airc-academic-revision

name	airc-academic-revision
description	AIRC workflow for revising markdown or txt academic writing while preserving headings, formulas, terminology, citations, numbers, captions, placeholders, English spacing, paths, and Markdown structure. Use when Codex needs to scan a paper, classify blocks by rewrite policy, revise safe narration blocks, review integrity, and pass a write gate before output.

AIRC

Artificial Intelligence Rewrite Content

AIRC is a structure-preserving academic rewrite skill. It helps an agent revise academic Markdown or TXT documents while keeping protected academic content stable.

• Product name: AIRC
• CLI: airc
• Skill ID: airc-academic-revision
• Python import: airc_skill

What This Skill Does

AIRC scans a document, assigns block policies, performs controlled developmental rewrite on editable body prose, reviews preservation and style metrics, and writes output only when the write gate allows it.

When To Use

Use AIRC when the user asks to revise an academic Markdown or TXT document and needs:

• protected headings, formulas, citations, numbers, paths, captions, image Markdown, checkpoints, and technical terms
• academic body-prose rewriting rather than simple synonym replacement
• report-backed review before writing
• optional style preset control
• optional target-style alignment against a reference document

When Not To Use

Do not use AIRC to:

• generate new facts, citations, datasets, experiments, metrics, or external background
• change formulas, captions, tables, references, code blocks, paths, checkpoints, image Markdown, or protected technical terms
• rewrite non-academic text freely
• imitate a target-style file as content or copy target wording
• bypass review or write gate checks

Workflow

1. guide
2. run
3. review report
4. write output

Command flow:

airc guide paper.md --as-agent-context
airc run paper.md --preset academic_natural

Target-style flow:

airc run paper.md --target-style-file target-style.md

The target file is a style reference only. It is not an evidence source. Never copy target text directly.

Block Policies

AIRC classifies blocks before rewriting:

• do_not_touch: preserve verbatim or compare only protected spans.
• high_risk: keep original unless a very narrow cleanup is safe.
• light_edit: small sentence-level cleanup.
• rewritable: body prose eligible for deeper sentence-cluster and discourse-level revision.

Agent instructions may include required_sentence_actions, required_cluster_actions, clarity_constraints, preserve_items, and forbidden patterns.

Preservation Rules

Never change:

• titles, headings, and heading numbering
• formulas, variables, and formula numbering
• citations, citation order, years, numeric values, thresholds, metrics, sample sizes, and model names
• technical terms, dataset names, algorithm names, mechanism names, and system names
• captions, figures, image Markdown, links, placeholders, tables, references, and code blocks
• paths, checkpoints, filenames, and version strings
• Markdown structure and symbols

If a rewrite risks any protected content, keep the source block and report the reason.

Target-Style Alignment Rule

When --target-style-file is provided:

• use the target file only to fit language-distribution signals
• preserve the source document's facts, terms, numbers, formulas, citations, paths, figures, and Markdown
• never import target claims or external evidence
• never copy target sentences directly
• Do not copy target wording or target sentence structure as content.
• reject or warn if terminology_drift, evidence_drift, grammar risk, or protected-content drift appears

Target-style metrics include:

• style_distribution_match_ratio
• class_aware_style_match_ratio
• terminology_drift
• evidence_drift
• paragraph-class breakdown

Output Contract

AIRC writes:

• revised .airc.md, or the requested output path
• .report.json, or the requested report path

The final response should summarize:

• output path
• report path
• write_gate.decision
• core warnings
• body_rewrite_coverage
• target-style metrics when enabled
• any blocker if the write gate rejects

The candidate should be treated as usable only when the write gate decision is pass or pass_with_minor_risk.

Presets

• academic_natural: default controlled academic revision.
• zh_academic_l2_mild: mildly non-native Chinese academic prose. It may be slightly more explanatory and less compressed, but must remain grammatical, academic, factual, and protected-content safe.

Other local presets may exist for specialized workflows, but public use should start with academic_natural.

Agent Prompt Template

Use the airc-academic-revision skill to revise my Markdown/TXT academic document.
Run guide first, then run the revision pipeline, then review the report.
Preserve headings, formulas, citations, numbers, paths, figure references, image markdown, checkpoints, and technical terms.
If I provide a target-style file, use it only for language-distribution alignment, not as a source of facts.
Write the revised file next to the source file and provide the report path.