AGENTS.md · diff
git:20260823.18e96ee to git:20260828.6e8a8ec
14 added, 2 removed. Audit A to A.
# CVPR-Figure — agent entry point
Portable instructions for agents that do not read `SKILL.md` frontmatter
(Codex, Cursor, Aider, plain API loops). Behaviour is identical; `SKILL.md` is
the same document with Claude Code routing metadata.
## What this is
A deterministic layout engine plus a house style for the framework, pipeline,
teaser and module figures in CVPR / ICCV / ECCV / NeurIPS / ICLR / ICML / AAAI /
ACL papers. Output is SVG, PDF, PNG, EMF, native Visio `.vsdx` and native
PowerPoint `.pptx`.
Pure Python standard library. No install step. Python 3.8+.
## Use it when
The user asks for an architecture diagram, method overview, framework figure,
pipeline diagram, teaser or paradigm-comparison figure, attention/token
diagram, module zoom-in, or an editable Visio/PowerPoint version of any of
those. Also: 论文框架图、流程图、方法图、网络结构图、teaser 图、Visio 可编辑图.
Do **not** use it for bar charts, curves, heatmaps, scatter plots or ablation
plots — those are matplotlib work.
## The loop
```bash
- # 1. start from the closest archetype
+ # 0. OPTIONAL -- if the user has the model code, draft from it instead of from
+ # a reading of it. Nothing is imported or executed; both readers use ast.
+ python3 scripts/from_code.py path/to/models/ --list
+ python3 scripts/from_code.py path/to/models/ --model MyNet -o figures/src/overview.yaml
+ python3 scripts/from_code.py configs/model.py --mm -o figures/src/overview.yaml
+ # The output is a first draft. Rename boxes to the paper's names, delete
+ # what the paper does not discuss, attach real crops, mark the contribution
+ # `role: core`. See references/from-code.md for the known limits.
+
+ # 1. otherwise start from the closest archetype
cp templates/pipeline-4stage.yaml figures/src/overview.yaml
# architecture: pipeline-4stage, module-detail, attention-tokens,
# streaming-worldmodel, dual-branch, gated-module,
# teaser-comparison
+ # with real data: framework-with-io, surroundview-pipeline,
+ # trainable-frozen, teacher-student, wrapped-pipeline
# analysis: blindspot-teaser, study-overview, factorial-2x2, taxonomy
# 2. edit the spec (see references/spec-language.md)
# 3. render
python3 scripts/render.py figures/src/overview.yaml -o figures/overview \
-f svg,pdf,png,vsdx,pptx --dpi 600
# 4. audit, and fix every FAIL
python3 scripts/validate.py figures/src/overview.yaml --svg figures/overview.svg
# 5. look at it
# (convert to PNG and open it; the auditor cannot judge communication)
```
## Read these, in this order
| File | When |
|---|---|
| `static/core/stance.md` | always, first |
| `static/core/contract.md` | always, before writing a spec |
| `references/archetypes.md` | choosing the figure's shape |
| `references/spec-language.md` | writing the spec |
| `references/house-style.md` | colours, sizes, stroke weights, and why |
| `references/case-studies.md` | matching a specific paper's figures |
| `references/anti-ai-checklist.md` | before delivering |
| `references/visio-workflow.md` | hand-editing the output |
| `references/latex-integration.md` | putting it in the paper |
## Hard rules
1. Never hand-write SVG or PDF coordinates. Write a spec.
2. Never generate the figure with an image model.
- 3. Type sizes are final printed points. Under 5.5 pt is a failure.
+ 3. Type sizes are final printed points. Under 5.0 pt is a failure,
+ 5.0-5.6 pt a warning you need a reason for.
4. Colour means something or it is grey. One concept, one colour, across the
whole paper.
5. `image` slots stay placeholders until the user supplies real crops. Never
invent an output render or a number.
6. Run the auditor before saying the figure is done.
## Environment notes
- `.svg`, `.vsdx`, `.pptx` are written by bundled code — always available.
- `.pdf` / `.png` / `.tiff` / `.emf` need one of: Inkscape, LibreOffice
(`soffice`), `rsvg-convert`, `cairosvg`, ImageMagick. `render.py` reports
what it used, takes rasters from the PDF, and verifies the pixel size
against `--dpi` (600 = camera-ready, 1200 = poster).
- PyYAML is used when importable; otherwise a bundled parser handles the same
subset. Specs may also be written as JSON.