Check Review Alignment

Applicable Scenarios

You have
```
{topic}_review.tex
```
and the corresponding
```
.bib
```
file, and need to verify whether each in-text citation "actually cites the paper", and make minimal rewrites if necessary to eliminate mismatched/hallucinated citations.
Need to automatically render and generate PDF and Word after optimization.

Input

```
work_dir
```
: The working directory of the review (contains
```
{topic}_review.tex
```
and
```
.bib
```
, may include PDFs).
Optional:
```
--tex
```
specifies the tex file name (by default, the first
```
*_review.tex
```
in the directory is selected; if multiple candidates exist in the directory, the script will issue a warning and prompt to use
```
--tex
```
to specify explicitly).

Output

```
{work_dir}/.check-review-alignment/ai_alignment_report.md
```
: A single report including Summary / Critical Fixes (P0) / Warnings (P1) / Rendering Result (PDF/Word path or error summary is attached at the end).
```
{work_dir}/.check-review-alignment/ai_alignment_input.json
```
: Structured input generated by the script, which includes "citation + literature meta-information (including DOI/URL) + PDF abstract segment", facilitating quick verification by the host AI.
Optimized
```
{topic}_review.tex
```
(saved in the root directory of
```
work_dir
```
); newly generated
```
{topic}_review.pdf
```
and
```
{topic}_review.docx
```
(saved in the root directory of
```
work_dir
```
).

Workflow

Step 0: Dependency Check (Mandatory Only for Rendering Path)

This skill supports two types of deterministic actions:

```
--prepare
```
: Generate structured input (does not depend on rendering skill)
```
--render
```
: Render PDF/Word (mandatory dependency on
```
systematic-literature-review
```
)

Only when you need to perform rendering (Step 5 or
--render
), you must check the following dependencies:

Is the

systematic-literature-review

skill available?

If not available: Stop immediately and prompt:

❌ Missing dependency: check-review-alignment relies on the systematic-literature-review skill for PDF/Word rendering. Please install the systematic-literature-review skill first.

If available: Proceed

Step 1: Pre-check and File Location

Locate
```
*_review.tex
```
and the corresponding
```
.bib
```
in
```
work_dir
```
(or specify with
```
--tex
```
).
If any file is missing: Stop and explain the missing files.

Step 2: Extract Structured Context (Recommended)

Run the script to generate

ai_alignment_input.json

for the host AI to verify each item quickly and traceably:

bash

# Enter the skill root directory (usually ~/.codex/skills/check-review-alignment or ~/.claude/skills/check-review-alignment after installation)
cd /path/to/check-review-alignment
python3 scripts/run_ai_alignment.py --work-dir "/path/to/work_dir" --prepare

Step 3: AI Semantic Verification and Priority Classification

Verify each citation (bibkey):

Identify Error Type: Determine if it belongs to P0 (Fatal)/P1 (Warning)/P2 (Forbidden)
Evidence Source: Use PDF abstract segment > BibTeX abstract/title > inference only from sentences (lowest priority) in order of priority
Process by Priority:
- P0: Rewrite the "sentence containing the citation" with minimal changes to restore consistency (only adjacent sentences in the same paragraph that also contain the citation can be rewritten if necessary; whole paragraph rewriting is prohibited)
- P1: Only record to Warnings, no rewriting
- P2: Skip, no modification
Mandatory Constraints:
- Keep LaTeX command integrity (such as
```
\cite{}
```
  ,
```
\ref{}
```
  ,
```
\label{}
```
  )
- Do not introduce new bibkeys (unless fixing wrong bibkeys)
- Do not forge paper content
- Prohibit rewriting sentences that do not contain citations
- Keep the original content when uncertain ("Do Not Modify When Unsure" principle)

Step 4: Generate Report

Generate a report in

{work_dir}/.check-review-alignment/ai_alignment_report.md

, which must include the following sections:

Summary: Statistics such as number of paragraphs, number of citations, number of P0 modifications, number of P1 warnings, number of P2 skips, etc.
Detailed Information (Mandatory): Detailed verification records for each citation, including:
- Citation: Literature title
```
{title}
```
  ; DOI number
```
{DOI}
```
  (mark "No DOI" if not available)
- Original Content: Original sentence containing the citation
- Actual Literature Content: Actual content of the paper (extracted from PDF abstract segment / BibTeX abstract / title)
- Citation Rationality Evaluation: Semantic consistency analysis, explaining whether the original description accurately reflects the literature content
- Issue Level: P0 (Fatal)/P1 (Warning)/P2 (Forbidden)/No Issue
Critical Fixes (P0): Fatal errors that must be fixed (original sentence/cause/new sentence/line number/priority)
Warnings (P1): Issues that only require warnings (original sentence/cause/suggestion/line number/priority)
Rendering Result: PDF/Word path or error summary

Example Format for Detailed Information Section:

markdown

## Detailed Information

### Citation: [bibkey]

| Field | Content |
|------|------|
| Literature Title | {title} |
| DOI | {DOI or "No DOI"} |
| Original Content | {Original sentence containing the citation} |
| Actual Literature Content | {Actual paper content extracted from PDF/BibTeX} |
| Citation Rationality Evaluation | {Semantic consistency analysis} |
| Issue Level | P0 / P1 / P2 / No Issue |

Step 5: Render PDF/Word

Render after modifying the tex file:

bash

cd /path/to/check-review-alignment
python3 scripts/run_ai_alignment.py --work-dir "/path/to/work_dir" --render

Modification Boundaries and Priorities

Must Fix (P0)

The following situations must be modified (with minimal changes):

Fake Citation (
```
fake_citation
```
)
- Performance: The cited literature does not exist at all (missing in .bib or wrong bibkey)
- Handling: Mark
```
missing_in_bib: true
```
  in the report and explain, suggest the user check the bibkey
Wrong Citation (
```
wrong_citation
```
)
- Performance: The bibkey in the citation command does not match the context (mixing up references)
- Handling: Rewrite the sentence to match the literature content of the bibkey, or replace with the correct bibkey
Contradictory Citation (
```
contradictory_citation
```
)
- Performance: The main body description contradicts the paper content (e.g., the paper says male, but the text writes female)
- Handling: Rewrite the sentence to match the paper content

Only Warn, No Modification (P1)

The following situations are only recorded in the report's Warnings, no rewriting:

Weak Support (
```
weak_support
```
): Suggest the user add stronger citations
Positioning Deviation (
```
overclaim
```
): Suggest the user reduce the intensity of the statement (e.g., avoid "first-ever/initial/core")

Prohibit Modification (P2)

The following situations are completely prohibited from being modified:

Style Optimization (
```
style_issue
```
): Word order/wording/polishing
Sentences Without Citations: Sentences that do not contain any
```
\cite{}
```
Paragraph Rewriting: Even if all citations in the paragraph are fixed, the whole paragraph cannot be rewritten

Modification Principles

Minimal Change Principle: Only rewrite necessary sentences, do not rewrite adjacent irrelevant sentences or the whole paragraph
Do Not Modify When Unsure Principle: Keep the original content and record to Warnings when uncertain whether it is a fatal error
Evidence Priority Principle: Judgment basis priority: PDF abstract segment > BibTeX abstract/title > inference only from sentences

Configuration (See

check-review-alignment/config.yaml

)

```
citation_commands
```
: Recognized citation commands
```
pdf.*
```
: Whether to extract PDF text and page limit
```
render.*
```
: Dependent skill name and override strategy
```
ai.input_limits.*
```
: Text truncation limit for structured input (
```
ai_alignment_input.json
```
)
```
ai.modification.*
```
: Modification strategy (executed by the host AI; the script only packages it into
```
ai_alignment_input.json
```
)
- ```
ai.modification.error_priority
```
  : P0/P1/P2 classification (only modify P0; P1 only warns; P2 skips)
- ```
ai.modification.non_fatal_handling
```
  : Handling strategy for P1/P2
```
ai.paragraph_optimization.*
```
: Paragraph optimization strategy (executed by the host AI; disabled by default to avoid style rewriting)

Quick Start

bash

cd /path/to/check-review-alignment
# 1) Generate structured input (Recommended)
python3 scripts/run_ai_alignment.py --work-dir "/path/to/work_dir" --prepare
# 2) The host AI completes verification and tex rewriting according to this SKILL.md workflow, and writes ai_alignment_report.md
# 3) Render PDF/Word
python3 scripts/run_ai_alignment.py --work-dir "/path/to/work_dir" --render

Dependencies

Python packages:
```
PyYAML
```
,
```
bibtexparser
```
(reads config/bib),
```
pdfplumber
```
or
```
PyPDF2
```
(PDF extraction). It will degrade when missing (will not interrupt).
No additional dependencies for AI functions: AI is provided by the Agent Skills host environment (Claude/Codex); the script does not directly call the LLM API locally.
Reference directory: The current
```
check-review-alignment/references/
```
is empty (if more detailed strategies/templates are introduced in the future, they will be added to this directory).

Verification Checklist (Static Self-check)

```
{work_dir}/.check-review-alignment/ai_alignment_report.md
```
has been generated, and each modification includes: original sentence/cause/new sentence/location information.
```
{work_dir}/.check-review-alignment/ai_alignment_input.json
```
has been generated, containing structured citation and literature information.
The modified tex can be compiled without errors, and citation commands (such as
```
\cite{}
```
) are completely retained.
The path and status of the rendered PDF/Word have been written to the report.

Security Principles

This skill follows the following principles to ensure "Do not modify for the sake of modifying":

Principle	Description	Implementation
Deterministic Script Boundary	The script only performs parsing/extraction/rendering, no semantic judgment	The script does not call LLM API
Priority Classification	P0 Fix / P1 Warning / P2 Skip	`ai.modification.error_priority`
Do Not Modify When Unsure	Keep original content when uncertain	Workflow mandatory constraints + Warnings
Minimal Change	Only rewrite necessary sentences, no whole paragraph rewriting	`max_edits_per_sentence` + Prohibit paragraph rewriting
LaTeX Integrity	Retain all LaTeX command structures	`preserve_citations: true`

check-review-alignment

NPX Install

Tags

SKILL.md Content (Chinese)

Check Review Alignment

Applicable Scenarios

Input

Output

Workflow

Step 0: Dependency Check (Mandatory Only for Rendering Path)

Step 1: Pre-check and File Location

Step 2: Extract Structured Context (Recommended)

Step 3: AI Semantic Verification and Priority Classification

Step 4: Generate Report

Step 5: Render PDF/Word

Modification Boundaries and Priorities

Must Fix (P0)

Only Warn, No Modification (P1)

Prohibit Modification (P2)

Modification Principles

Configuration (See
`check-review-alignment/config.yaml`
)

Quick Start

Dependencies

Verification Checklist (Static Self-check)

Security Principles

check-review-alignment

NPX Install

Tags

SKILL.md Content (Chinese)

Check Review Alignment

Applicable Scenarios

Input

Output

Workflow

Step 0: Dependency Check (Mandatory Only for Rendering Path)

Step 1: Pre-check and File Location

Step 2: Extract Structured Context (Recommended)

Step 3: AI Semantic Verification and Priority Classification

Step 4: Generate Report

Step 5: Render PDF/Word

Modification Boundaries and Priorities

Must Fix (P0)

Only Warn, No Modification (P1)

Prohibit Modification (P2)

Modification Principles

Configuration (See check-review-alignment/config.yaml)

Quick Start

Dependencies

Verification Checklist (Static Self-check)

Security Principles

Configuration (See
`check-review-alignment/config.yaml`
)