grad-paragraph

Name: grad-paragraph
Author: WILLOSCAR

489stars39forks

How to install grad-paragraph?

willoscar/research-units-pipeline-skills/grad-paragraph

$npx -y skills add WILLOSCAR/research-units-pipeline-skills --skill grad-paragraph

Installs into the current project.

Summary

Write one survey-quality paragraph from evidence packs (tension → contrast → evaluation anchor → limitation). Trigger: grad paragraph, paragraph micro-structure, argument paragraph, 研究生段落, 论证段落, 对比段, 段落写作. Use when: you are drafting sections/S*.md (H3 body) and want sub

Files · 1

View on GitHub

SKILL.md

1	# Grad Paragraph (survey paragraph micro-skill)
2
3	Purpose: produce a single paragraph that reads like real survey prose, not “outline expansion”.
4
5	This is a writing micro-skill you can apply repeatedly inside `subsection-writer` (per H3 file under `sections/`).
6
7
8	## Role cards (use explicitly)
9
10	### Argument Planner
11
12	Mission: decide the paragraph’s tension/contrast/eval/limitation before writing.
13
14	Do:
15	- Write a 4-line plan (kept out of final prose).
16	- Ensure planned sentences can be grounded in evidence/citations.
17
18	Avoid:
19	- Writing from headings or generic axis labels.
20
21	### Paragraph Author
22
23	Mission: turn the plan into one content-bearing paragraph with embedded citations.
24
25	Do:
26	- Use explicit contrast markers (whereas/in contrast).
27	- Include at least one evaluation anchor token (task/metric/constraint).
28	- End with a limitation that changes interpretation.
29
30	Avoid:
31	- Narration and repeated template stems.
32
33
34	## Role prompt: Paragraph Author (argument move)
35
36	```text
37	You are writing one paragraph of a technical survey.
38
39	Your job is to perform one argument move under evidence:
40	- tension/question (why this matters here)
41	- explicit contrast (A vs B; not a list)
42	- evaluation anchor (task/metric/constraint)
43	- limitation (what breaks transfer or comparability)
44
45	Style:
46	- natural prose, content-bearing
47	- no narration (“This paragraph surveys…”)
48	- no repeated discourse stems across paragraphs
49
50	Constraints:
51	- do not invent facts or citations
52	- embed citations inside the sentence that needs them
53	- stay within the subsection’s citation scope
54	```
55
56	## What this paragraph must contain
57
58	In one paragraph (typically 4–6 sentences), cover:
59
60	- Tension / question: what this paragraph is trying to resolve (subsection-specific).
61	- Contrast: compare at least two approaches/routes/clusters (A vs B) with explicit contrast words.
62	- Evaluation anchor: name how comparisons are made (benchmark/dataset/metric/protocol), even if only abstract-level.
63	- Limitation / verification: state what is uncertain (missing protocol details, incomparable benchmarks, unclear constraints) without turning into boilerplate.
64	- If you include a number, also include: task type + metric definition + constraint (budget/cost/tool access), and cite it.
65
66	## Inputs (practical)
67
68	- `outline/subsection_briefs.jsonl` (for `rq`, `axes`, `clusters`, `paragraph_plan`)
69	- `outline/evidence_drafts.jsonl` (for evidence snippets + candidate comparisons)
70	- `outline/evidence_bindings.jsonl` (allowed citations for this H3)
71	- `citations/ref.bib`
72
73	## Outputs
74
75	- One paragraph (4–6 sentences) to paste into the target `sections/S<sub_id>.md` file.
76	- Optional (when debugging): a 4-line plan (tension/contrast/eval/limitation) kept out of the final prose.
77
78	## Roles (two-pass is more reliable)
79
80	### Role A: Argument Planner
81
82	Write a 4-line plan before prose:
83
84	1) Tension sentence (1 line)
85	2) Contrast sentence (1 line; A vs B)
86	3) Evaluation anchor sentence (1 line)
87	4) Limitation sentence (1 line)
88
89	Rules:
90	- Each line should be anchored by at least one citation key you intend to use.
91	- If evidence is abstract-only, avoid “dominant / clearly / state-of-the-art” style conclusions.
92
93	### Role B: Writer
94
95	Turn the plan into one natural paragraph.
96
97	Rules:
98	- Keep the paragraph subsection-specific (it should not be copy-pastable into other subsections).
99	- Place citations inside the sentence they support (not only at paragraph end).
100	- Do not mention pipeline internals (“working claim”, “axes we track”, “verification targets”).
101
102	## Paper voice (avoid template cadence)
103
104	- Keep tone calm and academic; avoid hype words (e.g., “clearly”, “obviously”, “breakthrough”).
105	- Vary sentence openings; don’t start every paragraph with the same connector (“However/Moreover/Taken together”).
106	- Avoid explicit labels like `Key takeaway:`; let the sentence carry the point.
107	- Prefer concrete nouns + mid-sentence ties (`...; however, ...`) over “PPT narration” signposting.
108
109
110
111	## Examples (what to write / what to avoid)
112
113	Bad (template narration + vague claims + cite dump):
114
115	```text
116	This subsection surveys how agents use memory. Taken together, these approaches improve performance across tasks [@example2023; @example2024; @example2025].
117	```
118
119	Why it is bad:
120	- Starts with narration ("This subsection ...").
121	- No explicit A-vs-B contrast (reads like a topic list).
122	- No evaluation anchor (benchmark/metric/protocol is missing).
123	- Citations are only used as a trailing tag list.
124
125	Good (tension -> contrast -> eval anchor -> limitation; citations embedded):
126
127	Plan (kept out of final prose):
128
129	1) Tension: Memory

People who liked this also viewed

6 picks

Type

grad-paragraph

489stars39forks

View on GitHub

How to install grad-paragraph?

willoscar/research-units-pipeline-skills/grad-paragraph

$npx -y skills add WILLOSCAR/research-units-pipeline-skills --skill grad-paragraph

Installs into the current project.

Summary

Files · 1

View on GitHub

SKILL.md

1	# Grad Paragraph (survey paragraph micro-skill)
2
3	Purpose: produce a single paragraph that reads like real survey prose, not “outline expansion”.
4
5	This is a writing micro-skill you can apply repeatedly inside `subsection-writer` (per H3 file under `sections/`).
6
7
8	## Role cards (use explicitly)
9
10	### Argument Planner
11
12	Mission: decide the paragraph’s tension/contrast/eval/limitation before writing.
13
14	Do:
15	- Write a 4-line plan (kept out of final prose).
16	- Ensure planned sentences can be grounded in evidence/citations.
17
18	Avoid:
19	- Writing from headings or generic axis labels.
20
21	### Paragraph Author
22
23	Mission: turn the plan into one content-bearing paragraph with embedded citations.
24
25	Do:
26	- Use explicit contrast markers (whereas/in contrast).
27	- Include at least one evaluation anchor token (task/metric/constraint).
28	- End with a limitation that changes interpretation.
29
30	Avoid:
31	- Narration and repeated template stems.
32
33
34	## Role prompt: Paragraph Author (argument move)
35
36	```text
37	You are writing one paragraph of a technical survey.
38
39	Your job is to perform one argument move under evidence:
40	- tension/question (why this matters here)
41	- explicit contrast (A vs B; not a list)
42	- evaluation anchor (task/metric/constraint)
43	- limitation (what breaks transfer or comparability)
44
45	Style:
46	- natural prose, content-bearing
47	- no narration (“This paragraph surveys…”)
48	- no repeated discourse stems across paragraphs
49
50	Constraints:
51	- do not invent facts or citations
52	- embed citations inside the sentence that needs them
53	- stay within the subsection’s citation scope
54	```
55
56	## What this paragraph must contain
57
58	In one paragraph (typically 4–6 sentences), cover:
59
60	- Tension / question: what this paragraph is trying to resolve (subsection-specific).
61	- Contrast: compare at least two approaches/routes/clusters (A vs B) with explicit contrast words.
62	- Evaluation anchor: name how comparisons are made (benchmark/dataset/metric/protocol), even if only abstract-level.
63	- Limitation / verification: state what is uncertain (missing protocol details, incomparable benchmarks, unclear constraints) without turning into boilerplate.
64	- If you include a number, also include: task type + metric definition + constraint (budget/cost/tool access), and cite it.
65
66	## Inputs (practical)
67
68	- `outline/subsection_briefs.jsonl` (for `rq`, `axes`, `clusters`, `paragraph_plan`)
69	- `outline/evidence_drafts.jsonl` (for evidence snippets + candidate comparisons)
70	- `outline/evidence_bindings.jsonl` (allowed citations for this H3)
71	- `citations/ref.bib`
72
73	## Outputs
74
75	- One paragraph (4–6 sentences) to paste into the target `sections/S<sub_id>.md` file.
76	- Optional (when debugging): a 4-line plan (tension/contrast/eval/limitation) kept out of the final prose.
77
78	## Roles (two-pass is more reliable)
79
80	### Role A: Argument Planner
81
82	Write a 4-line plan before prose:
83
84	1) Tension sentence (1 line)
85	2) Contrast sentence (1 line; A vs B)
86	3) Evaluation anchor sentence (1 line)
87	4) Limitation sentence (1 line)
88
89	Rules:
90	- Each line should be anchored by at least one citation key you intend to use.
91	- If evidence is abstract-only, avoid “dominant / clearly / state-of-the-art” style conclusions.
92
93	### Role B: Writer
94
95	Turn the plan into one natural paragraph.
96
97	Rules:
98	- Keep the paragraph subsection-specific (it should not be copy-pastable into other subsections).
99	- Place citations inside the sentence they support (not only at paragraph end).
100	- Do not mention pipeline internals (“working claim”, “axes we track”, “verification targets”).
101
102	## Paper voice (avoid template cadence)
103
104	- Keep tone calm and academic; avoid hype words (e.g., “clearly”, “obviously”, “breakthrough”).
105	- Vary sentence openings; don’t start every paragraph with the same connector (“However/Moreover/Taken together”).
106	- Avoid explicit labels like `Key takeaway:`; let the sentence carry the point.
107	- Prefer concrete nouns + mid-sentence ties (`...; however, ...`) over “PPT narration” signposting.
108
109
110
111	## Examples (what to write / what to avoid)
112
113	Bad (template narration + vague claims + cite dump):
114
115	```text
116	This subsection surveys how agents use memory. Taken together, these approaches improve performance across tasks [@example2023; @example2024; @example2025].
117	```
118
119	Why it is bad:
120	- Starts with narration ("This subsection ...").
121	- No explicit A-vs-B contrast (reads like a topic list).
122	- No evaluation anchor (benchmark/metric/protocol is missing).
123	- Citations are only used as a trailing tag list.
124
125	Good (tension -> contrast -> eval anchor -> limitation; citations embedded):
126
127	Plan (kept out of final prose):
128
129	1) Tension: Memory

People who liked this also viewed

6 picks

Type