rewrite-literature

name: rewrite-literature description: Orchestrate a deep rewrite of Chapter 2 (Literature Review) using multi-agent pipeline with RAG-backed evidence

Literature Review Rewrite Orchestrator

Rewrite Chapter 2 from scratch with deep extraction from ~20 cited papers. Uses a 5-phase pipeline with 7 specialized agents.

Arguments

• all — Run the full 5-phase pipeline (default)
• phase1 through phase5 — Run a specific phase (assumes prior phases completed)
• section N — Draft only section 2.N (for iterating on one section)

Workspace

All intermediate artifacts go in thesis/chapters/.litreview_workspace/. Create it at the start:

mkdir -p /Users/alessandro/Projects/Tesi/thesis/chapters/.litreview_workspace/drafts

Target Structure (~400-500 lines)

Chapter 2: Literature Review
├── Intro paragraph (preview sections using \Cref)
├── 2.1 Automation and the Labor Market
│   ├── 2.1.1 The Task-Based Framework
│   ├── 2.1.2 Labor Market Polarization
│   ├── 2.1.3 Task Displacement and Rent Dissipation
│   └── 2.1.4 Bridge paragraph → politics
├── 2.2 Economic Shocks and Political Attitudes
│   ├── 2.2.1 The China Shock and Political Consequences
│   ├── 2.2.2 Automation and Voting Behavior
│   └── 2.2.3 Methodological Parallels (NEW)
├── 2.3 Mechanisms: From Economic Disruption to Political Change
│   ├── 2.3.1 The Economic Insecurity Thesis
│   ├── 2.3.2 The Cultural Backlash Thesis
│   ├── 2.3.3 Status Threat and the 2016 Election
│   ├── 2.3.4 Psychological Foundations (DEEPENED)
│   └── 2.3.5 Identity Threat as Integrating Framework (NEW)
└── 2.4 Contribution of This Thesis

PHASE 1: RESEARCH (parallel)

Launch TWO agents in parallel using the Task tool:

Agent 1: Extraction Assembler (sonnet, read-only)

subagent_type: general-purpose
model: sonnet

Prompt:

You are assembling a structured evidence dossier for a thesis literature review rewrite.

Read ALL files in thesis/references/literature_analysis/paper_extractions/*.json (12 files), the concept graph at thesis/references/literature_analysis/concept_graph.json, and the mechanism extraction at thesis/references/literature_analysis/mechanism_extraction.json.

Produce TWO JSON files in thesis/chapters/.litreview_workspace/:

dossier.json — Organize ALL extracted evidence by the target section structure:

{
  "2.1.1_task_framework": {
    "papers": ["autor_levy_murnane_2003"],
    "key_quotes": [{"text": "...", "page": N, "bibkey": "autor2003skill"}],
    "claims": [{"claim": "...", "evidence": "...", "strength": "strong"}],
    "concept_edges": ["task_model → routine_manual", "task_model → substitution"],
    "constructs": [{"name": "...", "definition": "..."}]
  },
  "2.1.2_polarization": { ... },
  ...
}

gaps.json — Identify areas where extractions are thin and RAG queries would help:

{
  "gaps": [
    {
      "section": "2.1.3",
      "topic": "rent dissipation mechanism details",
      "query": "rent dissipation automation wage premium loss",
      "reason": "Extraction has claim but no detailed mechanism quote"
    },
    ...
  ]
}

Map evidence to these sections:

• 2.1.1: autor_levy_murnane_2003 (task framework, DOT, routine/nonroutine)
• 2.1.2: autor_dorn_2013 (polarization, U-shape, commuting zones, shift-share)
• 2.1.3: acemoglu_restrepo_2025 (task displacement, labor share, rent dissipation)
• 2.2.1: autor_dorn_hanson_majlesi_2020, colantone_stanig_2018 (China shock politics)
• 2.2.2: frey_berger_chen_2018, anelli_colantone_stanig_2019 (robots and voting)
• 2.2.3: methodological parallels (shift-share across studies, pseudo-panel)
• 2.3.1: economic insecurity evidence (Colantone & Stanig, Norris/Inglehart)
• 2.3.2: inglehart_norris_2016, norris_inglehart_2018 (cultural backlash)
• 2.3.3: mutz_2018, morgan_2018 (status threat debate)
• 2.3.4: osborne_2023, stenner_2005 (psychological mechanisms, dual-process)
• 2.3.5: manunta_2025, gidron_hall_2017 (identity threat integration)
• 2.4: contribution positioning

Include ALL verbatim quotes with page numbers. Be exhaustive — this is the evidence base for the entire chapter.

Agent 2: RAG Deepener (sonnet, Bash access)

subagent_type: general-purpose
model: sonnet

Prompt:

You supplement the literature dossier with additional evidence from the thesis RAG corpus (105 papers, 10,217 chunks in ChromaDB).

First, if thesis/chapters/.litreview_workspace/gaps.json exists, read it and use those queries. If not yet available, use the default queries below.

For each query, run:

cd /Users/alessandro/Projects/Tesi/thesis/references/RAG && source venv/bin/activate && python /Users/alessandro/Projects/Tesi/.claude/skills/verify-claims/scripts/query_rag.py "QUERY" --top-k 5 --json

Run 20-30 queries targeting these underdeveloped topics:

Section 2.1 (Automation & Labor):

1. "task is routine if can be accomplished by machines following explicit programmed rules"
2. "Dictionary of Occupational Titles DOT task measures"
3. "labor market polarization U-shape employment"
4. "commuting zone shift-share Bartik instrument"
5. "task displacement labor share decline automation"
6. "rent dissipation wage premium automation"
7. "between-group inequality automation 52 percent"

Section 2.2 (Shocks & Politics): 8. "China import shock political polarization electoral consequences" 9. "trade exposure radical right nationalist party Europe" 10. "economic nationalism protectionism import competition" 11. "robot exposure Trump presidential election swing states" 12. "counterfactual occupation assignment automation voting" 13. "robot exposure distinct from trade exposure"

Section 2.3 (Mechanisms): 14. "economic insecurity hypothesis pocketbook egotropic sociotropic" 15. "cultural backlash silent revolution post-materialist values" 16. "status threat group dominance racial resentment" 17. "Morgan critique Mutz 51 vote switchers fixed effects" 18. "dual-process model RWA SDO authoritarianism" 19. "authoritarian dynamic normative threat activation" 20. "identity threat mediates economic cultural populism" 21. "subjective social status populist right" 22. "dangerous world belief right-wing authoritarianism" 23. "competitive jungle social dominance orientation"

Section 2.4 (Contribution): 24. "pseudo-panel repeated cross-section Deaton synthetic cohort" 25. "IRT ideology estimate ideal point continuous measure"

Save results to thesis/chapters/.litreview_workspace/rag_supplements.json:

{
  "queries": [
    {
      "query": "...",
      "target_section": "2.1.1",
      "results": [
        {"source": "Author (Year)", "page": N, "text": "...", "similarity": 0.85}
      ]
    }
  ]
}

Prioritize results with similarity > 0.70 that contain verbatim-quotable text with specific numbers, definitions, or methodological details.

Wait for both agents to complete before proceeding to Phase 2.

PHASE 2: ARCHITECTURE (single agent)

Agent 3: Narrative Architect (opus)

subagent_type: general-purpose
model: opus

Prompt:

You are designing the narrative architecture for a thesis literature review chapter. You have access to:

1. Evidence dossier: thesis/chapters/.litreview_workspace/dossier.json
2. RAG supplements: thesis/chapters/.litreview_workspace/rag_supplements.json
3. Current Ch.2: thesis/chapters/02_literature.tex (the version being replaced)
4. Literature synthesis: thesis/references/literature_analysis/literature_synthesis.md
5. Chapter openings for forward refs:
   • thesis/chapters/03_data.tex (first 30 lines)
   • thesis/chapters/04_methodology.tex (first 30 lines)
   • thesis/chapters/05_results.tex (first 30 lines)

Read all of these files. Then produce thesis/chapters/.litreview_workspace/outline.md with this structure:

# Chapter 2 Outline: Literature Review

## Introductory Paragraph
- Preview all four sections using \Cref{sec:lit_automation}, etc.
- ~8-10 lines

## Section 2.1: Automation and the Labor Market
\label{sec:lit_automation}
Estimated: ~100-120 lines

### 2.1.1 The Task-Based Framework \label{subsec:task_model}
- ~25-30 lines
- Topic sentence: [exact sentence]
- Paper ordering: ALM 2003 only
- Key quotes to include: [list with page numbers]
- Concepts to define: routine vs nonroutine, DOT operationalization, substitution/complementarity
- Forward ref: \Cref{ch:data} (O*NET as modern DOT successor)
- Transition to 2.1.2: [sentence bridging to polarization]

### 2.1.2 Labor Market Polarization \label{subsec:polarization}
[same structure...]

[continue for ALL subsections...]

## Citation Plan
| BibTeX key | Sections used | Role (primary/supporting) |

## Forward Reference Map
| \Cref target | Section where introduced | Context |

## Estimated Total: ~420 lines

Critical requirements:

• Every subsection must have a specific topic sentence drafted
• Every transition between subsections must be planned
• Every direct quote must be assigned to a specific location with page citation
• The outline must specify which concept graph edges inform each transition
• Section 2.2.3 (Methodological Parallels) is NEW — design it to bridge empirics → your methodology
• Section 2.3.5 (Identity Threat as Integrating Framework) is NEW — must synthesize the mechanisms debate
• Section 2.4 must connect to ALL of Ch.3, Ch.4, Ch.5

DISCURSIVE POSITIONING requirement: The literature review must read as a CONVERSATION with other researchers, not a catalog of findings. For each subsection, the outline must plan at least one "positioning passage" that follows this pattern (common in top economics journals like QJE, AER, JEPS):

Pattern A — "Closely related / distinct from": Name 2-3 closest papers → describe what they do → pivot with "distinct from" / "I depart from" / "unlike" → explain your approach. Example from Autor et al. (2024, QJE): "Our work is closely related to Webb (2020) and Kogan et al. (2021), who use NLP tools to identify innovations... Distinct from this literature, we develop a method to..."

Pattern B — "While X... I instead...": Compare a specific methodological choice by naming what others did and explaining your alternative. Example: "While Frey et al. (2018) aggregate robot exposure at the commuting-zone level, and Anelli et al. (2019) exploit individual-level variation through counterfactual occupation assignment, I adopt a group-level pseudo-panel..."

Pattern C — "Building on X, but extending...": Acknowledge what you inherit and add. Example: "Following the shift-share logic of Autor & Dorn (2013), I aggregate industry-level shocks to demographic groups using base-year employment weights—but unlike their fixed-baseline approach, I use rolling windows that capture recent automation dynamics."

Pattern D — "None of these studies... This is the gap": After reviewing 3-4 related papers, identify what is missing that your thesis addresses.

The outline must explicitly mark WHERE each positioning passage goes, using tags like: [POSITION: Pattern B — compare Frey CZ-level vs Anelli individual vs my group-level]

Key comparisons to plan (at minimum):

• Automation measure: robot counts (A-R 2020) vs RTI proxies (Autor-Dorn) vs task displacement (A-R 2025) → yours
• Unit of analysis: commuting zones (ADH, Frey) vs individual (Anelli) vs demographic groups (yours)
• Political outcome: vote share (Frey, ADHM) vs party ID vs IRT ideology (yours)
• Identification: county-level shift-share (Frey) vs individual counterfactual (Anelli) vs group-level TWFE (yours)
• Geography: US only (Mutz, Frey, ADHM) vs Europe (Colantone-Stanig, Anelli) vs your US state-level
• Temporal design: cross-section (Inglehart-Norris) vs panel (Mutz, Morgan) vs pseudo-panel (yours)

Write in the voice of a Bocconi Master's student: precise, measured, first person singular.

Wait for completion before Phase 3.

PHASE 3: DRAFTING (4 parallel agents)

Launch FOUR section-drafter agents in parallel using the Task tool. Each writes one .tex file.

Agent 4a-4d: Section Drafters (opus, 4 instances)

For each section (2.1, 2.2, 2.3, 2.4), use:

subagent_type: general-purpose
model: opus

Prompt template (adjust section number, label, and line target):

You are drafting Section 2.X of the Literature Review chapter for a Bocconi Master's thesis.

Read these files:

1. Your section in the outline: thesis/chapters/.litreview_workspace/outline.md (Section 2.X only)
2. Evidence for your section from: thesis/chapters/.litreview_workspace/dossier.json (keys starting with "2.X.")
3. RAG supplements: thesis/chapters/.litreview_workspace/rag_supplements.json (entries targeting "2.X.*")
4. BibTeX file: thesis/references.bib (to verify all keys exist)
5. Notation registry: .claude/rules/notation-registry.md
6. Citation standards: .claude/rules/literature-citations.md
7. Forward reference requirements: .claude/rules/forward-references.md

Write thesis/chapters/.litreview_workspace/drafts/sec_2.X.tex containing ONLY the LaTeX content for this section (no \chapter, no preamble).

Requirements:

• First person singular: "I review", "I examine", "I identify"
• Present tense for methodology descriptions, past tense for data/findings
• Use \citet{key} for author-as-subject, \citep{key} for parenthetical
• Every direct quote in ...'' `` must have \citep[p.~N]{key}
• Include \label{subsec:...} for each subsection
• Use \Cref{target} for forward references to Ch.3-5
• Section separators: % === for section header, % --- for subsections
• Target: ~[80-120] lines for Section 2.X
• End with a transition sentence to the next section (except 2.4)

DISCURSIVE STYLE — CRITICAL: Write as a CONVERSATION with the literature, not a paper-by-paper summary. Follow the patterns used in top economics journals (QJE, AER, JEPS). Specifically:

1. Compare and position: When discussing related papers, group them by what they share and where they differ. Example: "While \citet{frey2018robot} aggregate robot exposure at the commuting-zone level, \citet{anelli2019robots} exploit individual-level variation through a counterfactual occupation assignment. I adopt a different unit of analysis---demographic groups defined by education, gender, and age---motivated by the pseudo-panel design of \citet{deaton1985panel}."

2. Build intellectual genealogy: Show how ideas evolved across papers. Example: "The task displacement measure I employ builds on \citeauthor{autor2003skill}'s framework but operationalizes it through labor share changes rather than occupational task proxies. This direct measurement approach, developed by \citet{acemoglu2025automation}, captures all forms of automation---not just industrial robots."

3. Discuss methodology discursively: Don't just state what a paper found. Explain HOW they identified the effect and what that implies. Example: "\citet{mutz2018status} exploits within-person variation in a two-wave panel, finding that status threat---not personal economic hardship---predicts vote switching. \citet{morgan2018status} challenges this identification: with only 51 vote switchers in the panel, the fixed-effects estimates lack statistical power."

4. Use transition patterns that compare: Instead of "Next, I review X", write "Polarization research established that automation reshapes the labor market along predictable lines, but it relied on occupational task intensity as a proxy. A more direct measurement approach has recently become available."

5. NEVER write sequential summaries: "Paper A found X. Paper B found Y. Paper C found Z." Instead, weave them into a narrative: "A growing body of evidence links automation to right-wing politics. \citet{frey2018robot} show that US counties with greater robot exposure swung toward Trump, an effect distinct from the China trade shock documented by \citet{autor2020importing}. In Europe, \citet{anelli2019robots} confirm this pattern at the individual level, using a counterfactual occupation assignment that avoids the endogeneity of current job choice."

For Section 2.1: Cover ALM task framework, Autor-Dorn polarization, A-R task displacement + rent dissipation. Bridge to politics. Include positioning passages comparing automation measures (robot counts vs RTI vs task displacement). For Section 2.2: Cover China shock (ADHM, Colantone-Stanig), robot voting (Frey, Anelli), and NEW methodological parallels subsection. Emphasize discursive comparison of identification strategies across papers. For Section 2.3: Cover economic insecurity, cultural backlash, status threat + Morgan debate, psychological mechanisms (dual-process), and NEW identity threat integration. Present the Mutz-Morgan debate as a real intellectual exchange, not two separate summaries. For Section 2.4: Contribution positioning — task displacement > robot exposure, IRT ideology, pseudo-panel, mechanism adjudication. This section should be ENTIRELY structured around "closely related / distinct from" comparisons with prior work. Forward refs to Ch.3, 4, 5.

Do NOT verify BibTeX keys exist — just use the keys from the citation standards rule. If a key might be missing, add a % TODO: verify bibkey comment.

Wait for all 4 agents to complete before Phase 3.5.

PHASE 3.5: STYLE PARAPHRASING (4 parallel agents)

The section drafters (Phase 3) produce content with correct structure, citations, and arguments---but each agent writes in its own voice. This phase normalizes all four sections to the author's actual writing style, as demonstrated in Chapters 3-5.

Launch FOUR style-paraphraser agents in parallel using the Task tool (one per section):

Agent 4.5a-4.5d: Style Paraphraser (sonnet, 4 instances)

For each section (2.1, 2.2, 2.3, 2.4), use:

subagent_type: style-paraphraser
model: sonnet

Prompt template (adjust section number):

You are the style-paraphraser agent. Your task is to rewrite Section 2.X of the literature review draft so that it matches the author's voice from Chapters 3-5.

Step 1 — Internalize the author's style. Read 80 lines from each of:

• thesis/chapters/03_data.tex (lines 1-80)
• thesis/chapters/04_methodology.tex (lines 1-80)
• thesis/chapters/05_results.tex (lines 1-80)

Step 2 — Read the draft. Read: thesis/chapters/.litreview_workspace/drafts/sec_2.X.tex

Step 3 — Rewrite. Transform the prose paragraph by paragraph:

• Match the author's sentence rhythm (complex, clause-rich, 25-35 words, em-dash parentheticals)
• Use first person singular active voice ("I review", "I examine", not "we" or passive)
• Apply the author's transition patterns ("While X..., Y demonstrated that...")
• Use inline enumeration: "(i)..., (ii)..., and (iii)..." for method steps
• Use "Three patterns emerge. First,... Second,... Third,..." for findings
• Hedge with "consistent with", "suggesting", "may reflect" (never "clearly", "obviously")
• Introduce technical terms with \emph{} on first use
• No \textbf{} in running prose, no footnotes, no rhetorical questions

Step 4 — Preserve invariants. The rewrite MUST keep:

• Every \citet{}, \citep{}, \citeauthor{} citation (same keys, same positions)
• Every equation and its \label{}
• Every \Cref{} and \cref{} forward/back reference
• Every \label{subsec:...} and \label{sec:...}
• The same argument structure (claim ordering, evidence flow)
• All ...'' `` direct quotes with their page citations

Step 5 — Write output. Overwrite thesis/chapters/.litreview_workspace/drafts/sec_2.X.tex with the style-normalized version.

Before writing, verify: same number of \cite commands in input vs output. If any are missing, fix before saving.

Wait for all 4 agents to complete before Phase 4.

PHASE 4: VERIFICATION (parallel)

Launch TWO agents in parallel:

Agent 5: Claim Verifier (sonnet)

subagent_type: general-purpose
model: sonnet

Prompt:

You verify every factual claim in the new literature review draft.

Read all draft sections:

• thesis/chapters/.litreview_workspace/drafts/sec_2.1.tex
• thesis/chapters/.litreview_workspace/drafts/sec_2.2.tex
• thesis/chapters/.litreview_workspace/drafts/sec_2.3.tex
• thesis/chapters/.litreview_workspace/drafts/sec_2.4.tex

For each factual claim with a citation:

1. Check against paper extractions in thesis/references/literature_analysis/paper_extractions/*.json
2. For claims not covered by extractions, query RAG:

   cd /Users/alessandro/Projects/Tesi/thesis/references/RAG && source venv/bin/activate && python /Users/alessandro/Projects/Tesi/.claude/skills/verify-claims/scripts/query_rag.py "CLAIM" --top-k 3 --json
   

Produce thesis/chapters/.litreview_workspace/verification_report.md:

# Claim Verification Report

## Summary
- Total claims: N
- SUPPORTED: N (%)
- PARTIALLY SUPPORTED: N (%)
- UNSUPPORTED: N (%)
- CONTRADICTED: N (%)

## Detailed Results

### Section 2.1
| # | Claim | Citation | Status | Note |
|---|-------|----------|--------|------|
| 1 | "routine tasks..." | autor2003skill p.1283 | SUPPORTED | Exact quote verified |
...

## Corrections Needed
[List any claims that need rewording, with suggested corrections]

Agent 6: Forward Compatibility Checker (sonnet)

subagent_type: general-purpose
model: sonnet

Prompt:

You check that the new literature review properly sets up all concepts used in Chapters 3-5.

Read:

1. All draft sections in thesis/chapters/.litreview_workspace/drafts/
2. Full Chapter 3: thesis/chapters/03_data.tex
3. Full Chapter 4: thesis/chapters/04_methodology.tex
4. Full Chapter 5: thesis/chapters/05_results.tex

Check:

1. Every concept first used in Ch.3-5 that should be introduced in Ch.2 IS introduced
2. Every \Cref forward reference in the drafts points to a label that exists in Ch.3-5
3. Every BibTeX key used in the drafts exists in thesis/references.bib
4. Notation consistency: any math symbols used in Ch.2 match the notation registry

Produce thesis/chapters/.litreview_workspace/compat_report.md:

# Forward Compatibility Report

## Missing Concept Setups
[Concepts used in Ch.3-5 but not introduced in Ch.2]

## Broken Forward References
[Any \Cref targets that don't exist]

## Missing BibTeX Keys
[Any keys used in drafts but not in references.bib]

## Notation Issues
[Any symbol inconsistencies]

## Status: [PASS / ISSUES FOUND]

Wait for both agents to complete before Phase 5.

PHASE 5: REFINEMENT (single agent)

Agent 7: LaTeX Polisher (sonnet, write access)

subagent_type: general-purpose
model: sonnet

Prompt:

You are the final assembler for the literature review chapter. You combine section drafts into a polished 02_literature.tex.

Read:

1. All section drafts in thesis/chapters/.litreview_workspace/drafts/
2. Verification report: thesis/chapters/.litreview_workspace/verification_report.md
3. Compatibility report: thesis/chapters/.litreview_workspace/compat_report.md
4. Current chapter header pattern: thesis/chapters/02_literature.tex (first 8 lines only — preserve the chapter/label format)
5. LaTeX conventions: .claude/rules/latex-conventions.md
6. Forward references: .claude/rules/forward-references.md

Tasks:

1. Apply corrections from the verification report (fix any UNSUPPORTED or CONTRADICTED claims)
2. Add missing forward refs from the compatibility report
3. Add missing BibTeX entries — if keys are used in the draft but missing from thesis/references.bib, list them as % TODO: add to references.bib comments
4. Write the intro paragraph previewing all sections with \Cref references
5. Assemble all sections into a single file with consistent formatting:
   • Chapter header with % === separators
   • Section headers with % ===, subsection with % ---
   • Consistent spacing
6. Write the final file to thesis/chapters/02_literature.tex

The file must:

• Start with % =====... / \chapter{Literature Review} / \label{ch:literature}
• Be ~400-500 lines total
• Use ONLY \citet, \citep, \citeauthor for citations (never manual "Author (Year)")
• Have all \label prefixes correct: sec:, subsec:
• End with no trailing blank lines after the last \section

POST-PIPELINE: Compilation Test

After Phase 5 completes, compile the thesis:

cd /Users/alessandro/Projects/Tesi/thesis && tectonic main.tex 2>&1

Check for:

• Errors: Must fix before declaring success
• Undefined references: List any broken \Cref targets
• Undefined citations: List any missing BibTeX keys → add them to references.bib

If there are undefined citations, add the missing entries to thesis/references.bib and recompile.

Cleanup

After successful compilation, optionally remove the workspace:

rm -rf /Users/alessandro/Projects/Tesi/thesis/chapters/.litreview_workspace

Ask the user before cleaning up — they may want to inspect intermediate artifacts.

BibTeX Key Reference

Additional keys available in references.bib: acemoglu2020robots, autor2013china, colantone2021backlash, colantone2019surge, manunta2025identity, stenner2005authoritarian, duckitt2009dual, duckitt2017dual, deaton1985panel, gidron2017politics, gest2018nostalgic, feldman1997perceived, stenner2018momentary, salmela2017emotional, jost2003political, ferwerda2025nostalgic