plotting-agent

Name: plotting-agent
Author: Ar9av

43installs86forks

How to install plotting-agent?

ar9av/paperorchestra/plotting-agent

$npx -y skills add Ar9av/PaperOrchestra --skill plotting-agent

Installs into the current project.

Summary

Step 2 of the PaperOrchestra pipeline (arXiv:2604.05018). Execute the visualization plan from outline.json — render plots and conceptual diagrams from experimental_log.md and idea.md, optionally refine via VLM critique loop, and produce context-aware captions. Runs in parallel wi

Files · 1

View on GitHub

SKILL.md

1	# Plotting Agent (Step 2)
2
3	Faithful implementation of the Plotting Agent from PaperOrchestra
4	(Song et al., 2026, arXiv:2604.05018, §4 Step 2 and App. F.1 p.45).
5
6	Cost: ~20–30 LLM calls. The paper uses PaperBanana (Zhu et al., 2026) as
7	the default backbone with a closed-loop VLM-critique refinement. This skill
8	expresses that loop in host-agent terms: you (the host agent) generate
9	matplotlib code with your own LLM, render via your Bash/Python tool,
10	optionally critique the rendered PNG with your vision model, redraw, and
11	finally caption.
12
13	## Inputs
14
15	- `workspace/outline.json` — specifically the `plotting_plan` array
16	- `workspace/inputs/idea.md` and `workspace/inputs/experimental_log.md` —
17	the source data
18	- `workspace/inputs/figures/` — optional pre-existing figures (`PlotOn` mode)
19
20	## Outputs
21
22	- `workspace/figures/<figure_id>.png` — one PNG per `plotting_plan` entry
23	(300 DPI, sized to the requested aspect ratio)
24	- `workspace/figures/captions.json` — `{figure_id: caption_text}` map
25
26	## Workflow
27
28	### Per figure (executed independently per `figure_id`)
29
30	1. Read the figure spec from `outline.json`:
31	```json
32	{
33	"figure_id": "fig_main_results",
34	"title": "Main Results on Dataset X",
35	"plot_type": "plot",
36	"data_source": "experimental_log.md",
37	"objective": "Visual summary (Grouped Bar Chart) demonstrating ...",
38	"aspect_ratio": "5:4"
39	}
40	```
41
42	2. Few-shot retrieval (visual planning): pick the matching pattern from
43	`references/chart-patterns.md` (for `plot_type=="plot"`) or
44	`references/diagram-patterns.md` (for `plot_type=="diagram"`).
45
46	3. Extract data: parse `idea.md` and/or `experimental_log.md`
47	(`data_source` field tells you which) to obtain the numeric values or
48	conceptual entities the figure needs. For `experimental_log.md`, the
49	`## 2. Raw Numeric Data` section contains markdown tables.
50
51	4. Render:
52
53	If `PAPERBANANA_PATH` is set — use the PaperBanana backbone
54	(Zhu et al., 2026). It runs a Retriever → Planner → Stylist → Visualizer
55	→ Critic loop and is especially good for `plot_type == "diagram"`.
56	See `references/paperbanana-cookbook.md` for setup (needs a Gemini API key).
57
58	```bash
59	python skills/plotting-agent/scripts/paperbanana_render.py \
60	--figure-id <figure_id> \
61	--caption "<objective from figure spec>" \
62	--content-file workspace/inputs/idea.md \
63	--task <diagram\|plot> \
64	--aspect-ratio <aspect_ratio> \
65	--out workspace/figures/<figure_id>.png
66	```
67
68	Otherwise — write a matplotlib script and run it via your Bash tool,
69	or use the bundled helper:
70	```bash
71	python skills/plotting-agent/scripts/render_matplotlib.py \
72	--spec spec.json \
73	--out workspace/figures/<figure_id>.png
74	```
75	The script must apply the academic style from `chart-patterns.md`, use the
76	correct pixel size from `aspect-ratios.md`, save at 300 DPI, and call
77	`plt.close()` after `savefig`.
78
79	5. VLM critique loop (optional, only if your host has vision):
80	- Reload the rendered PNG as a multimodal input to your LLM.
81	- Critique it against the figure's `objective` from the outline. Look for:
82	visual artifacts, mislabeled axes, illegible text, color clashes,
83	misleading scaling, missing legend, overlapping labels.
84	- If problems are found, regenerate the matplotlib script with corrections
85	and re-render. Cap at 3 critique iterations per figure.
86	- This is the closed-loop refinement step the paper inherits from
87	PaperBanana. See `references/plotting-pipeline.md` for the full loop
88	description.
89	- If your host has no vision input, skip this step entirely. The
90	figure will still render correctly, just without iterative refinement.
91
92	6. Generate the caption using the verbatim Caption Generation prompt at
93	`references/caption-prompt.md`. Inputs to the caption prompt:
94	- `task_name` — the section the figure belongs to (e.g., "Methodology",
95	"Experiments")
96	- `raw_content` — the surrounding section text (or content_bullets from
97	the section_plan if the section isn't drafted yet)
98	- `description` — the `objective` field from the figure spec
99	- `figure_desc` — a 1-sentence description of what the rendered figure
100	actually shows (from your VLM critique pass, or from the script's plan
101	if no vision)
102
103	Write the caption to `workspace/figures/captions.json` keyed by
104	`figure_id`. **Captions must NOT contain `Figu

People who liked this also viewed

6 picks

Type

plotting-agent

43installs86forks

View on GitHub

How to install plotting-agent?

ar9av/paperorchestra/plotting-agent

$npx -y skills add Ar9av/PaperOrchestra --skill plotting-agent

Installs into the current project.

Summary

Files · 1

View on GitHub

SKILL.md

1	# Plotting Agent (Step 2)
2
3	Faithful implementation of the Plotting Agent from PaperOrchestra
4	(Song et al., 2026, arXiv:2604.05018, §4 Step 2 and App. F.1 p.45).
5
6	Cost: ~20–30 LLM calls. The paper uses PaperBanana (Zhu et al., 2026) as
7	the default backbone with a closed-loop VLM-critique refinement. This skill
8	expresses that loop in host-agent terms: you (the host agent) generate
9	matplotlib code with your own LLM, render via your Bash/Python tool,
10	optionally critique the rendered PNG with your vision model, redraw, and
11	finally caption.
12
13	## Inputs
14
15	- `workspace/outline.json` — specifically the `plotting_plan` array
16	- `workspace/inputs/idea.md` and `workspace/inputs/experimental_log.md` —
17	the source data
18	- `workspace/inputs/figures/` — optional pre-existing figures (`PlotOn` mode)
19
20	## Outputs
21
22	- `workspace/figures/<figure_id>.png` — one PNG per `plotting_plan` entry
23	(300 DPI, sized to the requested aspect ratio)
24	- `workspace/figures/captions.json` — `{figure_id: caption_text}` map
25
26	## Workflow
27
28	### Per figure (executed independently per `figure_id`)
29
30	1. Read the figure spec from `outline.json`:
31	```json
32	{
33	"figure_id": "fig_main_results",
34	"title": "Main Results on Dataset X",
35	"plot_type": "plot",
36	"data_source": "experimental_log.md",
37	"objective": "Visual summary (Grouped Bar Chart) demonstrating ...",
38	"aspect_ratio": "5:4"
39	}
40	```
41
42	2. Few-shot retrieval (visual planning): pick the matching pattern from
43	`references/chart-patterns.md` (for `plot_type=="plot"`) or
44	`references/diagram-patterns.md` (for `plot_type=="diagram"`).
45
46	3. Extract data: parse `idea.md` and/or `experimental_log.md`
47	(`data_source` field tells you which) to obtain the numeric values or
48	conceptual entities the figure needs. For `experimental_log.md`, the
49	`## 2. Raw Numeric Data` section contains markdown tables.
50
51	4. Render:
52
53	If `PAPERBANANA_PATH` is set — use the PaperBanana backbone
54	(Zhu et al., 2026). It runs a Retriever → Planner → Stylist → Visualizer
55	→ Critic loop and is especially good for `plot_type == "diagram"`.
56	See `references/paperbanana-cookbook.md` for setup (needs a Gemini API key).
57
58	```bash
59	python skills/plotting-agent/scripts/paperbanana_render.py \
60	--figure-id <figure_id> \
61	--caption "<objective from figure spec>" \
62	--content-file workspace/inputs/idea.md \
63	--task <diagram\|plot> \
64	--aspect-ratio <aspect_ratio> \
65	--out workspace/figures/<figure_id>.png
66	```
67
68	Otherwise — write a matplotlib script and run it via your Bash tool,
69	or use the bundled helper:
70	```bash
71	python skills/plotting-agent/scripts/render_matplotlib.py \
72	--spec spec.json \
73	--out workspace/figures/<figure_id>.png
74	```
75	The script must apply the academic style from `chart-patterns.md`, use the
76	correct pixel size from `aspect-ratios.md`, save at 300 DPI, and call
77	`plt.close()` after `savefig`.
78
79	5. VLM critique loop (optional, only if your host has vision):
80	- Reload the rendered PNG as a multimodal input to your LLM.
81	- Critique it against the figure's `objective` from the outline. Look for:
82	visual artifacts, mislabeled axes, illegible text, color clashes,
83	misleading scaling, missing legend, overlapping labels.
84	- If problems are found, regenerate the matplotlib script with corrections
85	and re-render. Cap at 3 critique iterations per figure.
86	- This is the closed-loop refinement step the paper inherits from
87	PaperBanana. See `references/plotting-pipeline.md` for the full loop
88	description.
89	- If your host has no vision input, skip this step entirely. The
90	figure will still render correctly, just without iterative refinement.
91
92	6. Generate the caption using the verbatim Caption Generation prompt at
93	`references/caption-prompt.md`. Inputs to the caption prompt:
94	- `task_name` — the section the figure belongs to (e.g., "Methodology",
95	"Experiments")
96	- `raw_content` — the surrounding section text (or content_bullets from
97	the section_plan if the section isn't drafted yet)
98	- `description` — the `objective` field from the figure spec
99	- `figure_desc` — a 1-sentence description of what the rendered figure
100	actually shows (from your VLM critique pass, or from the script's plan
101	if no vision)
102
103	Write the caption to `workspace/figures/captions.json` keyed by
104	`figure_id`. **Captions must NOT contain `Figu

People who liked this also viewed

6 picks

Type