figmirror

Name: figmirror
Author: VILA-Lab

495stars35forks

How to install figmirror?

vila-lab/figmirror/figmirror

$npx -y skills add VILA-Lab/FigMirror --skill figmirror

Installs into the current project.

Summary

This FigMirror skill should be used when the user asks to "mirror this figure's style", "copy this figure's style", "make a chart that looks like this paper", "reproduce this figure with my data", "match this paper's aesthetic", "I want a NeurIPS-quality version of this", or any

Files · 1

View on GitHub

SKILL.md

1	# FigMirror (`figmirror`)
2
3	Transfer the visual style of a top-conference paper figure (NeurIPS / ICML /
4	ICLR / Nature / Science) onto user data via an iterative Drawer / Reviewer
5	loop. Output is a self-contained matplotlib script + PNG + type-42 PDF that
6	matches the reference's STYLE — not its data.
7
8	## When to use
9
10	Trigger when the user provides all three:
11
12	- A reference paper-figure screenshot.
13	- Their own data in any parseable form (pasted table, CSV, TSV, markdown
14	table, or dirty terminal text).
15	- An expectation that the output should look like the reference — even via
16	casual phrasing ("make this chart but with my numbers", "redo this in
17	matplotlib"). This includes 3D references when the reference or data is
18	actually 3D.
19
20	Do not trigger on plain matplotlib chart requests with no reference image.
21
22	## Required inputs
23
24	- Reference image (PNG / JPG), cropped or uncropped. It may include margins,
25	captions, neighboring panels, or page text; Stage 0 preprocesses it.
26	- User data (any parseable form).
27	- Optional working directory. Default: `<cwd>/figmirror-runs/<run-id>/`.
28
29	## 3D Insert Gate
30
31	Enable `references/three-d-prompting.md` only when the user asks for a 3D
32	figure, the reference is visibly 3D, or the parsed data requires a 3D encoding
33	such as `x/y/z`, surfaces, trajectories, layered profiles, closed objects, 3D
34	small multiples, 3D bars, or plane projections. Do not use this insert to turn
35	an ordinary 2D task into 3D.
36
37	## Architecture
38
39	Three bundled subagents drive the loop; the caller orchestrates from the main
40	thread:
41
42	- `figure-preprocessor` — Preprocessor. Stage 0: preserves the raw upload,
43	crops away margins/captions/page text/neighboring panels when safe, and writes
44	`inputs/reference_clean.png` plus a crop check/report.
45	- `figure-illustrator` — Drawer. Per iter: reads reference + data + L2
46	library, produces `figure_iter<N>.py`, `img_iter<N>.png`,
47	`notes_iter<N>.md`, `floor_selfcheck_iter<N>.txt`. Self-checks the layout
48	floor before returning.
49	- `figure-critic` — Reviewer. Per iter: vision-only audit on a
50	fresh-context view (reference + draft + L2 library + optional 3D insert +
51	prior audit only). Returns ONE strict JSON object per the review schema.
52
53	Subagents are stateless across dispatch; iter-to-iter state flows through
54	workdir files.
55
56	Prefer `subagent_type: figure-preprocessor` / `figure-illustrator` /
57	`figure-critic`. Fallback path
58	when those names don't resolve: see `references/iter-loop-spec.md` §
59	"Subagent dispatch fallback".
60
61	## Workflow
62
63	For each run:
64
65	1. Stage workdir. Pre-create every directory the loop will write into
66	(subagents Write into existing dirs only — workspace permission quirk).
67	Stage the uploaded reference image to `inputs/reference_raw.png` and also
68	to `inputs/reference_clean.png` as a temporary first-paint copy; stage parsed
69	data to `inputs/data.txt`, and the L2 library to
70	`inputs/aesthetic-library.md`; stage the 3D router plus
71	`references/three-d/` only when the 3D insert gate is enabled. The router
72	selects exactly one mode file: `three-d/style-transfer.md` for ordinary
73	user-data figures, or `three-d/strict-reproduction.md` for reproduction,
74	comparison, or candidate/control replacement. For strict 3D reproduction runs
75	that need quantitative candidate diagnosis, also stage the optional candidate
76	scorer. The top-level Orchestrator owns final selection and must run the
77	selected mode's rendered-image gates before copying any candidate to the
78	final figure.
79	2. Preprocess reference. Dispatch `figure-preprocessor` before data-gen,
80	Drawer, or Reviewer. It writes the clean L1 anchor to
81	`inputs/reference_clean.png` and records the before/after crop check.
82	3. Echo data parse to user (Decision-7). Show parsed shape (rows × cols,
83	columns, NaN cells, sample row); proceed when confirmed, or skip if the
84	user pre-authorized. Either way, persist the echo to `data_echo.md`.
85	4. Iterate with the caller-provided `max_iters`; default to 6 when the
86	caller gives no explicit limit. If the caller enables auto-until-shipped,
87	ignore `max_iters` and continue until `ship` or a real blocker. Each iter:
88	- Dispatch the Drawer.
89	- Stage the Reviewer's audit view (reference + new draft + L2 library +

People who liked this also viewed

6 picks

Type

figmirror

495stars35forks

View on GitHub

How to install figmirror?

vila-lab/figmirror/figmirror

$npx -y skills add VILA-Lab/FigMirror --skill figmirror

Installs into the current project.

Summary

Files · 1

View on GitHub

SKILL.md

1	# FigMirror (`figmirror`)
2
3	Transfer the visual style of a top-conference paper figure (NeurIPS / ICML /
4	ICLR / Nature / Science) onto user data via an iterative Drawer / Reviewer
5	loop. Output is a self-contained matplotlib script + PNG + type-42 PDF that
6	matches the reference's STYLE — not its data.
7
8	## When to use
9
10	Trigger when the user provides all three:
11
12	- A reference paper-figure screenshot.
13	- Their own data in any parseable form (pasted table, CSV, TSV, markdown
14	table, or dirty terminal text).
15	- An expectation that the output should look like the reference — even via
16	casual phrasing ("make this chart but with my numbers", "redo this in
17	matplotlib"). This includes 3D references when the reference or data is
18	actually 3D.
19
20	Do not trigger on plain matplotlib chart requests with no reference image.
21
22	## Required inputs
23
24	- Reference image (PNG / JPG), cropped or uncropped. It may include margins,
25	captions, neighboring panels, or page text; Stage 0 preprocesses it.
26	- User data (any parseable form).
27	- Optional working directory. Default: `<cwd>/figmirror-runs/<run-id>/`.
28
29	## 3D Insert Gate
30
31	Enable `references/three-d-prompting.md` only when the user asks for a 3D
32	figure, the reference is visibly 3D, or the parsed data requires a 3D encoding
33	such as `x/y/z`, surfaces, trajectories, layered profiles, closed objects, 3D
34	small multiples, 3D bars, or plane projections. Do not use this insert to turn
35	an ordinary 2D task into 3D.
36
37	## Architecture
38
39	Three bundled subagents drive the loop; the caller orchestrates from the main
40	thread:
41
42	- `figure-preprocessor` — Preprocessor. Stage 0: preserves the raw upload,
43	crops away margins/captions/page text/neighboring panels when safe, and writes
44	`inputs/reference_clean.png` plus a crop check/report.
45	- `figure-illustrator` — Drawer. Per iter: reads reference + data + L2
46	library, produces `figure_iter<N>.py`, `img_iter<N>.png`,
47	`notes_iter<N>.md`, `floor_selfcheck_iter<N>.txt`. Self-checks the layout
48	floor before returning.
49	- `figure-critic` — Reviewer. Per iter: vision-only audit on a
50	fresh-context view (reference + draft + L2 library + optional 3D insert +
51	prior audit only). Returns ONE strict JSON object per the review schema.
52
53	Subagents are stateless across dispatch; iter-to-iter state flows through
54	workdir files.
55
56	Prefer `subagent_type: figure-preprocessor` / `figure-illustrator` /
57	`figure-critic`. Fallback path
58	when those names don't resolve: see `references/iter-loop-spec.md` §
59	"Subagent dispatch fallback".
60
61	## Workflow
62
63	For each run:
64
65	1. Stage workdir. Pre-create every directory the loop will write into
66	(subagents Write into existing dirs only — workspace permission quirk).
67	Stage the uploaded reference image to `inputs/reference_raw.png` and also
68	to `inputs/reference_clean.png` as a temporary first-paint copy; stage parsed
69	data to `inputs/data.txt`, and the L2 library to
70	`inputs/aesthetic-library.md`; stage the 3D router plus
71	`references/three-d/` only when the 3D insert gate is enabled. The router
72	selects exactly one mode file: `three-d/style-transfer.md` for ordinary
73	user-data figures, or `three-d/strict-reproduction.md` for reproduction,
74	comparison, or candidate/control replacement. For strict 3D reproduction runs
75	that need quantitative candidate diagnosis, also stage the optional candidate
76	scorer. The top-level Orchestrator owns final selection and must run the
77	selected mode's rendered-image gates before copying any candidate to the
78	final figure.
79	2. Preprocess reference. Dispatch `figure-preprocessor` before data-gen,
80	Drawer, or Reviewer. It writes the clean L1 anchor to
81	`inputs/reference_clean.png` and records the before/after crop check.
82	3. Echo data parse to user (Decision-7). Show parsed shape (rows × cols,
83	columns, NaN cells, sample row); proceed when confirmed, or skip if the
84	user pre-authorized. Either way, persist the echo to `data_echo.md`.
85	4. Iterate with the caller-provided `max_iters`; default to 6 when the
86	caller gives no explicit limit. If the caller enables auto-until-shipped,
87	ignore `max_iters` and continue until `ship` or a real blocker. Each iter:
88	- Dispatch the Drawer.
89	- Stage the Reviewer's audit view (reference + new draft + L2 library +

People who liked this also viewed

6 picks

Type