | **DOT → `dot -Tsvg`** | graphviz auto-layout beats hand-placing; optionally use the result as a layout draft for TikZ | | Quantitative data: a time-series, a line/bar plot, a distribution, an option-payoff diagram | **mermaid `xychart-beta`** (renders natively in the app), or a labelled ASCII/Unicode plot in a code fence | native render, no toolchain; ASCII when the plot is small or xychart fights you — see the rules below, this is where wrong charts ship | ## Quantitative charts (a chart is a claim — the trap that shipped a wrong graph) A wrong chart is a wrong claim delivered with false confidence. On surfaces where you cannot compile-and-LOOK (mermaid in chat, the lounge), the LOOK step becomes a **trace-every-point check you run in your head before emitting**. **mermaid `xychart-beta` — all binding:** - **x-axis category labels MUST be unique.** The band axis keys on the label *string*: two identical labels collapse to ONE slot, and any later point carrying that label jumps **backward** onto the earlier slot. This is the exact bug that shipped — an implied-vol series with `x-axis ["pre-Q3","Q3 print","post","pre-Q4","Q4 print","post"]` rendered its final point regressed, because the second `"post"` collapsed onto the first. For repeating phases, disambiguate: `post-Q3`, `post-Q4`. - **Data length == x-axis count.** `line [...]` / `bar [...]` with N values needs exactly N x-categories, in the same order. - **y-axis range spans the data:** `y-axis "label" MIN --> MAX`, MIN below the smallest value, MAX above the largest — else points clip at the frame. - **Trace every point before shipping.** Read each value against its (unique) x-label and confirm the shape is what you meant: the monotone part is monotone, the sawtooth actually saws, the last point sits at the last x. This is the data-chart equivalent of the TikZ LOOK. **Prefer a labelled ASCII/Unicode plot** for a quick trend, ≤ ~8 points, or when xychart fights you: it renders everywhere (monospace code fence), cannot collapse labels, and matches the house classical-print style. Hand-place the marks against a labelled axis and trace them the same way, e.g.: ``` implied vol (pts) 58 ┤ ● │ \ ● rebuild is slow… 42 ┤ ●______/ \ 38 ┤ \ / \ 37 ┤ \__/ ● ← still below, mean-reverting up into the next print └────┬────┬────┬────┬────┬──── pre-Q3 post Q4 post-Q4 ``` **Renderer note:** LambdaMind's chat and session panes render mermaid natively (```mermaid fences). `xychart-beta` works there; the band-axis label-collapse is a mermaid property, **not an app bug** — uniqueness is on you. Never emit a chart you have not traced. ## The TikZ loop (mandatory) 1. Write the picture into the scratchpad using the standalone harness from `~/lattice/org-notes/uni/Anki.md` (binding — read it first in any session that draws for Anki): ``` \documentclass[tikz,border=3pt]{standalone} \usepackage{amsmath,amssymb} \usetikzlibrary{arrows.meta,positioning,calc,backgrounds,shapes.geometric,fit} \begin{document} ... \end{document} ``` 2. `lualatex -interaction=nonstopmode -halt-on-error` in the scratchpad. 3. `pdftoppm -png -r 150` the result. 4. **Read the PNG and inspect it**: arrowheads point the right way; no label/edge collisions; observed nodes actually shaded; slices aligned; nothing orphaned or overlapping. Fix and recompile. 5. Budget ≤3 compile-view cycles; if it still fights you, simplify the figure (fewer nodes, drop the plate, straight edges) rather than fight TikZ. A clean simple figure beats a mangled ambitious one. 6. Only a viewed, passing figure may be embedded or pushed to a card. ## PGM drawing conventions - Latent variables: `\node[vertex] (X1) {$X_1$};` with `vertex/.style={circle, draw, minimum size=8mm}`. - Observed variables: `observed/.style={vertex, fill=gray!30}`. - DBN/HMM time slices: one column per slice, transition arrows horizontal, emission arrows vertical, trailing `\node {$\cdots$}` for unrolling. 2-TBN form: two slices + the inter-slice arrows only. - Plates: `backgrounds` layer + `fit`, label bottom-right. - Undirected (Markov networks): plain `--` edges, no arrowheads. - Style names: never bare English words TikZ may own (`out`, `in`, `at`, `to`, `node`, `edge`, `every`, `cap` have all bitten). Use `vertex`, `observed`, `slice`, `blanket`. - No pgfplots; hand-draw curves if ever needed. - Course-faithful notation (check the course's own slides for symbol conventions before labelling). ## Slide excerpts (fourth tier — not drawing, extracting) When a teaching step cites slide/page $N$ of a course PDF, show it: 1. `pdftoppm -png -r 150 -f N -l N /slide` (single page). 2. If only a region matters, crop: `magick -crop WxH+X+Y +repage out.png` — then **view the result** to confirm the crop caught the right region. 3. Copy to `/sessions/assets/-s.png`, embed relatively in the session file next to the teaching step. Course slides are copyrighted: excerpts live in the private vault only — never in anything public, never in the OSS repo, never in shared decks. ## Destinations - **Vault embed**: save PNG to `/courses//figs/.png` and embed relatively (`![...](../courses//figs/.png)` from a session file, or wiki-embed). Keep the `.tex` source next to it. - **Anki card**: the card carries the **bare** `\begin{tikzpicture}...` source (never wrapped in `\[...\]`) — only after the local compile passed. All Anki.md TikZ rules binding. - **Session file quick-structure**: mermaid inline; no files. ## Compile hygiene All compiles run in the scratchpad, never in the vault. Copy in only the final PNG + tex source. If the vault is a git repo, figures ride the session's end-of-session commit.