section-writing-agent

Name: section-writing-agent
Author: Ar9av

41installs86forks

How to install section-writing-agent?

ar9av/paperorchestra/section-writing-agent

$npx -y skills add Ar9av/PaperOrchestra --skill section-writing-agent

Installs into the current project.

Summary

Step 4 of the PaperOrchestra pipeline (arXiv:2604.05018). ONE single multimodal LLM call that drafts the remaining paper sections (Abstract, Methodology, Experiments, Conclusion), extracts numeric values from experimental_log.md into LaTeX booktabs tables, splices the generated f

Files · 1

View on GitHub

SKILL.md

1	# Section Writing Agent (Step 4)
2
3	Faithful implementation of the Section Writing Agent from PaperOrchestra
4	(Song et al., 2026, arXiv:2604.05018, §4 Step 4, App. F.1 pp. 47–49).
5
6	Cost: ONE LLM call (App. B: "Section Writing Agent (1 call): A single,
7	comprehensive multimodal call to draft and compile the complete LaTeX
8	manuscript"). Do NOT split this into per-section calls — the paper
9	explicitly designs it as one comprehensive call so the model can maintain
10	global coherence across sections.
11
12	## Inputs
13
14	- `workspace/outline.json` — the master plan
15	- `workspace/inputs/idea.md` — technical details
16	- `workspace/inputs/experimental_log.md` — raw data for tables and qualitative analysis
17	- `workspace/drafts/intro_relwork.tex` — the template **with Intro + Related
18	Work already filled in by Step 3**. This is your starting point. The
19	preamble, package list, style, and the two pre-filled sections must be
20	preserved verbatim.
21	- `workspace/citation_pool.json` — the citation map (`{key, title, abstract}`
22	for each verified paper)
23	- `workspace/refs.bib` — the BibTeX file
24	- `workspace/inputs/conference_guidelines.md` — formatting rules
25	- `workspace/figures/` — the actual PNG files from Step 2 (used as
26	multimodal vision input!)
27	- `workspace/figures/captions.json` — caption text per figure_id
28	- `workspace/tex_profile.json` — TeX package availability flags (written by
29	`check_tex_packages.py` at Step 0). **Read this before generating any
30	LaTeX.** It tells you which packages are installed so you select the right
31	cross-reference pattern, font packages, etc. before you write — not after
32	you try to compile.
33
34	## Output
35
36	- `workspace/drafts/paper.tex` — the complete LaTeX paper, with all sections
37	filled. The Step 5 Refinement Agent will iterate on this file.
38
39	## How to do it
40
41	### 0.5. Read tex_profile.json and select LaTeX patterns
42
43	Before composing the prompt, read `workspace/tex_profile.json` and apply
44	these rules to every LaTeX choice in the generated paper:
45
46	\| Profile flag \| True → use \| False → use instead \|
47	\|---\|---\|---\|
48	\| `use_cleveref` \| `\cref{fig:X}`, `\cref{tab:Y}` \| `Figure~\ref{fig:X}`, `Table~\ref{tab:Y}` \|
49	\| `use_nicefrac` \| `\nicefrac{a}{b}` \| `$a/b$` \|
50	\| `use_microtype` \| `\usepackage{microtype}` \| omit the line \|
51	\| `use_t1_fontenc` \| `\usepackage[T1]{fontenc}` \| omit the line \|
52
53	If `tex_profile.json` does not exist (old workspace), default to the safe
54	fallback column (no cleveref, no nicefrac, no microtype, no T1 fontenc).
55
56	### 1. Pre-extract metrics from the experimental log
57
58	Run the deterministic helper:
59
60	```bash
61	python skills/section-writing-agent/scripts/extract_metrics.py \
62	--log workspace/inputs/experimental_log.md \
63	--out workspace/metrics.json
64	```
65
66	This parses the `## 2. Raw Numeric Data` section's markdown tables into
67	structured JSON. The Section Writing Agent uses this to construct LaTeX
68	booktabs tables without re-deriving values from raw text. Read
69	`references/latex-table-patterns.md` for the booktabs conventions.
70
71	### 2. Compose the prompt and make ONE multimodal call
72
73	Load `references/prompt.md` (verbatim Section Writing Agent prompt from App.
74	F.1). Prepend the Anti-Leakage Prompt from
75	`../paper-orchestra/references/anti-leakage-prompt.md`.
76
77	The user message contains:
78
79	- `outline.json` — full content
80	- `idea.md` — full content
81	- `experimental_log.md` — full content (tables AND prose)
82	- `intro_relwork.tex` — full content (this becomes `template.tex` for the prompt)
83	- `citation_pool.json` — full content (becomes `citation_map.json`)
84	- `conference_guidelines.md` — full content
85	- `figures_list` — array of `{figure_id, filename, caption}` from
86	`captions.json` and the file listing
87	- The actual figure PNGs as multimodal image inputs, so the model can
88	visually inspect them and write accurate descriptions / refer to them
89	correctly in the prose.
90
91	If your host LLM has no vision input, fall back to text-only mode: pass the
92	captions in `captions.json` as descriptions and tell the agent it cannot see
93	the images directly. Quality drops noticeably (the paper notes that visual
94	grounding measurably improves figure-text alignment), but the pipeline
95	still completes.
96
97	### 3. Save the output
98
99	The agent's response is wrapped in `\`\`\`latex ... \`\`\`` fences. Extract
100	the LaTeX code and save to `workspace/drafts/paper.tex`.
101
102	### 4. Run the deterministic gates
103
104	```bash
105	# Orphan citation gate: ever

People who liked this also viewed

6 picks

Type

section-writing-agent

41installs86forks

View on GitHub

How to install section-writing-agent?

ar9av/paperorchestra/section-writing-agent

$npx -y skills add Ar9av/PaperOrchestra --skill section-writing-agent

Installs into the current project.

Summary

Files · 1

View on GitHub

SKILL.md

1	# Section Writing Agent (Step 4)
2
3	Faithful implementation of the Section Writing Agent from PaperOrchestra
4	(Song et al., 2026, arXiv:2604.05018, §4 Step 4, App. F.1 pp. 47–49).
5
6	Cost: ONE LLM call (App. B: "Section Writing Agent (1 call): A single,
7	comprehensive multimodal call to draft and compile the complete LaTeX
8	manuscript"). Do NOT split this into per-section calls — the paper
9	explicitly designs it as one comprehensive call so the model can maintain
10	global coherence across sections.
11
12	## Inputs
13
14	- `workspace/outline.json` — the master plan
15	- `workspace/inputs/idea.md` — technical details
16	- `workspace/inputs/experimental_log.md` — raw data for tables and qualitative analysis
17	- `workspace/drafts/intro_relwork.tex` — the template **with Intro + Related
18	Work already filled in by Step 3**. This is your starting point. The
19	preamble, package list, style, and the two pre-filled sections must be
20	preserved verbatim.
21	- `workspace/citation_pool.json` — the citation map (`{key, title, abstract}`
22	for each verified paper)
23	- `workspace/refs.bib` — the BibTeX file
24	- `workspace/inputs/conference_guidelines.md` — formatting rules
25	- `workspace/figures/` — the actual PNG files from Step 2 (used as
26	multimodal vision input!)
27	- `workspace/figures/captions.json` — caption text per figure_id
28	- `workspace/tex_profile.json` — TeX package availability flags (written by
29	`check_tex_packages.py` at Step 0). **Read this before generating any
30	LaTeX.** It tells you which packages are installed so you select the right
31	cross-reference pattern, font packages, etc. before you write — not after
32	you try to compile.
33
34	## Output
35
36	- `workspace/drafts/paper.tex` — the complete LaTeX paper, with all sections
37	filled. The Step 5 Refinement Agent will iterate on this file.
38
39	## How to do it
40
41	### 0.5. Read tex_profile.json and select LaTeX patterns
42
43	Before composing the prompt, read `workspace/tex_profile.json` and apply
44	these rules to every LaTeX choice in the generated paper:
45
46	\| Profile flag \| True → use \| False → use instead \|
47	\|---\|---\|---\|
48	\| `use_cleveref` \| `\cref{fig:X}`, `\cref{tab:Y}` \| `Figure~\ref{fig:X}`, `Table~\ref{tab:Y}` \|
49	\| `use_nicefrac` \| `\nicefrac{a}{b}` \| `$a/b$` \|
50	\| `use_microtype` \| `\usepackage{microtype}` \| omit the line \|
51	\| `use_t1_fontenc` \| `\usepackage[T1]{fontenc}` \| omit the line \|
52
53	If `tex_profile.json` does not exist (old workspace), default to the safe
54	fallback column (no cleveref, no nicefrac, no microtype, no T1 fontenc).
55
56	### 1. Pre-extract metrics from the experimental log
57
58	Run the deterministic helper:
59
60	```bash
61	python skills/section-writing-agent/scripts/extract_metrics.py \
62	--log workspace/inputs/experimental_log.md \
63	--out workspace/metrics.json
64	```
65
66	This parses the `## 2. Raw Numeric Data` section's markdown tables into
67	structured JSON. The Section Writing Agent uses this to construct LaTeX
68	booktabs tables without re-deriving values from raw text. Read
69	`references/latex-table-patterns.md` for the booktabs conventions.
70
71	### 2. Compose the prompt and make ONE multimodal call
72
73	Load `references/prompt.md` (verbatim Section Writing Agent prompt from App.
74	F.1). Prepend the Anti-Leakage Prompt from
75	`../paper-orchestra/references/anti-leakage-prompt.md`.
76
77	The user message contains:
78
79	- `outline.json` — full content
80	- `idea.md` — full content
81	- `experimental_log.md` — full content (tables AND prose)
82	- `intro_relwork.tex` — full content (this becomes `template.tex` for the prompt)
83	- `citation_pool.json` — full content (becomes `citation_map.json`)
84	- `conference_guidelines.md` — full content
85	- `figures_list` — array of `{figure_id, filename, caption}` from
86	`captions.json` and the file listing
87	- The actual figure PNGs as multimodal image inputs, so the model can
88	visually inspect them and write accurate descriptions / refer to them
89	correctly in the prose.
90
91	If your host LLM has no vision input, fall back to text-only mode: pass the
92	captions in `captions.json` as descriptions and tell the agent it cannot see
93	the images directly. Quality drops noticeably (the paper notes that visual
94	grounding measurably improves figure-text alignment), but the pipeline
95	still completes.
96
97	### 3. Save the output
98
99	The agent's response is wrapped in `\`\`\`latex ... \`\`\`` fences. Extract
100	the LaTeX code and save to `workspace/drafts/paper.tex`.
101
102	### 4. Run the deterministic gates
103
104	```bash
105	# Orphan citation gate: ever

People who liked this also viewed

6 picks

Type