review-landscape

name	review-landscape
description	Strategic field reconnaissance for a domain you know a little but not enough. Produces an 8-dimension landscape map ending in a gap table ranked by a 3-filter comparative-advantage test against the user's own research pillars. Domain coverage: complexity science, urban science, urban sustainability, industrial ecology, economic geography. Scouting mode only — no prose drafting, no PDF ingestion. Conductor pattern: delegates verification and deep-read work to other Lu Lab skills.
allowed-tools	["Agent","WebSearch","WebFetch","Read","Write","Edit","Glob","Grep","Bash","AskUserQuestion"]
user-invocable	true

Review-Landscape: Field Reconnaissance Orchestrator

Usage: /review-landscape "<topic>" [mode] [domain...]

You are the conductor for a field reconnaissance task. You do not do the searching or reading yourself — you dispatch parallel subagents at the right model tier, aggregate their findings, and produce an 8-dimension landscape map ending in a gap table ranked by the user's stated research-pillar comparative advantage.

References (load on demand):

• references/venues.md — domain journal + preprint + database lists
• references/theoretical-lineages.md — founder chains + named-law catalog per domain
• references/strength-fit-rubric.md — 3-filter comparative-advantage test + attack-mode catalog

---

Hard Constraints

These are inviolable. Refuse any attempt to relax them.

• No fabricated citations. Every DOI must be CrossRef-verifiable. Flag unverifiable as [CLAIM UNVERIFIED].
• No vague gaps. "Understudied" is banned; every gap must name the closest prior paper + why it fails.
• No publication-potential ranking. Only comparative-advantage ranking against the user's pillars.
• No prose draft output. Scouting only. The user can run a drafting skill afterward.
• No scoop-risk check. Reconnaissance is reconnaissance.
• No auto-detect of domain. User specifies.
• PRISMA counting only in full mode.
• No PDF downloads. Reconnaissance ≠ ingestion. Suggest /read-ken for deep-read of any retained paper.
• No bib-note creation. Hand off to the user's bib pipeline.

---

Phase 0: Parse Inputs & Initialize

Required inputs

Parse from the invocation:

• Topic: quoted phrase (required)
• Mode: rapid | targeted (DEFAULT) | full
• Domain: one or more of complexity, urban, urban-sustainability, industrial-ecology, economic-geography (required; if ambiguous, ask the user via AskUserQuestion)

Pillar elicitation (one-time per run)

The 3-filter test in Phase 4h needs the user's three research pillars (e.g., "evolution × networks × complex systems", "thermodynamics × ecology × policy"). Ask once via AskUserQuestion:

1. Pillars: "What three concepts define your research edge? These will be used to score gaps for comparative-advantage fit."
2. Active edges (optional): "Any 2–3 active project keywords I should check gaps against?" (skip if user has nothing in flight)

Save these to $OUTPUT_DIR/pillars.md for Phase 4h.

Optional inputs

• --output-dir /path/ — override default output location (default: ./output/review-landscape/$SLUG-$DATE/)

Create output directory

SLUG=<kebab-case topic>
DATE=<YYYY-MM-DD>
OUT_DIR="${OUTPUT_DIR:-./output/review-landscape/$SLUG-$DATE}"
mkdir -p "$OUT_DIR"

Files to create:

• $OUT_DIR/search-log.md — seed with frontmatter: topic, mode, domains, date, pillars
• $OUT_DIR/landscape-map.md — created empty; populated in Phase 4
• $OUT_DIR/pillars.md — user's pillars + active edges, written from Phase 0 elicitation

Initialize task tracking

Use the task tracking tool to register phases 1–6 as top-level tasks with status pending. Per-subagent sub-tasks are created as each phase fires.

Seed the search log

Write initial metadata to search-log.md:

# Search Log — [topic] ([date])

- **Mode**: [rapid/targeted/full]
- **Domains**: [list]
- **Output**: [OUT_DIR]
- **Pillars**: [user's three pillars]

## Queries
(appended after each wave)

---

Phase 1: Local Inventory Search (Haiku, parallel — if available)

If the working tree contains markdown index files (e.g., a curated reading list, a notes index, a project-local bibliography), grep them first — local knowledge beats web search.

Detection:

find . -maxdepth 4 -type f \( -name "*inventory*.md" -o -name "*index*.md" -o -name "references.md" -o -name "reading-list.md" \) 2>/dev/null

If files are found, dispatch one Haiku 4.5 subagent per file. Each subagent:

• Greps the file for topic keywords [K1, K2, K3].
• For each hit, follows links to the source note (if any) and extracts YAML metadata (authors, year, DOI, summary).
• Returns: at most 20 hits sorted by relevance.

Aggregation: collect all results. Append to search log. If all subagents return 0 hits, flag: "Field is cold in local notes — weight Phase 2 web coverage heavier."

If no index files exist: log "No local index — web-first mode" in search log, skip Phase 1, proceed to Phase 2.

PDF policy: never re-read a PDF when a digest already exists in the local note. Trust prior summaries.

Mark Phase 1 task as completed.

---

Phase 2: Multi-Wave Web Search (Haiku + Sonnet, parallel)

Dispatch parallel subagents per wave × per source. Wave composition by mode:

Wave	Rapid	Targeted	Full
1. Core topic	✓	✓	✓
2. Theoretical pillars	✓	✓	✓
3. Mechanisms / frontier	—	✓	✓
4. Contested / null	—	✓	✓
5. Cross-level bridge	—	—	✓

Wave 1 — Core topic (Haiku × 4 sources)

Four parallel Haiku subagents, one each for Google Scholar, Semantic Scholar, arXiv, Web of Science. Query template:

"[main concept]" AND ["mechanism" OR "scaling" OR "dynamic"] AND "[domain keyword]"

Each subagent returns: title, authors, year, venue, DOI, abstract snippet, 1-sentence relevance.

Target: 10–15 total hits, retain 5–8 after dedup.

Wave 2 — Theoretical pillars (Haiku × domain)

One Haiku subagent per domain queried. Each reads references/theoretical-lineages.md for that domain's founder chain, then searches for recent extensions. Target: 5–8 hits, retain 3–5.

Example dispatch for economic-geography domain:

Founder chain (from references/theoretical-lineages.md):
Marshall, Krugman, Taylor, Derudder, Bettencourt, Lobo, Jacobs, Glaeser.
Search for: (a) recent works citing 2+ of these founders, (b) named-law papers
(e.g., Krugman 1991 core-periphery, Zipf rank-size, Christaller central-place),
(c) Annual Review articles in economic geography.
Return at most 10 hits.

Wave 3 — Mechanisms / frontier (Sonnet × preprint sources, targeted+full only)

Sonnet because mechanism identification needs more careful reading. Parallel dispatch: arXiv (physics.soc-ph, nlin.AO), SSRN, SocArXiv, bioRxiv. Query adds recent year filter (last 5y).

Wave 4 — Contested / null findings (Sonnet × sources, targeted+full only)

Sonnet for careful distinction between genuine null and methodological critique. Queries include "no effect", "null", "contradicts", "comment on", "revisited".

Wave 5 — Cross-level bridge (Sonnet × adjacent levels, full only)

If topic lives at level N, query level N-1 and N+1 for porting candidates. Sonnet handles level-mapping judgment.

After each wave: append query, hit count, retained-paper rationale to search-log.md. Paper inventory table grows incrementally.

Mark Phase 2 task as completed.

---

Phase 3: Review Articles & Citation Chains (Haiku, parallel)

Select top 3 retained papers from Waves 1–2 by citation count + theoretical centrality.

For each, dispatch a Haiku subagent running Semantic Scholar forward+backward citation via API:

PAPER_DOI="..."
curl -s "https://api.semanticscholar.org/graph/v1/paper/DOI:$PAPER_DOI?fields=title,year,authors,citationCount,citations.title,citations.year,citations.authors,references.title,references.year&limit=50" | python3 -m json.tool

Subagent returns: top-10 forward citations (sort by year desc), top-10 backward references (sort by citation count desc), flagged foundational works (pre-2000 or 200+ citations).

DOI verification (mandatory): for every new DOI about to be added to the inventory, run:

curl -s "https://api.crossref.org/works/$DOI" | python3 -m json.tool

Verify title + authors + year match claim. Flag mismatches. For high-stakes runs, invoke /manage-refs.

Review-venue sweep (Haiku × review venues): search Annual Review of Environment and Resources, Physics Reports, Nature Reviews Earth & Environment, etc. (see references/venues.md §Review venues) for review articles on the topic.

Mark Phase 3 task as completed.

---

Phase 4: Build the Landscape Map (Opus synthesis)

Model: Opus 4.7. This is the deliverable — do not delegate to a cheaper model.

Write the 8 dimensions to $OUT_DIR/landscape-map.md.

4a. Field Evolution Timeline (prose paragraph)

Founder paper + consolidation date, 1–2 paradigm shifts, current frontier.

4b. Theoretical Landscape

Two forms, pick based on field maturity:

Form A — Multi-trunk ancestry tree (preferred for mature fields with ≥2 parent disciplines, ancestors pre-1950). Produce a Haberl-2019-style tree grouped by parent trunk. See references/theoretical-lineages.md §"Tree-rendering template" for the exact markdown structure. Each ancestor node: **Author YEAR** — [contribution in <15 words].

Form B — Framework table (for young fields, <30 years of ancestry):

| Framework | Proponents | Core prediction | Empirical support | Scope conditions | Named-law anchor | Known weakness |

Named-law anchor is critical — it's the handle for attack. Catalog in references/theoretical-lineages.md §"Named-law anchor catalog".

When in doubt, produce BOTH forms.

4c. Established Findings

5–10 well-replicated claims with effect sizes, 3+ replications, remaining ambiguity.

4d. Contested Findings

Both sides, source of disagreement, what would settle it.

4e. Null / Absent Findings

Strictly distinguish null (tested, no effect) from absent (never tested).

4f. Mechanisms Inventory (table — ATTACK SURFACE)

| Mechanism | Proposed by | Type | Empirical status (TESTED / INDIRECT / ASSUMED / CONTESTED) | Testing method | What's missing |

ASSUMED is the primary attack surface.

4g. Methodological Landscape

Dominant data types, designs, computational tools, unused-tool gap, unused-data gap.

4h. Research Gaps — Pillar-Ranked

Pre-flight: read $OUT_DIR/pillars.md to load the user's three pillars + active edges.

Pass 1: abstract 3-filter test using user's pillars (see references/strength-fit-rubric.md). Pass 2: active-edge overlap — if the user supplied active edges in Phase 0, check each gap for DIRECT / ADJACENT / NEW TERRITORY overlap. If no active edges supplied, skip Pass 2.

| # | Gap statement | Closest prior paper | Why prior fails | F1 (pillars) | F2 (competition) | F3 (multi-path) | Score | Active-edge overlap | Attack mode | Notes |

Full protocol in references/strength-fit-rubric.md — including the attack-mode catalog. Sort by combined score, then by connection strength (DIRECT > ADJACENT > NEW TERRITORY).

4i. Theory Handoff (for top-2 gaps by score)

Sketch minimal model: state variables (target: 2–3), simplest mechanism, empirical calibration. Do not create new project notes autonomously — surface the sketch and let the user decide.

Mark Phase 4 task as completed.

---

Phase 5: Verification (Sonnet)

Dispatch one Sonnet subagent to audit the landscape map on 5 criteria:

1. Coverage — founder chain for each domain queried is represented
2. Synthesis — findings synthesized, not just listed
3. Gap specificity — every gap names closest prior paper
4. Mechanism discipline — every mechanism has a 4-level status
5. Balance — both supporting and contesting evidence present

Return PASS or specific revision items. Failed dimensions re-run via targeted Opus subagent.

Mark Phase 5 task as completed.

---

Phase 6: Save Outputs & Report

1. Confirm $OUT_DIR/search-log.md, $OUT_DIR/landscape-map.md, and $OUT_DIR/pillars.md exist and verification passes.
2. Print summary to user:
   • Output file paths
   • N papers retained
   • Top-3 gaps by score
   • Recommended next steps (see below)

Post-run suggestions (print, do not auto-execute)

• New DOIs worth deep-reading → suggest /read-ken on top-priority ones
• Citation pool needs verification → suggest /manage-refs
• Cross-level port candidates → suggest /trace-origin on the upstream framework
• Score-3 gaps that map to a new project → suggest /launch-project

Mark Phase 6 task as completed.

---

Parallelism Architecture Summary

Opus orchestrator (you) + worker subagents (Haiku/Sonnet, dispatched via Agent tool with explicit model parameter):

Task	Model	Count
Phase 1 inventory grep	Haiku	1 per detected inventory file
Phase 2 Wave 1 search	Haiku	4 parallel (one per source)
Phase 2 Wave 2 lineage	Haiku	1 per domain
Phase 2 Wave 3 mechanisms	Sonnet	4 parallel (preprint sources)
Phase 2 Wave 4 contested	Sonnet	2 parallel
Phase 2 Wave 5 cross-level	Sonnet	2 parallel
Phase 3 citation chain	Haiku	3 parallel (one per top paper)
Phase 3 review venue sweep	Haiku	1–3 parallel
Phase 4 landscape synthesis	Opus	1 (you)
Phase 4h gap ranking	Opus	1 (you)
Phase 5 verification	Sonnet	1

Target parallelism:

• rapid: 5–8 subagents total
• targeted: 10–15 subagents
• full: 20+ subagents

Always register each subagent dispatch in the task list so the user sees progress live.

---

Error Handling

• No inventory files found: skip Phase 1, log as "No local inventory — web-first mode."
• API rate limit (Semantic Scholar 100 req/5min without key): back off, serialize the affected wave.
• CrossRef mismatch: flag the paper, do NOT include in inventory, note in search log.
• 0 hits in Phase 2 Wave 1: try broader keyword set; if still 0, report "Topic unrecognizable in literature — check spelling / try related terms."
• Landscape map fails Phase 5 verification twice: return failures to user; do not auto-retry more than once.
• PDF URL returns paywall: skip silently, note DOI in inventory.

---

Invocation Examples

/review-landscape "urban infrastructure industrial ecology" targeted urban-sustainability industrial-ecology

/review-landscape "world cities network" full economic-geography urban

/review-landscape "circular economy at city scale" rapid industrial-ecology

/review-landscape "closure laws in growing urban systems" full complexity urban urban-sustainability

---

Tone & Output Style

• Be terse. The user will read many landscape maps; don't pad.
• Tables over prose where possible.
• Cite author-year inline; full DOI in the search log only.
• Do not editorialize about papers; report what the literature says, flag what's contested.
• For the final summary: 3–5 bullet points max. The user reads the map itself.

---

This skill is complete only when: (a) landscape-map.md exists with all 8 dimensions populated, (b) Phase 5 verification returns PASS, (c) pillars.md records the user's stated pillars for traceability. Otherwise the skill failed.

---

Upstream / Provenance

This skill is a fork of the scholar-lit-review skill in open-scholar-skill by Yongjun Zhang (@joshzyj · joshzyj.github.io). The core design is his: the 8-dimension landscape map (field evolution → theoretical landscape → established / contested / null findings → mechanisms inventory → methodological landscape → research gaps → theory handoff) and the search → map → verify pipeline.

What the Lu Lab changed for its use case: retargeted from social science to five domains (complexity, urban, urban-sustainability, industrial ecology, economic geography); swapped the gap ranking from publication-potential to a configurable comparative-advantage test scored against the user's own research pillars (elicited per run, not hard-coded); restricted scope to scouting (no review-draft writing); and rewired the handoffs to Lu Lab skills. Thank you, Yongjun.

Released under Yongjun's Open Scholar Skill License (Academic Use): free for academic, educational, and non-commercial research; commercial use requires his written permission. The full copyright and permission notice is retained in NOTICE and travels with this skill.