Academic Writing

Industrial AI Research

Run a lean, source-aware research workflow for Industrial AI.

Capability Summary

• Structured literature research for Industrial AI and automation topics
• Mandatory four-question intake before any search or synthesis
• Venue-aware source prioritization (arXiv, IEEE, automation venues)
• Four deliverable modes: research-brief, literature-map, venue-ranked survey, research-gap memo
• Contrarian synthesis pass to surface contradictions and under-explored gaps
• Survey draft generation: outline-first writing with per-section evidence packs and optional LaTeX export

Triggering

Use this skill when the user wants to:

• Survey Industrial AI literature on a specific subtopic
• Compare papers across venues or methods within Industrial AI
• Identify research gaps in predictive maintenance, scheduling, anomaly detection, or smart manufacturing
• Produce a structured research report with source-backed evidence
• Draft a structured survey on an Industrial AI subtopic
• Produce a survey manuscript with taxonomy, evidence packs, and section-by-section writing

Do Not Use

• Writing or compiling LaTeX/Typst papers (use latex-paper-en, latex-thesis-zh, or typst-paper).

Note: survey-draft mode produces Markdown by default; for LaTeX output, it delegates final formatting to latex-paper-en.

• Auditing paper quality or formatting (use paper-audit)
• Systematic reviews or meta-analyses requiring IRB or clinical ethics
• Topics outside the Industrial AI and automation domain
• Auditing an existing paper's quality or formatting (use paper-audit)
• Editing LaTeX/Typst source files (use the appropriate writing skill)

Safety Boundaries

• Never fabricate paper metadata (title, authors, venue, year, DOI)
• Never present preprints as peer-reviewed publications
• Never start synthesis before intake questions are answered
• Never suppress contradictions or conflicting evidence
• Never use Tier 4 sources (blogs, press releases) as primary evidence

Core Rules

1. Ask the user the four intake questions (see references/question-flow.md) before starting any search or synthesis. 2. Keep the skill workflow in English only, even when the requested report language is not English. 3. Prefer recent arXiv plus top IEEE and automation venues over generic web articles. 4. Default to the last 3 years, but keep seminal older work when it is still necessary for context. 5. Cite every substantive claim and separate verified evidence from inference. 6. In survey-draft mode, complete all structure and evidence phases before generating any prose. Structure phases produce YAML/tables only.

Intake Contract

Always start by asking the four intake questions defined in references/question-flow.md: 1. Report language (English / Simplified Chinese / Bilingual summary) 2. Deliverable mode (research-brief / literature-map / venue-ranked survey / research-gap memo / survey-draft) 3. Time window (last 12 months / last 3 years / last 5 years / custom) 4. Industrial AI emphasis (predictive maintenance / intelligent scheduling / industrial anomaly detection / smart manufacturing and process optimization / CPS and edge AI / robotics crossover)

If the user does not choose, default to last 3 years and the subdomain implied by their prompt.

Required Inputs

• A concrete Industrial AI topic or question.
• User choices for report language, deliverable mode, time window, and domain emphasis.
• Optional preferences on peer-reviewed-only filtering, benchmarks vs deployment evidence, or desired output format.

If any intake item is missing, ask the mandatory questions from references/question-flow.md before you search.

Source Strategy

Read these files before searching:

• references/source-priority.md
• references/venue-map.md

Primary sources:

• arXiv: eess.SY, cs.AI
• IEEE and automation anchors: T-ASE, CASE

Supporting crossover sources:

• arXiv: cs.RO, cs.LG
• IEEE robotics venues: ICRA, IROS, RA-L, T-RO
• Adjacent industrial and control venues listed in references/venue-map.md

When the user asks for the latest work, prefer: 1. arXiv recent streams for rapid updates 2. top IEEE and automation venues for stronger publication filtering 3. secondary crossover venues only when they materially improve coverage

Workflow

Phase 1. Scope

• Rewrite the request as a precise Industrial AI research objective.
• Lock the report language, deliverable mode, time window, and domain emphasis.
• State explicit in-scope and out-of-scope boundaries.

Phase 2. Search Plan

• Build venue buckets and keyword groups from references/source-priority.md.
• Separate primary sources from secondary crossover sources.
• State the recency policy and any seminal-paper exceptions.

Phase 3. Source Collection

• Gather papers from the prioritized source buckets.
• Prefer official venue pages, arXiv recent listings, IEEE Xplore landing pages, and publisher or conference pages.
• Record why each paper was included.

Phase 4. Verification and Triage

• Check venue quality, publication type, year, and relevance.
• Remove weak matches, duplicates, and generic blog-style sources.
• Mark unreviewed preprints as preprints.

Phase 5. Synthesis

• Cluster the shortlisted papers by problem, method, dataset, deployment setting, and evaluation style.
• Surface trends, gaps, contradictions, and under-explored opportunities.
• Run a contrarian pass: what would challenge the dominant conclusion?

Phase 6. Report Assembly

Use the stable report structure from references/report-modes.md.

Every final report must include:

• search scope
• source buckets by venue
• shortlisted papers
• synthesis of trends and gaps
• recommended next reading or next experiments

Survey-Draft Workflow (Phases S1–S4)

When the user selects survey-draft, Phases 1–4 (Scope, Search Plan, Source Collection, Verification) execute as normal, then S1–S4 replace the original Phases 5–6.

Phase S1. Outline Building

Read references/modules/SURVEY_OUTLINE.md.

• Extract a taxonomy from the verified literature.
• Build the section skeleton as structured YAML.
• Present the outline to the user for approval.
• CHECKPOINT: do not enter S2 until the user approves the outline.

Phase S2. Evidence Pack Assembly

Read references/modules/SURVEY_EVIDENCE.md.

• Assemble an evidence pack for every H3 subsection.
• Lock the citation scope for each subsection.
• Produce structured evidence bundles (no prose).

Phase S3. Section-by-Section Writing

Read references/modules/SURVEY_WRITER.md.

• Draft each H3 independently, grounded in its evidence pack.
• Run the self-check gate on every H3 (depth, citation scope, tone).
• Produce one Markdown file per H2 section.

Phase S4. Merge and Quality Gate

Read references/modules/SURVEY_MERGE.md.

• Merge all section drafts into a single document.
• Run cross-section consistency checks.
• Apply the final quality checklist.
• If the user requested LaTeX output, delegate to latex-paper-en.

Deliverable Modes

Read references/report-modes.md and follow the selected mode exactly.

• research-brief: short, decision-ready overview
• literature-map: thematic map across methods and subproblems
• venue-ranked survey: grouped by source quality and venue tier
• research-gap memo: open problems, design space, and next-step opportunities
• survey-draft: taxonomy-driven survey manuscript with outline-first writing and optional LaTeX export

Output Contract

• State the locked intake choices and any defaults you applied before synthesis.
• Distinguish verified evidence from inference in every deliverable.
• Label preprints explicitly as preprints.
• For non-survey modes, produce a structured report that includes: scope, source buckets, shortlisted papers, synthesis, and next reading or next experiments.
• For survey-draft, keep stage outputs format-specific:
• S1: YAML outline only
• S2: evidence packs or tables only
• S3: section Markdown drafts grounded in the evidence packs
• S4: merged Markdown survey with cross-section consistency notes
• If sources are sparse, inaccessible, or off-scope, say so directly and report the exact fallback you used.

Module Router

Quality Bar

Read references/quality-checklist.md before finalizing.

Non-negotiable standards:

• no unsupported claims
• no venue-blind source mixing
• no hiding contradictions
• no synthesized report before intake questions are answered
• no generic "latest research says" language without source-backed evidence

Error Handling

• Zero results: Broaden keywords, relax the time window by one tier, and try adjacent venues. If still empty, report the negative result with the exact queries attempted.
• Off-subdomain topic: State that the topic falls outside Industrial AI scope, suggest the closest supported subdomain, and ask the user whether to proceed or abort.
• Inaccessible databases: Note which sources were unreachable, proceed with available sources, and flag the gap in the final report.
• Too few papers (<5 shortlisted): Lower the time window threshold, include Tier 2/3 venues, and explicitly note the thin evidence base in the synthesis.

Reference Map

Examples

• examples/predictive-maintenance.md
• examples/intelligent-scheduling.md
• examples/industrial-anomaly-detection.md
• examples/survey-predictive-maintenance.md

Example Requests

• “Research recent predictive maintenance papers from the last 3 years and return a research-brief.”
• “Compare industrial anomaly detection papers across arXiv and IEEE automation venues, and show contradictions in evaluation setups.”
• “Draft a survey on intelligent scheduling for researchers new to the subfield, but stop after the YAML outline for approval.”
• “My topic is warehouse picking robotics. If that is outside scope, tell me the closest supported Industrial AI framing and proceed only with that.”

Boundaries

This v1 skill does not implement:

• systematic review mode
• meta-analysis
• IRB-heavy or clinical ethics branches
• standalone automation scripts

If the user needs those, state the boundary and continue with the closest supported research mode.

Example: Industrial Anomaly Detection

User Prompt

Research recent industrial anomaly detection literature and summarize the gaps.

Expected Intake

• Ask for language, deliverable mode, time window, and Industrial AI emphasis.

Expected Search Bias

• Primary: T-II, T-ASE, cs.AI, eess.SY
• Secondary: cs.LG, T-IE

Good Output Shape

• Group papers by modality, supervision level, and industrial context
• Distinguish benchmark-only work from production-facing work
• End with open gaps in labels, drift handling, and deployment evidence

Example: Intelligent Scheduling

User Prompt

Compare latest scheduling RL papers from arXiv and IEEE for flexible job shops.

Expected Intake

• Ask for language, deliverable mode, time window, and domain emphasis.

Expected Search Bias

• Primary: CASE, T-ASE, eess.SY, cs.AI
• Secondary: cs.LG, SMC Systems

Good Output Shape

• Separate operations realism from algorithm novelty
• Make constraints and benchmark assumptions visible
• Flag when RL papers have weak industrial deployment evidence

Example: Predictive Maintenance

User Prompt

Research recent predictive maintenance papers for rotating machinery.

Expected Intake

• Ask for report language first.
• Ask for deliverable mode, time window, and Industrial AI emphasis.

Expected Search Bias

• Primary: T-ASE, T-II, eess.SY, cs.AI
• Secondary: Automatica, T-IE

Good Output Shape

• Identify sensor modality and maintenance target
• Separate peer-reviewed papers from preprints
• Highlight deployment realism, data scarcity, and generalization gaps

Example: Survey Draft — Predictive Maintenance with Deep Learning

This example demonstrates the complete survey-draft workflow from intake to final merge.

User Prompt

I need a survey on deep learning methods for predictive maintenance in manufacturing.

Focus on vibration-based and multi-modal approaches from the last 3 years.

Intake Answers

Phase S1: Outline

Expected taxonomy.md (excerpt)

# Taxonomy: Deep Learning for Predictive Maintenance

## Classification Axes

- Axis 1: Signal modality — the type of sensor data used as primary input
- Axis 2: Method family — the core deep learning architecture

## Paper-to-Cell Mapping

| Paper ID | First Author | Year | Signal Modality | Method Family | Primary Cell |
|----------|-------------|------|-----------------|---------------|-------------|
| pdm-01 | Zhang | 2024 | Vibration | CNN | Vibration × CNN |
| pdm-02 | Li | 2024 | Vibration | Transformer | Vibration × Transformer |
| pdm-03 | Wang | 2023 | Multi-modal | Hybrid | Multi-modal × Hybrid |
| pdm-04 | Kim | 2024 | Acoustic | CNN | Acoustic × CNN |
| pdm-05 | Chen | 2023 | Current | Physics-informed | Current × Physics-informed |

Expected outline.yml (excerpt)

title: "Survey: Deep Learning for Predictive Maintenance in Manufacturing"
audience: researchers_new
length_tier: standard
taxonomy:
  axis_1: "Signal modality"
  axis_2: "Method family"
sections:
  - id: S1
    title: "Introduction"
    type: front-matter
    guidance: "Motivation for DL-based PdM, limitations of traditional approaches, contribution of this survey"
  - id: S2
    title: "Background and Scope"
    type: front-matter
    guidance: "PdM problem formulation, DL basics for time-series, search methodology"
  - id: S3
    title: "Vibration-Based Methods"
    type: body
    subsections:
      - id: S3.1
        title: "CNN Architectures for Vibration Signals"
        paper_count: 8
        key_papers: ["pdm-01", "pdm-06", "pdm-11"]
      - id: S3.2
        title: "Transformer and Attention Models"
        paper_count: 6
        key_papers: ["pdm-02", "pdm-07"]
      - id: S3.3
        title: "Physics-Informed Hybrid Models"
        paper_count: 4
        key_papers: ["pdm-08", "pdm-12"]
  - id: S4
    title: "Acoustic and Current-Based Methods"
    type: body
    subsections:
      - id: S4.1
        title: "Acoustic Emission Analysis"
        paper_count: 5
        key_papers: ["pdm-04", "pdm-13"]
      - id: S4.2
        title: "Motor Current Signature Analysis"
        paper_count: 4
        key_papers: ["pdm-05", "pdm-14"]
  - id: S5
    title: "Multi-Modal Fusion Approaches"
    type: body
    subsections:
      - id: S5.1
        title: "Early and Late Fusion Strategies"
        paper_count: 5
        key_papers: ["pdm-03", "pdm-15"]
      - id: S5.2
        title: "Cross-Modal Attention Mechanisms"
        paper_count: 3
        key_papers: ["pdm-16", "pdm-17"]
  - id: S6
    title: "Comparative Analysis"
    type: analysis
  - id: S7
    title: "Open Challenges and Future Directions"
    type: discussion
  - id: S8
    title: "Conclusion"
    type: closing

Phase S2: Evidence Pack (excerpt for S3.1)

# Evidence Pack: S3.1 — CNN Architectures for Vibration Signals

## Claim Candidates

- **1D-CNN outperforms traditional features on CWRU**: pdm-01 — "The proposed 1D-CNN achieved 99.2% accuracy on the CWRU bearing dataset, outperforming SVM (94.1%) and random forest (92.7%)."
- **Multi-scale CNN captures frequency patterns**: pdm-06 — "Multi-scale convolutional filters at 32, 64, and 128 kernel sizes capture both local transient events and global frequency patterns."
- **Transfer learning reduces labeled data need**: pdm-11 — "Pre-training on a source domain with abundant labels and fine-tuning on the target domain with only 50 labeled samples achieved 96.8% accuracy."
- **Lightweight CNN for edge deployment**: pdm-18 — "The pruned CNN model (0.3M parameters) runs at 15ms inference on Raspberry Pi 4, enabling real-time bearing monitoring."
- **Attention-augmented CNN**: pdm-19 — "Adding channel attention to ResNet-18 improved F1-score from 0.94 to 0.97 on the Paderborn bearing dataset."

## Comparison Table

| Paper | Method | Dataset | Key Metric | Result | Deployment Evidence |
|-------|--------|---------|------------|--------|---------------------|
| pdm-01 | 1D-CNN | CWRU | Accuracy | 99.2% | None |
| pdm-06 | Multi-scale CNN | CWRU + Paderborn | Accuracy | 99.5% | Simulation |
| pdm-11 | Transfer CNN | CWRU → Factory | Accuracy | 96.8% | Pilot |
| pdm-18 | Pruned CNN | CWRU | Accuracy / Latency | 98.1% / 15ms | Production (edge) |
| pdm-19 | Attention-CNN | Paderborn | F1-score | 0.97 | None |

## Anchor Facts

- 99.2% accuracy on CWRU with 1D-CNN: pdm-01
- 50 labeled samples sufficient for transfer: pdm-11
- 15ms inference on Raspberry Pi 4: pdm-18
- 0.3M parameters after pruning: pdm-18

## Gaps and Limitations

- **Cross-domain generalization**: pdm-01, pdm-06 — models validated only on CWRU; real factory data may differ significantly.
- **Label scarcity**: pdm-11 — transfer learning helps but requires a well-labeled source domain.

## Allowed Citations

- **Primary**: pdm-01, pdm-06, pdm-11, pdm-18, pdm-19
- **Chapter-level**: pdm-02, pdm-08, pdm-12
- **Global**: pdm-03

## Evidence Density Flag

- Status: SUFFICIENT

Phase S3: Draft (excerpt for S3.1)

## CNN Architectures for Vibration Signals

Convolutional neural networks have become the dominant architecture for vibration-based
predictive maintenance, largely due to their ability to extract hierarchical features
directly from raw or minimally processed time-series signals.

Zhang et al. [pdm-01] demonstrated that a straightforward 1D-CNN achieves 99.2% accuracy
on the CWRU bearing dataset, substantially outperforming traditional machine learning
baselines including SVM (94.1%) and random forest (92.7%). Building on this foundation,
Li et al. [pdm-06] introduced multi-scale convolutional filters at kernel sizes of 32, 64,
and 128, capturing both local transient events and global frequency patterns, which pushed
accuracy to 99.5% across both CWRU and Paderborn benchmarks.

A persistent challenge in industrial PdM is the scarcity of labeled fault data. Wang et al.
[pdm-11] addressed this through transfer learning, pre-training on a source domain with
abundant labels and fine-tuning on a target factory domain with only 50 labeled samples,
achieving 96.8% accuracy — a result that suggests practical deployment is feasible even
with minimal target-domain annotation.

| Paper | Method | Dataset | Key Metric | Result | Deployment |
|-------|--------|---------|------------|--------|------------|
| pdm-01 | 1D-CNN | CWRU | Accuracy | 99.2% | None |
| pdm-06 | Multi-scale CNN | CWRU + Paderborn | Accuracy | 99.5% | Simulation |
| pdm-11 | Transfer CNN | CWRU → Factory | Accuracy | 96.8% | Pilot |
| pdm-18 | Pruned CNN | CWRU | Acc / Latency | 98.1% / 15ms | Production |
| pdm-19 | Attention-CNN | Paderborn | F1-score | 0.97 | None |

For edge deployment, Kim et al. [pdm-18] demonstrated that a pruned CNN with only 0.3M
parameters achieves 98.1% accuracy while running at 15ms inference on a Raspberry Pi 4,
enabling real-time bearing monitoring in resource-constrained environments. Park et al.
[pdm-19] further showed that augmenting ResNet-18 with channel attention improves F1-score
from 0.94 to 0.97 on the Paderborn dataset, suggesting that attention mechanisms offer a
low-cost accuracy boost without significant computational overhead.

Despite these advances, most CNN-based PdM studies rely heavily on the CWRU benchmark
[pdm-01, pdm-06], raising concerns about cross-domain generalization to real factory
conditions where noise profiles, operating regimes, and fault modes differ substantially.

Phase S4: Final Output Structure

output-dir/
├── outline.yml
├── taxonomy.md
├── evidence/
│   ├── s3-1-cnn-vibration.md
│   ├── s3-2-transformer-vibration.md
│   ├── s3-3-physics-informed.md
│   ├── s4-1-acoustic.md
│   ├── s4-2-current.md
│   ├── s5-1-fusion-strategies.md
│   └── s5-2-cross-modal-attention.md
├── citation-map.md
├── drafts/
│   ├── s1-introduction.md
│   ├── s2-background.md
│   ├── s3-vibration-methods.md
│   ├── s4-acoustic-current.md
│   ├── s5-multi-modal.md
│   ├── s6-comparative-analysis.md
│   ├── s7-challenges.md
│   └── s8-conclusion.md
├── survey-draft.md
└── quality-report.md

Module: Survey Evidence Pack Assembly (Phase S2)

Assemble structured evidence for every H3 subsection. This phase produces data artifacts only — no prose.

Prerequisites

• Phase S1 complete: outline.yml approved by user, taxonomy.md finalized.
• Verified paper set available with full metadata.

Evidence Pack Structure

For each H3 subsection in outline.yml, produce one evidence file at {output-dir}/evidence/{section-slug}.md:

# Evidence Pack: [Section ID] — [Sub-branch Title]

## Claim Candidates

- **[Claim 1]**: [Paper ID] — "[verbatim or close-paraphrase snippet from abstract/conclusion]"
- **[Claim 2]**: [Paper ID] — "[snippet]"
- ...

## Comparison Table

| Paper | Method | Dataset | Key Metric | Result | Deployment Evidence |
|-------|--------|---------|------------|--------|---------------------|
| ... | ... | ... | ... | ... | None / Simulation / Pilot / Production |

## Anchor Facts

Quantitative facts that the writer must use:

- [Fact 1]: [Paper ID] — [exact number and context]
- [Fact 2]: [Paper ID] — [exact number and context]
- ...

## Gaps and Limitations

- **[Gap 1]**: evidence from [Paper IDs] — [brief description]
- **[Limitation 1]**: [Paper IDs] — [brief description]
- ...

## Allowed Citations

- **Primary** (this H3): [list of paper IDs]
- **Chapter-level** (parent H2): [list of paper IDs]
- **Global** (use sparingly): [list of paper IDs]

## Evidence Density Flag

- Status: SUFFICIENT | THIN_EVIDENCE
- If THIN_EVIDENCE: [explanation of what is missing]

Key Rules

No Prose

Evidence packs contain only structured data: bullet lists, tables, and tagged citations. The writer module (Phase S3) converts these into narrative text.

No Fabrication

• Every claim candidate must trace to a specific snippet from a verified paper.
• Every anchor fact must include the exact quantitative value from the source.
• If a claim cannot be sourced, do not include it.

Citation Scope Locking

The writer (Phase S3) is restricted to citations listed in the evidence pack:

• Primary citations: directly relevant to this H3. Use freely.
• Chapter-level citations: relevant to the parent H2 but not specific to this H3. Use for context.
• Global citations: foundational or cross-cutting papers. Use sparingly (max 2 per H3).

Thin Evidence Handling

If a subsection has fewer claims or citations than the minimum required by the length tier, mark it as THIN_EVIDENCE and:

• Do NOT pad with filler claims or tangential papers.
• Flag explicitly so the writer can adjust depth expectations.
• Consider suggesting to the user that this subsection be merged with a sibling.

Evidence Density Requirements

Minimum thresholds per H3 subsection, by length tier:

Citation Map

After all evidence packs are assembled, produce {output-dir}/citation-map.md:

# Citation Map

## Section-to-Citation Binding

| Section ID | Section Title | Primary Citations | Chapter Citations | Global Citations |
|-----------|--------------|-------------------|-------------------|------------------|
| S3.1 | [title] | [IDs] | [IDs] | [IDs] |
| S3.2 | [title] | [IDs] | [IDs] | [IDs] |
| ... | ... | ... | ... | ... |

## Citation Frequency

| Paper ID | First Author | Year | Sections Referenced | Role |
|----------|-------------|------|--------------------|----- |
| ... | ... | ... | [list] | Primary / Chapter / Global |

Artifacts Produced

Module: Survey Merge and Quality Gate (Phase S4)

Merge all section drafts into a single survey document, run quality checks, and optionally hand off to LaTeX.

Prerequisites

• Phase S3 complete: all drafts/{section-slug}.md files produced.
• outline.yml and citation-map.md available.

Step 1: Merge Sections

Assemble the final document in outline.yml section order:

1. Title and Abstract: generate a concise abstract (150–300 words) summarizing scope, method, key findings, and implications. 2. Front-matter sections: Introduction, Background and Scope (from Phase S3 drafts). 3. Body sections: each H2 with its H3 subsections (from Phase S3 drafts). 4. Insert transition sentences: 1–2 sentences between each H2 section connecting the logical flow from the previous branch to the next. 5. Comparative Analysis: build from the cross-cutting comparison table.

• Aggregate per-H3 comparison tables into a unified cross-cutting table.
• Add 2–4 paragraphs of cross-branch analysis highlighting patterns, trade-offs, and dominant approaches.

6. Open Challenges and Future Directions (from Phase S3 draft). 7. Conclusion (from Phase S3 draft). 8. References: collect all cited paper IDs from all evidence packs, deduplicate, and format as a numbered reference list.

Output: {output-dir}/survey-draft.md

Step 2: Quality Gate

Run every check below and record results in {output-dir}/quality-report.md:

Quality Checklist

Global Citation Density Requirements

Quality Report Format

# Quality Report: [Survey Title]

## Summary

- Length tier: [short / standard / comprehensive]
- Total word count: [N]
- Total unique citations: [N]
- Overall status: ALL PASS / HAS FAILURES / HAS WARNINGS

## Detailed Results

| # | Check | Criterion | Status | Details |
|---|-------|-----------|--------|---------|
| 1 | Outline alignment | ... | PASS | All 12 sections present |
| 2 | Citation health | ... | PASS | 67 unique, 0 duplicates |
| ... | ... | ... | ... | ... |

## Recommendations

- [Any WARN or FAIL items with suggested fixes]

Handling Failures

• FAIL on checks 1–6, 8–9: do not deliver the draft. Fix the issue and re-run the quality gate.
• WARN on check 7: deliver with a note to the user about section imbalance.
• After fixing, re-run the full checklist to confirm all items pass.

Step 3: LaTeX Handoff (Optional)

If the user requested LaTeX output during intake:

1. Confirm that survey-draft.md passes all quality gate checks. 2. Inform the user:

The Markdown draft is complete and has passed all quality checks. I will now delegate to the latex-paper-en skill for LaTeX formatting.

3. Provide survey-draft.md path to latex-paper-en for:

• Template selection (IEEE / ACM / Springer / NeurIPS / ICML or user-specified).
• BibTeX generation from the reference list.
• Compilation to PDF.

4. This skill does NOT directly create or edit .tex files. All LaTeX work is handled by latex-paper-en.

Output (if LaTeX requested): {output-dir}/survey-draft.tex (produced by latex-paper-en).

Artifacts Produced

Module: Survey Outline (Phase S1)

Build the survey skeleton before any prose is written. This phase produces structured YAML and a taxonomy table — no narrative text.

Prerequisites

• Phase 1–4 (Scope, Search Plan, Source Collection, Verification) are complete.
• A verified paper set is available.

Step 1: Extract Taxonomy

1. Read references/SURVEY_WRITING_GUIDE.md — load the taxonomy pattern library. 2. Identify which subdomain the verified papers belong to. 3. Select two classification axes from the pattern library (or propose custom axes if none fit). 4. Assign every verified paper to exactly one primary cell in the axis matrix. 5. Produce taxonomy.md:

# Taxonomy: [Topic]

## Classification Axes

- Axis 1: [name] — [definition]
- Axis 2: [name] — [definition]

## Paper-to-Cell Mapping

| Paper ID | First Author | Year | Axis 1 Value | Axis 2 Value | Primary Cell |
|----------|-------------|------|--------------|--------------|-------------|

Taxonomy Quality Checks

• Top-level branches (H2): 3–7. Fewer than 3 means the taxonomy is too coarse; more than 7 means it is too fragmented.
• Sub-branches per H2 (H3): 2–5.
• No cell should contain more than 40% of all papers. If it does, split the dominant axis.
• Every paper must map to exactly one primary sub-branch.

Step 2: Build Outline YAML

Produce outline.yml following this schema:

title: "Survey: [Topic]"
audience: researchers_new | practitioners | reviewers
length_tier: short | standard | comprehensive
taxonomy:
  axis_1: "[name]"
  axis_2: "[name]"
sections:
  - id: S1
    title: "Introduction"
    type: front-matter
    guidance: "Problem scope, motivation, contribution of this survey, reading guide"
  - id: S2
    title: "Background and Scope"
    type: front-matter
    guidance: "Key definitions, inclusion/exclusion criteria, search methodology summary"
  - id: S3
    title: "[Taxonomy Branch 1]"
    type: body
    subsections:
      - id: S3.1
        title: "[Sub-branch]"
        paper_count: N
        key_papers: ["paper_id_1", "paper_id_2"]
      - id: S3.2
        title: "[Sub-branch]"
        paper_count: N
        key_papers: ["paper_id_3"]
  # ... more body sections ...
  - id: SN-2
    title: "Comparative Analysis"
    type: analysis
    guidance: "Cross-cutting comparison tables, quantitative summaries across branches"
  - id: SN-1
    title: "Open Challenges and Future Directions"
    type: discussion
    guidance: "Gaps, emerging trends, recommended research directions"
  - id: SN
    title: "Conclusion"
    type: closing
    guidance: "Key takeaways, limitations of this survey, final recommendations"

Outline Validation Rules

• Every body section (type: body) must have at least 2 subsections.
• Every subsection must list at least 2 key_papers.
• The sum of all paper_count values must equal or exceed the total verified paper count (a paper may appear in multiple subsections as secondary, but has exactly one primary).
• Front-matter sections (Introduction, Background) have no subsections.

Step 3: Human Checkpoint

Present both outline.yml and taxonomy.md to the user with this exact prompt:

Please review the survey outline and taxonomy. You can:

- Approve to proceed to evidence pack assembly (Phase S2).

- Adjust branches — request merging, splitting, or renaming taxonomy branches.

- Reorder sections — change the sequence of body sections.

- Add/remove papers — adjust paper assignments.

>

I will not proceed to Phase S2 until you approve the outline.

Checkpoint Rules

• NEVER proceed to Phase S2 without explicit user approval.
• If the user requests changes, regenerate the affected parts of outline.yml and taxonomy.md, then re-present for approval.
• Track the approval status: outline_approved: true | false.

Artifacts Produced

Module: Survey Section-by-Section Writer (Phase S3)

Draft each H3 subsection as narrative prose, grounded entirely in the evidence packs from Phase S2.

Prerequisites

• Phase S1 complete: outline.yml approved.
• Phase S2 complete: all evidence packs and citation-map.md finalized.

Writing Procedure

Process sections in outline.yml order. For each H2 body section:

1. Write the H2 Lead Block

Before writing any H3, draft a 1–2 paragraph lead block for the H2 section:

• Preview the organizing principle of this taxonomy branch.
• State the comparison axes and key findings that will emerge.
• Reference the number of papers covered and the time span.

2. Write Each H3 Subsection

For each H3 under the current H2:

1. Load the evidence pack from {output-dir}/evidence/{section-slug}.md. 2. Start from claim candidates — do not write from a blank page. 3. Draft paragraphs following these rules:

• Every paragraph must contain at least one citation.
• Comparison sentences must use data from the comparison table.
• Quantitative claims must use exact values from anchor facts.
• Citation scope: prefer Primary → Chapter-level → Global (sparingly).

4. Include the comparison table from the evidence pack (may be reformatted for flow). 5. End with a synthesis paragraph that connects findings to the broader taxonomy branch.

3. Run Self-Check Gate

After completing each H3, verify against these thresholds:

If any check fails, revise the H3 before moving to the next one.

Forbidden Patterns

The following patterns must never appear in the draft:

Generator Tone

• "In this section, we will discuss..."
• "It is worth noting that..."
• "As mentioned earlier..."
• "The following subsection presents..."

Template Phrase Overuse

• "Taken together" — max 2 occurrences in the entire draft.
• "Notably" — max 3 occurrences in the entire draft.
• "Interestingly" — max 2 occurrences in the entire draft.

Citation Format Errors

• Adjacent citation blocks: [1] [2] is forbidden; merge to [1, 2].
• Orphan citations: a citation that appears in the text but not in the evidence pack's allowed list.

Unsupported Generalizations

• Any claim without a citation that makes a general statement about the field.
• Phrases like "it is well known that..." or "research has shown that..." without a specific reference.

THIN_EVIDENCE Handling

When an evidence pack is flagged as THIN_EVIDENCE:

• Reduce the expected paragraph count by 40%.
• Explicitly acknowledge the limited evidence: "The literature on [topic] remains sparse, with only [N] studies addressing..."
• Do NOT pad with tangential content or speculative claims.

Output Format

Produce one Markdown file per H2 section at {output-dir}/drafts/{section-slug}.md:

# [H2 Title]

[H2 lead block — 1-2 paragraphs]

## [H3.1 Title]

[Narrative paragraphs with inline citations]

[Comparison table]

[Synthesis paragraph]

## [H3.2 Title]

...

Front-Matter and Back-Matter Sections

Introduction

• Problem scope and motivation (why this survey, why now).
• Contribution statement (what this survey adds beyond existing surveys).
• Reading guide (brief description of each major section).

Background and Scope

• Key definitions relevant to the subdomain.
• Inclusion/exclusion criteria used during source collection.
• Brief search methodology summary (venues, time window, query terms).

Open Challenges and Future Directions

• Synthesize gaps from all evidence packs.
• Organize by: near-term (1–2 years), medium-term (3–5 years), long-term (5+ years).
• Each challenge must cite at least one paper that identifies or implies the gap.

Conclusion

• Summarize key findings (one sentence per H2 branch).
• State limitations of this survey.
• Final recommendation for practitioners and researchers.

Artifacts Produced

Quality Checklist

Run this checklist before final delivery.

Evidence Rules

• Every important claim has a source.
• Every paper summary states venue or source type.
• Preprints are labeled as preprints.
• Contradictory findings are surfaced, not hidden.
• Generic web summaries are excluded from evidence claims.

Industrial AI Relevance Rules

• The report stays anchored in automation and Industrial AI.
• Robotics papers appear only when they materially support the topic.
• General AI method papers are not treated as central evidence without industrial linkage.
• Source buckets reflect the requested domain emphasis.

Report Quality Rules

• The opening intake questions were asked first.
• The final report uses the stable section order from report-modes.md.
• The time window and any seminal-paper exceptions are explicit.
• The report ends with concrete next reading or next experiment suggestions.

Contrarian Pass

Before finalizing, ask:

• Which dominant claim would fail under real industrial constraints?
• Which shortlisted papers are strongest method papers but weakest deployment papers?
• Which part of the field is over-represented because it is easier to publish?
• Which missing evidence would most change the conclusion?

Question Flow

Use this exact order before searching.

Mandatory Opening Questions

1. Which report language should I use: English, Simplified Chinese, or bilingual summary? 2. Which deliverable do you want: research-brief, literature-map, venue-ranked survey, research-gap memo, or survey-draft? 3. Which time window should I prioritize: last 12 months, last 3 years, last 5 years, or a custom range? 4. Which Industrial AI emphasis is closest to your need: predictive maintenance, intelligent scheduling, industrial anomaly detection, smart manufacturing and process optimization, CPS and edge AI, or robotics crossover?

Follow-up Only If Needed

Ask at most two additional clarifying questions if the prompt is still too broad:

• Do you want peer-reviewed work only, or should I include recent preprints?
• Do you care more about methods, benchmarks, deployment evidence, or open research gaps?

Defaulting Rules

If the user leaves something unspecified:

• language -> English
• deliverable -> research-brief
• time window -> last 3 years
• emphasis -> infer from the user prompt

State the defaults you used before synthesis.

Survey-Draft Follow-up Questions

Ask these only when the user selects survey-draft:

1. What is the target audience: researchers new to the subfield, experienced practitioners, or reviewers evaluating the state of the art? 2. Do you want Markdown output only, or also a LaTeX draft via latex-paper-en? 3. Approximate target length: short survey (3000-5000 words), standard survey (5000-10000 words), or comprehensive survey (10000-15000 words)?

Survey-Draft Defaulting Rules

• audience → researchers new to the subfield
• output format → Markdown only
• target length → standard survey (5000-10000 words)

Report Modes

Use the selected mode exactly.

Stable Sections for Every Report

Every final report must contain these sections in this order:

1. Search Scope 2. Source Buckets by Venue 3. Shortlisted Papers 4. Synthesis of Trends and Gaps 5. Recommended Next Reading / Next Experiments

Mode: research-brief

Use when the user wants a fast answer.

• Target length: 700 to 1200 words
• Focus: direct findings, paper shortlist, quick recommendation
• Keep tables compact

Mode: literature-map

Use when the user wants a structured overview.

• Target length: 1200 to 2200 words
• Focus: themes, clusters, methods, datasets, evaluation patterns
• Include a theme-by-paper or method-by-paper map

Mode: venue-ranked survey

Use when the user wants stronger source discrimination.

• Target length: 1200 to 2500 words
• Group papers by venue tier and source bucket
• Make venue quality and publication type highly visible

Mode: research-gap memo

Use when the user wants opportunities and next steps.

• Target length: 900 to 1800 words
• Focus: unresolved gaps, weak evidence areas, and concrete future work ideas
• End with an ordered list of the most promising next experiments or reading tracks

Mode: survey-draft

Use when the user wants a structured survey manuscript draft.

• Target length: 3000–15000 words (user-selected: short / standard / comprehensive)
• Focus: taxonomy-driven survey with per-section evidence packs and structured argumentation
• Output: Markdown by default; optional LaTeX via latex-paper-en handoff
• Requires human checkpoint after outline approval

Survey-Draft Stable Sections

1. Title and Abstract 2. Introduction (problem scope, motivation, contribution of the survey, reading guide) 3. Background and Scope (key definitions, inclusion/exclusion criteria, search methodology) 4. Taxonomy (organizing framework — method-based, problem-based, or hybrid) 5. Body Sections (one H2 per taxonomy branch, H3 per sub-branch) 6. Comparative Analysis (cross-cutting comparison tables, quantitative summaries) 7. Open Challenges and Future Directions 8. Conclusion 9. References

Survey-Draft Artifact Contract

Source Priority

Use this file to rank sources and build searches.

Search Order

1. Recent arXiv streams for speed:

• eess.SY
• cs.AI
• cs.RO
• cs.LG

2. Top IEEE and automation venues:

• T-ASE
• CASE
• ICRA
• IROS
• RA-L
• T-RO

3. Adjacent industrial and control venues from venue-map.md

Ranking Logic

Use this ranking logic when filtering sources:

Recency Policy

• Default window: last 3 years.
• Use last 12 months when the user explicitly wants the latest wave.
• Use last 5 years when the field is sparse or adoption cycles are slow.
• Allow older seminal papers only when they define the problem or benchmark lineage.

Verification Rules

• Confirm year, venue, and publication type before citing a paper as evidence.
• Distinguish preprints from peer-reviewed papers.
• Remove duplicates between arXiv and conference or journal versions when possible.
• Prefer official pages and publisher metadata over scraped summaries.

Topic-Specific Heuristics

Predictive maintenance

• Heavily weight deployment setting, sensor modality, and maintenance outcome.
• Do not over-index on general anomaly detection papers with no industrial asset context.

Intelligent scheduling

• Weight papers that expose constraints, real system assumptions, or shop-floor context.
• Treat generic RL scheduling papers as secondary unless they show industrial realism.

Industrial anomaly detection

• Weight papers with fault type definitions, industrial datasets, or production constraints.
• Flag work that is image-only or benchmark-only with no industrial deployment link.

Smart manufacturing and CPS

• Weight papers that tie learning to control, optimization, or operations outcomes.
• Prefer papers with explicit industrial process assumptions over generic AI optimization papers.

Survey Writing Guide — Industrial AI

This guide defines the writing philosophy for survey drafts produced by the survey-draft deliverable mode. Read this file during all survey phases (S1–S4).

Why Industrial AI Surveys Are Different

Industrial AI surveys must evaluate work along dimensions that pure ML surveys can ignore:

• Deployment realism: algorithm performance alone is insufficient; real-world deployment evidence (pilot lines, factory trials, fleet rollouts) is a first-class evaluation axis.
• Data scarcity and labeling cost: most industrial datasets are small, imbalanced, or expensive to annotate. Surveys must surface how each method handles this.
• Latency spectrum: acceptable inference latency varies by orders of magnitude (PdM tolerates minutes; scheduling may need sub-second; real-time control needs milliseconds).
• Sim-to-real gap: many methods are validated only in simulation. The survey must explicitly flag sim-only vs. real-deployment evidence.
• Safety and regulatory context: industrial deployments often intersect with safety standards (IEC 61508, ISO 13849). Note when papers address or ignore these.

Taxonomy Pattern Library

Use these recommended classification axes when building the survey outline (Phase S1). Combine two axes into a matrix when the literature is dense enough.

Choosing Axes

1. Start with the two axes that produce the most even distribution of papers across cells. 2. If one axis produces a single dominant cell (>50% of papers), consider splitting that axis or switching to a different one. 3. Hybrid taxonomies (e.g., method-based at H2, application-based at H3) are acceptable when a single pair of axes cannot cover the literature.

Comparison Table Conventions

Comparison tables are mandatory artifacts in survey drafts. Follow these rules:

Per-H3 Tables (in evidence packs and body sections)

Every H3 subsection must include at least one comparison table with these columns:

Cross-Cutting Table (in Comparative Analysis section)

The Comparative Analysis section (Phase S4) must include at least one cross-cutting table that compares methods across taxonomy branches. Additional columns:

Writing Tone

• Analytical, not promotional. State what the evidence shows, not what "promises" a method holds.
• Use hedging language ("the results suggest", "under the reported conditions") when evidence is limited to a single study.
• Avoid generator-tone phrases: "In this section, we will discuss...", "It is worth noting that...", "Taken together..." (limit to ≤2 occurrences in the entire draft).
• Prefer active constructions: "Zhang et al. [12] propose..." over "A method was proposed by Zhang et al. [12]..."

Literature Review Quality Standards

Survey papers must exemplify the highest literature review quality. These rules (cross-referenced from the latex-paper-en skill's LOGIC module) are especially critical in survey context:

A1: Thematic Clustering (Mandatory)

Literature MUST be organized by research themes, methodology families, or application domains — never by chronological order or author enumeration. In survey context, the taxonomy axes (see Taxonomy Pattern Library above) define the thematic clusters.

Anti-pattern: "In 2018, Smith proposed X. In 2019, Jones introduced Y. In 2020, Wang designed Z." Correct pattern: Group papers by method family (e.g., CNN-based → Transformer-based → Hybrid), then discuss each group's strengths and limitations.

A2: Critical Analysis After Each Cluster (Mandatory)

Each H3 subsection (theme cluster) MUST end with a critical synthesis paragraph that:

• Summarizes shared strengths and limitations of methods in that cluster
• Identifies open problems within the cluster
• Provides a transition to the next cluster or section

A3: Research Gap Derivation (Mandatory)

The final subsection of the literature review (or the Comparative Analysis section) MUST explicitly identify:

• Gaps in the current literature that remain unaddressed
• Under-explored combinations of methods and applications
• Missing evaluation dimensions (e.g., deployment evidence, latency, data efficiency)

A4: Citation Density Funnel

Survey papers naturally follow a funnel pattern: broad introduction citations → focused per-cluster citations → specific technique deep-dives. Maintain this pattern within each H2 section.

Full reference: See ../latex-paper-en/references/modules/LOGIC.md for detailed detection heuristics and automated check patterns.

Length Tiers

Venue Map

Use this file to decide where to search first for Industrial AI topics.

Primary Anchors

Use these first unless the prompt clearly points elsewhere.

Secondary Crossover Venues

Use these when the prompt crosses into robotics, embodied decision-making, or learning-heavy methods.

Adjacent Secondary Venues

Use these only after the primary anchors are covered.

Subdomain Mapping

Weighting Rule

When multiple venues appear relevant:

1. Prefer Industrial AI and automation venues over general AI venues. 2. Prefer official venue or publisher pages over tertiary summaries. 3. Keep preprints when they are recent and relevant, but label them clearly as preprints. 4. Do not let robotics crossover venues dominate unless the user prompt is robotics-heavy.