Physics Diagrams

Use this skill to produce paper-ready physics diagrams, not quick sketches. The default path is LaTeX-native: explicit-coordinate TikZ with shared pd* styles and manual label lanes. Use tikz-feynhand when its propagator idioms make a Feynman/process diagram clearer. Compile a standalone vector PDF, render a high-DPI preview, inspect the preview, then deliver both source and PDF.

Decision Tree

1. Feynman/process/tree/loop/exchange diagram: use explicit-coordinate TikZ; add tikz-feynhand when particle-line idioms such as \propag, pho, or sca are useful. Use hand-placed vertices. Do not rely on automatic graph layout for final figures.
2. Integral-equation, BSE/Faddeev, recursion, colored-block schematic: use TikZ nodes, paths, nested tikzpicture blocks, and tikz-feynhand only for local line idioms.
3. Geometry, coordinate frames, angle planes, projected vectors, kinematic sketches: use Matplotlib/Python when that makes geometry clearer. Keep TeX-like math labels and vector PDF output.
4. Existing data plotted as a figure: this skill may help with visual styling, labels, vector export, and QA, but it does not compute the numerical data. Use the project's own computation pipeline for physics results, and use the sibling figure-hygiene skill for data/results-plot correctness and legibility (data fidelity, chart choice by data shape, label economy, colour threading, render-then-verify).
5. FeynArts/QGRAF: use only upstream to enumerate or check model-generated topologies. Redraw selected publication figures in TikZ/feynhand or another final-layout backend.
6. Inkscape: use only for optional finishing, imported SVG cleanup, or manual repair. It is not the canonical source unless the user explicitly chooses an SVG-first workflow.

Read references/tool-selection.md when the backend is uncertain.

Required Workflow

1. Write a semantic mini-spec before drawing: diagram role, particles/objects, line styles, vertices/blocks, labels, preferred label side, panel layout, manuscript width.
2. For diagrams that summarize an existing calculation, read the calculation source first and write a calculation-to-figure map: each kernel term, channel, topology, subdiagram, zero/selection rule, and intermediate state must have a corresponding visual element or an explicit reason for omission.
3. Pick the narrowest backend from the decision tree.
4. Choose the abstraction level for each row: if a component is expanded in a lower row, use one representative glyph in the parent row unless the algebra truly contains multiple factors or terms.
5. Use one editable source of truth:
   • default: standalone figs/src/<name>.tex -> figs/<name>.pdf;
   • existing manuscript inline style: keep figs/<name>.tikz, but create a standalone wrapper for QA;
   • Python geometry/plot route: keep the Python script plus generated vector PDF.
6. Reserve label lanes before drawing lines. Place labels with explicit anchors and offsets.
7. Compile in a clean build directory with scripts/build_diagram.py.
8. Open or view the rendered PNG preview. Check every math label against the quality gate.
9. State whether the current agent can actually read the preview image. If not, mark the figure NOT publication-ready until a human or visual subagent/model reviews the preview.
10. For complex/submission-bound figures, or any figure with more than six labels, multiple panels, or nested tikzpicture blocks, ask an independent visual reviewer/subagent to inspect the preview. If policy and tools allow, a different model family may be used for this visual review.
11. Iterate coordinates, anchors, backing boxes, and spacing until the preview passes. When user critique reveals a reusable failure mode, update this skill or its references after fixing the figure.
12. If moving an approved scratch figure into a target paper repo, adapt it to that repo's figure convention (for example figs/src/*.tikz plus a build script) and use neutral filenames such as <subject>_schematic, not overclaiming names such as <subject>_publication.
13. Deliver: source path, vector PDF path, insertion snippet, preview path, mini-spec, image-reading verdict, verifier identity, and verification status.

Label Safety Rules

• Prefer manual coordinates and explicit anchors over automatic edge labels.
• Give near-line labels a white backing: fill=white, fill opacity=0.92, text opacity=1, inner sep=1pt.
• Offset labels from paths with above, below, left, right, near start, near end, xshift, and yshift; do not center labels on strokes unless the white backing is intentional.
• Use filled circles/boxes/diamonds for amplitudes, kernels, propagators, and operators only when the calculation treats them as unresolved effective objects. If a propagator is dressed or has an explicit self-energy/loop definition in the paper or code, draw that topology instead of a box.
• For compact filled vertex blobs, draw all incident lines to the same center coordinate first, then draw the filled circle last so the unphysical interior line pieces are hidden.
• For HEP particle-line schematics, draw individual particles as single lines. Use solid single lines for heavy mesons such as D and D^*; use dashed single lines for light exchanged particles such as K and \pi unless the manuscript defines another convention.
• Draw a dimer/isobar as two separate nearby lines, not TikZ's closed-ended double stroke. Each of the two dimer lines should inherit the line style of the constituent particle: e.g. DK is one solid plus one dashed line, while DD^* is two solid lines.
• At an exchange vertex, do not let a dimer continue through the vertex as a dimer. A dimer emits one constituent/exchanged particle and becomes the remaining single particle; on the other leg, a single particle plus the exchanged particle becomes the outgoing dimer.
• Keep default colors restrained: black/gray strokes with low-saturation fills. Use strong colors only when the user asks or the manuscript already has that palette.
• Draw lines up to a block boundary and resume after the block; do not let a line pass through a block and rely on the fill to hide it.
• If an interaction is physically a point vertex such as a WT vertex, draw it as a vertex blob/circle with incident lines connected to the center and hidden by the blob. Do not draw it as an operator box with visible gaps between the line ends and the symbol.
• Do not collapse a multi-loop kernel or box contribution into a single vertex/block. If the calculation uses a two-loop box, show the box topology and its live intermediate particles or include the existing verified box source/PDF as a vector inset.
• For a dressed dimer--spectator propagator such as G_i, draw the spectator as a straight line and the dimer leg with its self-energy bubble (dimer constituents propagate and recombine), matching the paper's self-energy figure and the implementation, rather than inventing a generic G block.
• For channel-dependent exchange kernels, draw separate exchange topologies for each allowed channel pair. Do not use one generic exchange diagram labeled with a list such as K,D^*,D when the exchanged particle depends on the initial/final channel.
• Avoid putting every kernel component or subdiagram inside its own decorative box. Use open topology, shared baselines, thin separators, and only physically meaningful filled blocks/vertices.
• Keep ordinary topology vertices consistent across a figure: either show all of them with the same small-blob style or omit all ordinary blobs. Reserve distinct blob sizes/fills for genuinely different physical objects.
• In hierarchy/equation figures, keep +, =, and arrow glyphs on their own whitespace lanes. They should never touch a diagram or read as a line label. Balance the visual distance from an operator to the diagrams on both sides; inconsistent operator gaps make the row read as hand-spaced rather than typeset.
• Draw hierarchy arrows to named panel/box boundaries so the arrowhead touches the target boundary cleanly. Do not leave arrows hovering above the target, and do not let them pierce into the physics content.
• Size panel boxes from the actual content bounding box with modest, consistent padding. Oversized bottom padding makes the figure read as disconnected modules and should be corrected in preview QA.
• Keep related topology glyphs on comparable vertical scales. Expanded loop/box subdiagrams may need a little more room than exchange diagrams, but their upper and lower external channels should not be separated so much more that the figure reads as a different scale.
• Align repeated channels by shared coordinates, not by eye.
• Do not assemble one continuous physical propagator from multiple nested tikzpicture nodes; tiny coordinate mismatches create visible breaks. Draw continuous physical lines in one coordinate system, and use nested nodes only for separate subdiagrams or algebraic factors.
• For loops, keep particle labels outside the loop on normal directions; keep arrow marks away from labels.
• For multi-panel diagrams, reserve a panel-label corner before placing physics labels.
• A white backing may cover an incidental line behind text; it must not hide a physical vertex, arrowhead, crossing, or topology-defining connection.

Verification Gate

Compilation is not completion. A diagram is deliverable only when all applicable checks pass:

• standalone source compiles to PDF without LaTeX errors;
• final vector PDF exists and is nonempty;
• high-DPI PNG preview is rendered and inspected;
• the agent states whether it can read images in this session;
• complex/submission-bound figures, multi-panel figures, nested-equation figures, and figures with more than six labels get an independent visual review when a reviewer surface is available;
• every label is legible at the intended manuscript width;
• no math label, panel label, or symbol overlaps a line, arrowhead, vertex, or another label;
• labels and arrowheads are not clipped;
• line styles, arrow semantics, colors, and block styles are consistent across panels;
• pdfimages -list reports no unexpected raster image objects for vector schematics;
• pdffonts reports embedded fonts when available;
• scratch logs, previews, and JSON reports live outside the repo unless the user explicitly asks to keep them.

Use references/quality-gate.md for the full checklist.

Pattern References

• Universal presentation rules shared with data/results figures — colour threading, CVD safety, label economy floor and ceiling, role-mapped typography, render-then-verify — live in the sibling figure-hygiene skill and apply to schematics too; this skill stays scoped to schematic/topological content.
• TikZ/feynhand templates and idioms: references/tikz-patterns.md.
• Python/Matplotlib geometry schematics and existing-data figure styling: references/python-geometry.md.
• Tool choice and non-default backends: references/tool-selection.md.
• QA and publication acceptance: references/quality-gate.md.

Smoke Examples

The default golden example is assets/examples/basic_process.tex: a deliberately simple, low-saturation process diagram with aligned labels and a vertex blob whose incident lines meet at the center but are hidden inside the filled circle.

assets/examples/feynhand_process.tex is the same visual grammar implemented with real tikz-feynhand vertices and \propag particle-line commands.

assets/examples/neutral_schematic.tex is a complex stress example for nested equation blocks and hierarchy arrows; do not treat it as the default visual style.

Run:

python3 skills/physics-diagrams/scripts/build_diagram.py \
  skills/physics-diagrams/assets/examples/basic_process.tex \
  --out-dir /tmp/physics-diagrams-basic

For direct-PDF routes such as Matplotlib:

python3 skills/physics-diagrams/scripts/build_diagram.py \
  --qa-only /tmp/physics-diagrams-matplotlib-example/geometry.pdf \
  --out-dir /tmp/physics-diagrams-geometry-qa

Then inspect the rendered preview before trusting the PDF.