nsfc-humanization

Remove the "AI-generated tone" from NSFC grant proposal text, making it read as if written by a senior domain expert.

Skill Positioning

This skill focuses on text polishing, without changing content, supplementing information, or adjusting formatting. Input a section of proposal text with an "AI-generated tone", and output a polished version that sounds natural and demonstrates strong professional judgment.

Applicable to: Main text of various NSFC grant applications (both plain text and mixed LaTeX text are acceptable).

Optional Control Parameters (Intelligent Defaults if Not Provided by User)

To enhance controllability and cross-section consistency, users can explicitly declare the following parameters in their requests (can be described in natural Chinese or directly specify values):

Parameter	Values	Default	Function
`section_type`	`General` / `Rationale` / `Research Content` / `Research Foundation` / `Working Conditions` / `Risk Response` / `Other`	`General`	Section awareness: Different sections have different "expert-style" objectives
`field`	`general` / `cs` / `engineering` / `medicine` / `life_science`	`general`	Domain awareness: Adjust the focus of narration and judgment (but do not introduce terms/facts not present in the original text)
`strength`	`minimal` / `moderate` / `aggressive`	`minimal`	Strength control: Granularity of changes and rewrite allowances
`output_mode`	`text_only` / `text_with_change_summary` / `diagnosis_only` / `text_with_change_summary_and_style_card`	`text_only`	Output mode: Whether to attach "change summary/style card/diagnostic report"
`self_eval_rounds`	`1` / `2` (upper limit)	`1`	Upper limit of self-evaluation revision rounds (used to eliminate residual AI-generated tone)

Section Style Objectives (section_type)

Without adding new information, adjust expression focus according to section type:

```
Rationale
```
: Problem-driven + evidence chain + gap identification; avoid overemphasizing methodological details in high-level narratives
```
Research Content
```
: Clear task boundaries, verifiable steps, explicit comparison criteria; avoid writing "objectives" as "process completion"
```
Research Foundation
```
: Complete evidence chain of achievements, robust feasibility tone, clear boundary conditions; avoid exaggeration and unbounded assertions
```
Working Conditions
```
: Align resources/platforms/conditions with research content item by item, express in a more "implementable, verifiable" manner
```
Risk Response
```
: Risk → trigger conditions → impact → alternative plans/mitigation measures; pragmatic and actionable tone

Domain Style Objectives (field)

This parameter only affects "expression methods and judgment frameworks", and must not introduce domain terms, data, or facts not present in the original text:

```
cs
```
: Emphasize setup/comparison/boundaries and failure modes; use fewer vague "significance" statements, more "valid under which constraints" expressions
```
engineering
```
: Emphasize constraint conditions, implementable paths, and indicator criteria; avoid conceptual stacking
```
medicine
```
: Emphasize evidence hierarchy and conclusion boundaries; avoid presenting inferences as established facts (unless the original text does so)
```
life_science
```
: Emphasize causal boundaries and verifiability of mechanism chains; avoid generalized expressions

Strength Control (strength)

```
minimal
```
: Only modify obvious AI-generated tones (conjunction stacking/clichés/stylized enumeration/symmetric structures/template sentences), try not to change sentence structure
```
moderate
```
: Allow sentence rewriting and word order adjustment, but keep paragraph structure and line structure (line breaks/blank lines/indentation) unchanged
```
aggressive
```
: Allow intra-paragraph reorganization of expressions (e.g., merging/splitting clauses within sentences, reordering information), but still keep original paragraph and line structure unchanged, and must not add new information

Hard Constraints

LaTeX commands/environments/macros: Command names, environment names, and parameter structures must not be modified (retain the structure of
```
\xxx{...}
```
/
```
\begin{...}...\end{...}
```
)
List environment markers:
```
\begin{itemize}
```
/
```
\end{itemize}
```
,
```
\begin{enumerate}
```
/
```
\end{enumerate}
```
,
```
\item
```
keyword itself (but natural language after
```
\item
```
can be edited)
Citation and cross-reference tokens:
```
\cite{...}
```
,
```
\ref{...}
```
,
```
\label{...}
```
,
```
\eqref{...}
```
and their curly brace contents (keys/labels must be unchanged word for word)
Math mode:
```
$...$
```
,
```
$$...$$
```
,
```
$...$
```
,
```
\[...\]
```
, and content within math environments such as
```
equation/align/...
```
Comments: All content after
```
%
```
on the same line
Important "unchangeable strings": Numbers, units, variable names, abbreviations (case-sensitive), proper nouns, project/grant numbers, file paths, URLs, emails, DOIs
Special characters and escapes:
```
# $ % & _ { } ~ ^ \
```
etc. (including their escaped forms)

Protected vs. Editable Segments (Mandatory)

First, divide the input into two types of segments: Protected Segments (Uneditable) and Editable Segments (Polishable).

Protected Segments (Uneditable, Must Be Exact Matches)

LaTeX structures and control sequences: Command names and backslash sequences, environment names, curly brace/bracket structures themselves
List environment markers:
```
\begin{itemize}
```
/
```
\end{itemize}
```
,
```
\begin{enumerate}
```
/
```
\end{enumerate}
```
,
```
\item
```
keyword itself (but natural language after
```
\item
```
can be edited)
Citation and cross-reference tokens:
```
\cite{...}
```
,
```
\ref{...}
```
,
```
\label{...}
```
,
```
\eqref{...}
```
and their curly brace contents (keys/labels must be unchanged word for word)
Math mode:
```
$...$
```
,
```
$$...$$
```
,
```
$...$
```
,
```
\[...\]
```
, and content within math environments such as
```
equation/align/...
```
Comments: All content after
```
%
```
on the same line
Important "unchangeable strings": Numbers, units, variable names, abbreviations (case-sensitive), proper nouns, project/grant numbers, file paths, URLs, emails, DOIs
Special characters and escapes:
```
# $ % & _ { } ~ ^ \
```
etc. (including their escaped forms)

Editable Segments (Polishable, But Must Not Alter Facts or Structure)

Natural language text outside protected segments (including paragraph body text, and natural language parts within command parameter curly braces)
Natural language within command parameters such as
```
\caption{...}
```
/
```
\subsection{...}
```
/
```
\section{...}
```
: Editable (but command names and curly brace structures must remain unchanged)
List items: Natural language text after
```
\item
```
can be edited (list items are high-incidence areas for "stylized enumeration/clichés/symmetric structures" and should be prioritized for checking)
Allowed: Synonym replacement, minor sentence adjustments, minor word order adjustments, weakening template conjunctions
Prohibited: Adding uncertainties such as "unclear/Controversial/possible/speculative"; unless the original text explicitly expresses uncertainty/controversy

AI-generated Tone Identification Checklist

Polish is required when the following features are present (detailed comparison examples can be found in

references/machine-patterns.md

Stylized enumeration: Heavy use of "First... Second... Finally..." structures
Highly repetitive sentence patterns: Multiple sentences in a paragraph start with the same sentence pattern
Conjunction stacking: Frequent use of "therefore/thus/furthermore/in summary"
Flat wording: Lack of professional judgment tone, like stating a fact list
Lack of implicit consensus: Does not reflect "self-evident" judgments and trade-offs in the domain
Empty macro openings: Opening sentences like "With the rapid development of X, the problem of Y is increasingly important"
Importance clichés: Generic statements like "has important theoretical and practical value"
Excessive symmetric structures: Artificially creating symmetric frameworks such as "three elements" "four dimensions"
Lack of dialectical turns: The entire text is smooth, almost no thinking tension brought by "however/but"
Meta-comment word stacking: Frequent use of "It is worth noting that/It should be pointed out that/It is not difficult to find that"
Mechanical citation methods: Consecutive sentences like "Studies show[X]... Studies show[Y]...", lack of comprehensive interpretation
Template sentence stacking: Sentences like "This project intends to carry out... research on the basis of..." are repeated, with low information density
Passive voice overuse: Consecutive use of "is widely used in.../has been proven to...", diluting the subject and judgment
Number listing without interpretation: Consecutive listing of multiple data/improvement ranges, but lack of caliber consistency and comprehensive expression
Confusion between research objectives and research content: Writing objectives as steps, steps as objectives, unclear hierarchical relationships
Bracket nesting and information stacking: Stuffing data sources/data scales/annotations into the same pair of brackets, with semicolons/commas listing inside brackets (disrupts readability)

Senior Expert Writing Style

Polishing target style:

Diverse sentence patterns: Alternation of long and short sentences, avoiding single sentence patterns
Natural embedding of professional terms: No deliberate explanation, reflecting that the default reader is a peer
Visible trade-offs: Under the premise of not adding new information, make the existing priorities/trade-offs in the original text clearer (e.g., rewrite "the key lies in" into a more natural judgment sentence)
Natural logical transitions: Reduce explicit conjunctions, replace with semantic cohesion
Reflect domain implicit consensus: Reflect the judgment framework commonly recognized by researchers in the domain
Precise qualifiers: Know the boundaries of claims, use qualifiers like "under... conditions" "based on current evidence" instead of unbounded assertions
Acknowledge uncertainty: Only when the original text already expresses uncertainty/controversy, allow rewriting into more natural expressions (must not add out of thin air)
Narrative tension: First establish the problem, then introduce complexity, finally propose a solution, instead of flatly describing research steps
Lightweight brackets: Brackets only bear "short prompts", avoid stuffing multiple pieces of information into brackets, and even more avoid bracket nesting; when multiple pieces of information (especially with
```
;
```
) appear in brackets, prioritize rewriting into continuous narrative sentences (without adding new information)

Strength Control Reminder

Strength control is subject to

strength

; under any strength, "structure protection + zero semantic loss" must be observed, and avoid changing fact calibers in order to "sound more like an expert".

Input Format

NSFC grant proposal text segments, supporting:

Plain text paragraphs
Mixed LaTeX text (including commands, environments, etc.)

Recommendation: Input the entire proposal in batches by paragraph/section, which facilitates paragraph-by-paragraph verification of "structure protection + zero semantic loss".

Output Format

Line breaks/blank lines/indentation/list structure: Exactly the same as the original text (processed line by line, no automatic line breaks)
Only the text expression of editable segments can be changed; protected segments must be exact matches
LaTeX structure remains unchanged (commands/environments/citation keys/labels/mathematical content are not modified)

Additional Output (output_mode)

Default

text_only

: Only output the polished text (most suitable for direct pasting back into LaTeX source code).

When the user selects the following modes, append the corresponding content after the polished text (the polished text itself still retains the original format):

```
text_with_change_summary
```
: Append "Change Summary" (statistics of change types + representative change points, facilitating quick verification of semantic retention)
```
diagnosis_only
```
: Only output "Diagnostic Report" (no polished text output), including: Identified AI-generated tone patterns, severity, recommended strength/section type
```
text_with_change_summary_and_style_card
```
: Append "Change Summary" + "STYLE_CARD"

Recommended Change Summary Format

The change summary is used to help users quickly verify "zero semantic loss", and should be as short as verifiable:

Statistics of change types: e.g., "Removed 2 clichés, weakened 3 conjunction stacks, rewrote 1 sentence pattern (
```
strength=moderate
```
)"
Representative change points (1–5 items): Each item provides a fragment comparison of "original phrase → new phrase" (no more than 10 words/terms), avoiding long retellings
Risk reminder (optional): Only prompt "Sentences requiring manual confirmation" when ambiguity is found in the original text and polishing may trigger caliber deviation

Recommended Diagnostic Report Format

The diagnostic report is used to explain "where it sounds AI-generated, how much change is expected, and how to recommend modification" before polishing, and should include:

Identified patterns:

Pattern name + severity (low/medium/high) + triggering fragment (short)

Expected modification volume:
```
small/medium/large
```
Recommended settings:
```
section_type/field/strength/output_mode
```

STYLE_CARD (Cross-Paragraph Consistency Mechanism)

To solve cross-paragraph consistency (D7), when the output includes a STYLE_CARD:

If the user pastes the STYLE_CARD in subsequent batch inputs: Must follow the style card constraints first to ensure consistent overall readability of the same proposal
If the user does not provide a STYLE_CARD: Generate a style card with 6–10 "reusable style constraints" from the current polishing result without adding new information

Inapplicable Scenarios

Non-NSFC grant proposal content
Format or layout modification required
New research content supplementation required
Verification of scientific fact accuracy required

Examples

Detailed comparison examples can be found in

references/machine-patterns.md

Execution Process

Read the text provided by the user, and parse/infer parameters (
```
section_type
```
/
```
field
```
/
```
strength
```
/
```
output_mode
```
/
```
self_eval_rounds
```
)
If
```
output_mode=diagnosis_only
```
: First generate the "Diagnostic Report" and output it directly (no polishing)
Mark protected segments (LaTeX tokens/math/citation keys/labels/number units/comments, etc.), others are considered editable segments
Polish editable segments line by line: Execute according to
```
strength
```
, prioritize removing AI-generated tones and enhancing professional judgment expression
- Bracket rewriting priority: When brackets carry multiple pieces of information such as "data source + scale/scope/filter conditions", or bracket nesting/semicolon concatenation occurs, rewrite the bracket information into 1-3 consecutive narrative sentences (e.g., "Data was obtained from... The sample size is..."), and only keep necessary extremely short prompts in brackets
Structure self-check: Check line by line whether line breaks/indentation are retained; whether protected segments are exact matches
Semantic self-check: Must not add uncertainty, causality, comparison, conclusions; information not present in the original text must not appear
Style self-evaluation (mandatory, up to
```
self_eval_rounds
```
rounds, default 1; revise if problems are found, stop if no problems are found):
- Review item by item against the "AI-generated Tone Identification Checklist": Whether residual stylized enumeration/clichés/conjunction stacking/template sentences still exist
- Review against the "Senior Expert Writing Style": Whether it still "looks like listing facts rather than making judgments"
- If residual problems are found: Conduct a second round of minimal revisions without touching "structure protection + zero semantic loss" (must not introduce new information)
Output the polished result (retain original format); if required by
```
output_mode
```
, append change summary and/or STYLE_CARD

nsfc-humanization

NPX Install

Tags

SKILL.md Content (Chinese)

nsfc-humanization

Skill Positioning

Optional Control Parameters (Intelligent Defaults if Not Provided by User)

Section Style Objectives (section_type)

Domain Style Objectives (field)

Strength Control (strength)

Hard Constraints

Protected vs. Editable Segments (Mandatory)

Protected Segments (Uneditable, Must Be Exact Matches)

Editable Segments (Polishable, But Must Not Alter Facts or Structure)

AI-generated Tone Identification Checklist

Senior Expert Writing Style

Strength Control Reminder

Input Format

Output Format

Additional Output (output_mode)

Recommended Change Summary Format

Recommended Diagnostic Report Format

STYLE_CARD (Cross-Paragraph Consistency Mechanism)

Inapplicable Scenarios

Examples

Execution Process