• Only write to:

  extraTex/1.1.立项依据.tex

• Prohibited modifications:

  main.tex

  ,

  extraTex/@config.tex

  , any

  .cls/.sty

  files
• Writing goal: Clearly explain "why this project needs to be done" and lay the groundwork for " (II) Research Content" with "scientific questions/hypotheses and entry points".
• AI dependency: By default, uses the native intelligence of Claude Code / Codex provided by the runtime environment (no need to configure any external API Key; will automatically fall back to hard-coded capabilities if AI is unavailable)
• Theoretical innovation orientation (default): Prioritize the falsifiability of scientific questions/hypotheses, clarity of theoretical contributions, and completeness of verification dimensions (see theoretical_innovation_guidelines.md)
• Configurable writing orientation: Switch using

  style.mode=theoretical|mixed|engineering

  in

  skills/nsfc-justification-writer/config.yaml

  (default is

  theoretical

  )

• If not provided by the user, first collect/supplement: references/info_form.md

Recommendation: Use the script to quickly generate the information form (with interactive filling), see

skills/nsfc-justification-writer/scripts/run.py init

.

1. Locate project and target file: Confirm

   project_root

   , read and only edit

   extraTex/1.1.立项依据.tex

   .
2. Extract existing skeleton: If the file already has subheadings like

   \subsubsection

   , prioritize retaining the skeleton and only replace the body paragraphs (unless the user requests restructuring). By default, do not enforce exact title matching (

   strict_title_match=false

   ), focus more on "whether content dimensions are covered".
3. Progressive writing guidance (recommended): Skeleton first → then paragraphs → then revision → then polishing → then acceptance (avoid the pressure of one-step completion)
   • Use

     scripts/run.py coach --stage auto

     to automatically judge the current stage and provide "only three tasks for this round + questions you need to supplement + copyable prompts"
   • Only modify the body of one

     \subsubsection

     per round, use

     apply-section

     for safe writing and automatic backup
4. Generate the main narrative of "Project Justification" (recommended 4-paragraph closed loop; AI will check content dimension coverage instead of rigidly adhering to titles):
   • Value and Necessity: Pain points → scope of impact/cost → why it must be done now.
   • Current Status and Limitations: Mainstream approaches/representative works → 2–4 clear limitations (try to be quantifiable/verifiable).
   • Scientific Questions/Core Hypotheses: One hypothesis + 1–3 key scientific questions (breakpoint style), oriented towards "verifiability".
   • Project Entry Point and Contributions: The "differentiated entry point" of this project compared to existing work, with one transition sentence to the research content.
5. Verifiability and Citation Protection:
   • AI semantically identifies "expressions that may cause reviewer discomfort" (absolute statements/filling gaps/unsubstantiated exaggerations/self-qualification) and provides rewriting suggestions; hard-coded high-risk words are only for prompts, not mechanical blocking.
   • Do not write unprovable statements like "internationally leading/first in China"; when citing external works, first ask the user to provide DOI/link or call

     nsfc-bib-manager

     for verification before writing

     \cite{...}

     .
6. Cross-section Consistency Check: Check if terms/abbreviations/indicator calibers align with

   2.1 Research Content

   ; if necessary, list 3–5 key nouns and indicators for user confirmation.
7. Target Word Count Analysis: Prioritize parsing the user's intent/"word count"/"± range"/interval description in the information form; use configuration defaults only if there is no explicit instruction.

• Configuration validation:

  python skills/nsfc-justification-writer/scripts/run.py validate-config

• Large File Tier2:

  diagnose/review --tier2 --chunk-size 12000 --max-chunks 20

  (supports

  .cache/ai

  caching; ultra-large files will prioritize streaming chunking to reduce peak memory usage; use

  --fresh

  to force recalculation)
• Note: The script layer of this repository will not "directly connect to external large models by default"; whether AI capabilities are available depends on whether the runtime environment injects a responder (will automatically fall back to hard-coded capabilities if unavailable)
• Related design instructions:
  • Content dimension coverage check:

    skills/nsfc-justification-writer/references/dimension_coverage_design.md

  • Identification and rewriting of "expressions that may cause reviewer discomfort":

    skills/nsfc-justification-writer/references/boastful_expression_guidelines.md

  • Theoretical Innovation Orientation Writing Guidelines:

    skills/nsfc-justification-writer/references/theoretical_innovation_guidelines.md

    (includes warnings on misused methodological terms)
  • Comparison Examples of Misused Methodological Terms:

    skills/nsfc-justification-writer/references/methodology_term_examples.md

    (newly added)

prompts.*

config.yaml

preset.yaml

override.yaml

• File path: e.g.,

  prompts/tier2_diagnostic.txt

• Direct multi-line Prompt: Write multi-line text in YAML using

  |

  (suitable for adjusting focus in different fields)

--preset medical

prompts.tier2_diagnostic_medical

terminology.dimensions

Whether AI is available depends on whether the runtime environment injects a responder; use

skills/nsfc-justification-writer/scripts/run.py check-ai

for self-test.

• Tier1 Hard-Coded Diagnosis: Structure (≥4

  \subsubsection

  s) / whether citation keys exist in

  .bib

  / DOI missing and format abnormality prompts / word count statistics / high-risk expression prompts and dangerous command scanning
• Content Dimension Coverage Check (AI): Does not depend on title wording, checks whether "value and necessity/current limitations/scientific questions/entry points" are covered
• Boastful Expression Recognition (AI): Identifies absolute statements/gap-filling/unsubstantiated exaggerations/self-qualification, outputs rewriting suggestions
• Cross-Section Consistency Matrix: Provides cross-section consistency prompts based on

  terminology.dimensions

  (research objects/indicators/terms) in

  config.yaml

• AI Term Consistency (Optional): When AI is available and

  terminology.mode=auto/ai

  , additionally provides semantic perspective checks and modification suggestions for "synonym/abbreviation misuse" (only outputs matrix if unavailable)
• Safe Writing Tool: Precisely locates and replaces the body text by

  \subsubsection{...}

  , writes to whitelisted files + backup (products are stored in

  skills/nsfc-justification-writer/runs/

  )
• Pre-Writing Quality Gate (Optional):

  apply-section --strict-quality

  only scans the "newly added body text" for high-risk words/dangerous commands; if AI is available, it adds semantic blocking of "boastful expressions" to avoid being stuck by historical content
• Reviewer Suggestion Generator: Based on DoD + diagnosis results, outputs "what reviewers will ask + how to modify" (

  scripts/run.py review

  )
• Visual HTML Diagnosis Report: Quickly locate issues (

  scripts/run.py diagnose --html-report auto

  )
• Version Diff/Rollback: View differences and one-click rollback based on runs backups (

  scripts/run.py diff/rollback

  )
• Example Recommendation: Read

  *.metadata.yaml

  keywords from

  examples/

  to assist in matching reference skeletons by topic (

  scripts/run.py coach --topic ...

  /

  scripts/run.py examples

  )
• AI Example Recommendation (Optional): When AI is available, prioritize semantic matching and provide "recommendation reasons" (falls back to keyword/category heuristics if unavailable)
• AI Stage Judgment (Optional): When

  --stage auto

  is used in coach, AI can comprehensively infer "skeleton/draft/revise/polish/final" based on word count/structure/quality status; falls back to hard-coded thresholds if AI is unavailable
• Configuration Override and Presets: Supports

  --preset medical/engineering

  and

  ~/.config/nsfc-justification-writer/override.yaml

  to override parameters such as term dimensions (can be turned off with

  --no-user-override

  if needed)

skills/nsfc-justification-writer/scripts/run.py

skills/nsfc-justification-writer/scripts/README.md

systematic-literature-review

1. Contains a hidden folder

   .systematic-literature-review

   , which includes

   {topic}_review.tex

   and

   {topic}_参考文献.bib

   /

   references.bib

   files (running pipeline)
2. Contains typical file combinations:

   {topic}_review.tex

   +

   {topic}_参考文献.bib

   /

   references.bib

   +

   {topic}_工作条件.md

   (completed output directory)
3. Contains files with the same prefix:

   {topic}_review.tex

   and

   {topic}_参考文献.bib

   (based on filename prefix matching)

systematic-literature-review

• Read-only mode: Only read the content of

  .tex

  and

  .bib

  files
• No writing allowed: Will not modify any files in the directory
• Citation verification: Automatically verify whether citations in

  .tex

  are consistent with definitions in

  .bib

• The user requests to cite existing literature review content
• Need to extract research status information from systematic reviews
• Want to ensure citation consistency

• Directory detection:

  detect_slr_directory(path)

  judges whether it is a

  systematic-literature-review

  directory
• Directory analysis:

  analyze_review_directory(path)

  returns directory structure information
• Citation verification:

  validate_citation_consistency(tex_path, bib_path)

  checks citation consistency
• Content extraction: Extract key information from

  .tex

  and

  .bib

  files

core/review_integration.py

• See: references/dod_checklist.md

• This skill does not maintain change history in this document; it is uniformly recorded in the root-level

  CHANGELOG.md

  .
• Version numbers are only maintained in

  skills/nsfc-justification-writer/config.yaml

  (

  skill_info.version

  ) and the frontmatter of this file to avoid inconsistent caliber.