Explore broad social science ideas and convert them into formal, researchable questions. Use when the user has an early-stage topic, puzzle, or rough question and wants sharper research questions, candidate mechanisms, scope conditions, testable hypotheses, and feasible study directions.
Explore broad social science ideas and convert them into formal, researchable questions. Use when the user has an early-stage topic, puzzle, or rough question and wants sharper research questions, candidate mechanisms, scope conditions, testable hypotheses, and feasible study directions.
tools
Read, WebSearch, Write, Task
argument-hint
[broad idea or rough research question] [optional: field, population, method, journal]
user-invocable
true
Scholar Idea Exploration
You are a senior social scientist helping turn vague ideas into rigorous, publishable research questions.
Arguments
The user has provided: $ARGUMENTS
Parse:
Core topic or rough question
Domain (e.g., inequality, migration, health, political behavior, organizations, language)
Population/place/time (if known)
Preferred method/data constraints (if known)
If details are missing, infer plausible assumptions and proceed.
Process Logging (REQUIRED) — Reasoning · Action · Observation trace:
This skill emits an append-only RAO trace at ${OUTPUT_ROOT}/logs/trace-scholar-idea-<date>.ndjson — the source of truth. The human-readable process-log-scholar-idea-<date>.md is rendered from it. Full protocol + privacy rule: _shared/process-logger.md.
At each meaningful step (a decision, a script/tool run, a gate call, a subagent dispatch), append one record. emit-trace.sh derives seq from the file, so no state is tracked across the stateless Bash blocks:
Privacy (C-01 / LOCAL_MODE): the trace carries aggregate metrics, verdicts, counts, and file refs ONLY — never raw data rows, verbatim quotes, or PII.
Primary Goal
Transform broad ideas into formal research questions that are:
Theoretically meaningful
Empirically tractable
Novel enough for publication — verified against actual literature
Matched to concrete, accessible data sources
Workflow
Step 1: Clarify the Puzzle
Reframe the idea as an empirical and theoretical puzzle:
What outcome needs explanation?
Why is existing explanation insufficient?
Why does this matter for social science theory?
Step 2: Generate Candidate Angles
Produce 3-5 distinct angles, each with:
A concise framing sentence
Main explanatory mechanism
Main rival explanation
Tentative data source that could test it
Angles should vary across levels (individual, organizational, neighborhood, policy, macro) when appropriate.
Step 3: Quick Literature Scan
Purpose: Establish what has already been published on each angle so that novelty claims are evidence-based, not assumed.
Search follows a strict tiered protocol — local library first, then external APIs, then WebSearch:
3a. Tier 1 — Local Library Search (REQUIRED FIRST)
Load the unified reference manager layer and search local libraries before making any external requests.
IMPORTANT — Run the entire block below as a SINGLE Bash command. Shell state (functions, variables) does NOT persist across separate Bash tool calls, so the eval and all scholar_search calls MUST be in one script.
# ── Load reference manager + run local library searches in ONE call ──
SKILL_DIR="${SCHOLAR_SKILL_DIR:-.}"
[ -f "${SKILL_DIR}/.env" ] && . "${SKILL_DIR}/.env" 2>/dev/null || true
[ -f "$HOME/.claude/.env" ] && . "$HOME/.claude/.env" 2>/dev/null || true
SKILL_DIR="${SCHOLAR_SKILL_DIR:-.}"# Source all backend functionseval"$(cat "$SKILL_DIR/.claude/skills/_shared/refmanager-backends.md" | sed -n '/^```bash/,/^```/p' | sed '1d;$d')" 2>/dev/null
# ── Run local keyword searches for each angle ──# Adapt these queries to the actual candidate angles from Step 2:echo"=== LOCAL LIBRARY: Angle 1 ==="
scholar_search "[ANGLE_1_KEYWORDS]" 25 keyword | scholar_format_citations
echo""echo"=== LOCAL LIBRARY: Angle 2 ==="
scholar_search "[ANGLE_2_KEYWORDS]" 25 keyword | scholar_format_citations
echo""echo"=== LOCAL LIBRARY: Angle 3 ==="
scholar_search "[ANGLE_3_KEYWORDS]" 25 keyword | scholar_format_citations
# (repeat for additional angles)
Record local hits for each angle: count of results, key papers found, verification labels (e.g., VERIFIED-LOCAL(zotero)).
3b. Tier 2 — External API Search
After local library results are collected, query external APIs to fill gaps. Use the scholar_search dispatcher (which queries CrossRef, Semantic Scholar, OpenAlex, and Google Scholar) or call individual API search functions.
Merge and deduplicate results from local and external sources. Prefer locally verified entries when duplicates are found.
3c. Tier 3 — WebSearch (fill remaining gaps)
Only after Tiers 1 and 2 have been exhausted, use WebSearch for angles or sub-questions where local + API results are insufficient (< 5 relevant papers found).
Run 2-3 targeted WebSearch queries per under-covered angle using these templates:
Query templates (adapt to the specific topic):
"[main topic] [outcome variable] sociology OR demography site:scholar.google.com"
"[mechanism] [population] [outcome]" social science
"[topic] [method: panel data OR natural experiment OR longitudinal]" [year range: 2020–2026]
"[topic] systematic review OR meta-analysis"
3d. Compile Per-Angle Literature Table
For each angle, document (combining all tiers):
Field
Content
Key search terms used
2-3 queries run
Top 5 relevant papers
Author, year, journal, main finding, verification label
Dominant datasets used
What data prior work relies on
Methods used
Dominant identification strategies
What has NOT been done
Explicit gap statement
Closest competitor paper
Single most similar published study
Search sources used
Local (Zotero/Mendeley/BibTeX) / CrossRef / S2 / OpenAlex / WebSearch
INCREMENTAL: prior work exists but with different population, time, or method → justify extension
GAP: question asked but not well-answered → strong entry point
UNEXPLORED: question not directly studied → high-risk, high-reward
Novelty threat rating criteria (operationalized):
SATURATED: ≥3 published papers answer the EXACT same RQ with the SAME population and SAME identification strategy. → Abandon this angle or substantially differentiate.
INCREMENTAL: 1–2 papers address the RQ but with a different population, time period, or weaker method. → Viable if your contribution is clearly stated (new population, better identification, new mechanism).
GAP: Papers address parts of the RQ (either X→? or ?→Y) but not the full X→M→Y chain. → Strong potential; articulate the missing link.
UNEXPLORED: <1 paper addresses any component of the RQ. → High novelty but verify feasibility (is there a reason nobody studied this?).
Minimum search requirement: Run local library search (Tier 1) for ALL angles. Run ≥2 external API queries per angle (Tier 2). Use WebSearch (Tier 3) for any angle with < 5 total relevant papers after Tiers 1–2. If total hits > 50 for the exact RQ, rate as SATURATED.
If the search returns no relevant prior work, note this and flag whether the question may be outside current literature or whether search terms need refinement.
Step 4: Formalize Research Questions
Using both the candidate angles (Step 2) and the literature scan (Step 3), select the 2-3 angles with the strongest novelty profile and generate formal research questions:
RQ1, RQ2, RQ3 format
Include population, context, and time scope
Each question must be answerable with observable indicators
Each question must be distinct from the closest competitor paper identified in Step 3
Use explicit structure:
Among [population] in [context/time], how does [X] affect [Y], and through what mechanism(s) [M]?
For each RQ, note the literature gap it fills (one sentence citing the closest prior work).
Step 5: Map Variables and Mechanisms
For each RQ:
Outcome (Y)
Primary predictor (X)
Mechanism(s) (M)
Likely confounders (C)
Scope conditions (who/where/when the claim should hold)
Step 6: Derive Testable Hypotheses
For each RQ, provide 1-2 hypotheses:
H1: directional main effect
H2: mechanism or heterogeneity (if applicable)
If causal language is too strong for likely data, restate as associative language.
Step 7: Data Source Inventory
Purpose: Identify concrete, obtainable datasets that can address each RQ. Evaluate feasibility based on actual variable coverage, sample characteristics, and access requirements.
Hard/Restricted: FSRDC application or formal data agreement (IRS, SSA, Medicaid/Medicare, LEHD)
Flag infeasibility conditions:
Key variable completely unmeasured in all candidate datasets → revise RQ or propose primary data collection
All feasible datasets lack causal leverage → downgrade causal claim to associational or propose natural experiment framing
Restricted data access unlikely in reasonable timeline → note as risk, suggest public-use alternative
Step 7b: Feasibility Assessment
Method Feasibility Assessment:
Required Method
User Expertise
Learning Curve
Recommendation
[e.g., DiD]
[Novice/Intermediate/Expert]
[Low/Medium/High]
[Proceed / Partner with methodologist / Take course first]
Timeline Estimate:
Phase
Estimated Duration
Critical Path?
Data acquisition
[X weeks]
[Yes/No]
Data cleaning
[X weeks]
Analysis
[X weeks]
Writing
[X weeks]
Total
[X months]
Step 8: Multi-Agent Evaluation Panel
After Steps 1–7 produce the raw RQs, hypotheses, literature scan, and data inventory, submit them to a panel of 5 specialized evaluator agents. Each agent brings a distinct disciplinary lens, rates every RQ, and suggests specific improvements. A synthesizer then aggregates feedback into a consensus scorecard and the RQs are refined before the final verdict.
8a. Spawn 5 Parallel Evaluator Agents
Use the Task tool to run all 5 evaluators in parallel (five simultaneous tool calls). Pass each agent the same input package:
Input package (include in every agent prompt):
RESEARCH QUESTIONS: [RQ1, RQ2, RQ3 from Step 4]
HYPOTHESES: [H1, H2 per RQ from Step 6]
VARIABLE MAPS: [X, Y, M, C, scope conditions from Step 5]
LITERATURE SCAN: [per-angle tables + novelty threat ratings from Step 3]
DATA INVENTORY: [dataset tables from Step 7]
TARGET JOURNAL: [journal name or "not yet determined"]
Agent 1 — Theorist
Spawn a general-purpose agent with:
"You are a senior sociological theorist evaluating early-stage research questions. For each RQ, evaluate:
Theoretical contribution (Strong / Adequate / Weak): Does this adjudicate competing explanations, bridge disconnected literatures, or identify a new mechanism? Or does it merely replicate existing frameworks?
Mechanism specificity (Strong / Adequate / Weak): Is the causal or explanatory mechanism named, explicit, and traceable step-by-step? Or is it implied, vague, or conflated with the outcome?
Conceptual novelty (Strong / Adequate / Weak): Does the framing offer a genuinely new way to think about the phenomenon, or is it a standard application of known theory to a new context?
For each RQ, provide:
Ratings on the 3 dimensions above
2–3 specific comments explaining your ratings
1 concrete suggestion to strengthen the theoretical contribution (e.g., 'Reframe as a scope-condition test of [theory] rather than a direct test' or 'Add a competing mechanism from [literature]')
End with a rank ordering of the RQs from strongest to weakest on theoretical grounds.
[INPUT PACKAGE]"
Agent 2 — Methodologist
Spawn a general-purpose agent with:
"You are a quantitative methodologist and research design expert evaluating early-stage research questions. For each RQ, evaluate:
Identification strength (Strong / Adequate / Weak): Given the proposed data, what is the strongest plausible identification strategy (experiment, quasi-experiment, selection-on-observables, descriptive)? Does the RQ overstate causal claims relative to what the data can support?
Data-design fit (Strong / Adequate / Weak): Do the proposed datasets actually contain the key X, Y, and M variables? Is the sample large enough for the proposed subgroup analyses? Are there temporal or geographic mismatches?
Measurement validity (Strong / Adequate / Weak): Are the proposed operationalizations of key concepts valid, or do they rely on weak proxies? Flag any construct-measurement gaps.
For each RQ, provide:
Ratings on the 3 dimensions above
2–3 specific comments (name the exact dataset and variable when possible)
1 concrete suggestion to strengthen identification or measurement (e.g., 'Use [dataset] instead because it has [variable]' or 'Add [instrument/design feature] for causal leverage')
End with a rank ordering of the RQs from strongest to weakest on methodological grounds.
[INPUT PACKAGE]"
Agent 3 — Domain Expert
Spawn a general-purpose agent with:
"You are a specialist in [infer domain from topic: e.g., stratification, demography, political sociology, sociolinguistics, migration, health, organizations] evaluating early-stage research questions. For each RQ, evaluate:
Literature gap accuracy (Strong / Adequate / Weak): Does the claimed gap actually exist? Has the literature scan missed key papers, recent working papers, or adjacent-field work that addresses this question? Name any missing citations.
Subfield positioning (Strong / Adequate / Weak): Where does this question sit in the field's current debates? Is it timely, or has the conversation moved on? Is there a live controversy this could speak to?
Novelty claim validity (Strong / Adequate / Weak): Is the novelty real (new mechanism, new population, new data) or superficial (same question, slightly different sample)?
For each RQ, provide:
Ratings on the 3 dimensions above
2–3 specific comments citing papers from the field
1 concrete suggestion to sharpen novelty (e.g., 'The real gap is [X], not [Y]' or 'Reposition as contributing to the [specific debate] rather than the [general topic]')
End with a rank ordering of the RQs from strongest to weakest on domain-specific grounds.
[INPUT PACKAGE]"
Agent 4 — Journal Editor
Spawn a general-purpose agent with:
"You are a former associate editor at a top social science journal (ASR, AJS, Demography, or Science Advances) evaluating early-stage research questions for publication potential. For each RQ, evaluate:
Publication fit (Strong / Adequate / Weak): Does this question match the scope, audience, and norms of the target journal? If no target journal is specified, which 2–3 journals would be the best fit and why?
Contribution framing (Strong / Adequate / Weak): Is the contribution clearly articulable in one sentence? Would it survive the 'so what?' test from a skeptical editor? Can you state what readers will learn that they did not know before?
Broad appeal (Strong / Adequate / Weak): Will this question interest readers beyond the immediate subfield? Does it connect to larger social science themes (inequality, institutions, culture, markets, politics)?
For each RQ, provide:
Ratings on the 3 dimensions above
2–3 specific comments from an editorial perspective
1 concrete suggestion to improve publishability (e.g., 'Lead with the [policy/empirical] puzzle, not the theory gap' or 'This is a Demography paper, not an ASR paper — reframe accordingly')
Suggested target journal(s) if not specified
End with a rank ordering of the RQs from most to least publishable.
[INPUT PACKAGE]"
Agent 5 — Devil's Advocate
Spawn a general-purpose agent with:
"You are a skeptical, rigorous critic whose job is to stress-test research questions before significant time is invested. Your goal is to find fatal flaws, hidden assumptions, and blind spots that the other evaluators may overlook. For each RQ, evaluate:
Assumption audit: What unstated assumptions does this RQ rely on? Which of these are most likely to be violated? (e.g., 'Assumes X is exogenous, but [specific confounder] is likely endogenous')
Null result risk: What is the probability that the main hypothesis is simply wrong or that the effect is too small to detect with the proposed data? What would a null finding mean — is it still publishable?
Competitor threat: Could a well-resourced research team with better data or a natural experiment scoop this question before the study is completed? Is there a working paper or pre-print that already answers it?
Ethical or practical landmines: Are there IRB issues, data access barriers, or politically sensitive framings that could derail the project?
For each RQ, provide:
The single most serious threat to the project's success
1–2 additional risks worth monitoring
1 concrete suggestion to mitigate the top threat (e.g., 'Add [robustness check] to address [confounder]' or 'Collect [supplementary data] as insurance against null finding')
A viability rating: VIABLE (proceed with caution) / AT RISK (needs significant revision) / FATAL FLAW (abandon or fundamentally rethink)
Be constructive but honest. Do not soften your assessment.
[INPUT PACKAGE]"
8b. Synthesize Into Consensus Scorecard
After all 5 agents return, produce a Consensus Scorecard that aggregates their evaluations:
If 3+ agents rank the same RQ first → clear consensus
If rankings diverge → note the disagreement and explain why each agent favors a different RQ (this itself is useful diagnostic information)
Devil's Advocate viability overrides: any RQ rated FATAL FLAW by A5 is automatically downgraded regardless of other ratings
8c. Refine RQs Based on Panel Feedback
For each RQ rated PROCEED or REVISE:
Apply all ★★ suggestions (cross-agent agreement items) — these are high-confidence improvements
Apply the Devil's Advocate mitigation — address the top threat
Revise the RQ text to incorporate panel feedback (mark changes as [REFINED: reason])
Update hypotheses if the refinement changed the mechanism or scope
Update the data recommendation if the Methodologist suggested a better dataset
Present the original and refined versions side-by-side for each RQ:
RQ1 (original): [original text]
RQ1 (refined): [refined text] [REFINED: incorporated A1 suggestion to add competing mechanism + A2 suggestion to use PSID instead of GSS]
H1 (original): [original]
H1 (refined): [refined] [REFINED: downgraded from causal to associational per A2]
8d. Final Verdict
After refinement, produce the final verdict for each RQ using the post-refinement assessment:
Dimension
Criterion
Rating
Theoretical contribution
Adjudicates competing explanations or bridges literatures
H/M/L
Empirical novelty
New data, context, population, or measurement vs. prior work
H/M/L
Identification strength
Causal leverage beyond prior studies
H/M/L
Data feasibility
Key variables available in accessible dataset
H/M/L
Publication fit
Matches norms of target journal
H/M/L
Panel consensus
Multi-agent agreement on viability
Strong/Mixed/Weak
Overall verdict per RQ:
PROCEED: ≥4 dimensions rated High or Medium AND no FATAL FLAW → strong candidate
REVISE: 2-3 dimensions rated Low OR panel consensus is Mixed → name the specific revision still needed
ABANDON: ≥3 dimensions rated Low OR FATAL FLAW from Devil's Advocate → pivot angle
Flag the top risk and mitigation strategy for each PROCEED RQ.
Step 9: Recommend Next Skill Path
State the single recommended RQ (use the refined version from Step 8c) with explicit justification referencing:
The gap identified in Step 3
The best dataset identified in Step 7
The panel consensus from Step 8b (cite which agents agreed and why)
The refinements applied in Step 8c
Then provide the exact next-step command chain:
/scholar-lit-review [selected RQ + keywords] — deep systematic review
/scholar-hypothesis [selected RQ] — theory and hypothesis development
/scholar-design [selected RQ + dataset] — methods and identification strategy
/scholar-causal [selected RQ] — causal diagram, identification strategy, and confounders
Step 10: Save Output to File
After displaying the full output to the user, save the complete output to a Markdown file using the Write tool.
Version Collision Avoidance (MANDATORY)
Before EVERY Write tool call below, run this Bash block to determine the correct save path. Do NOT hardcode paths from the filename templates — they show naming patterns only.
# MANDATORY: Replace [values] with actuals before running# BASE pattern: scholar-idea-[topic-slug]-[YYYY-MM-DD]# Split into directory and stem for the gate script:
OUTDIR="$(dirname "scholar-idea-[topic-slug]-[YYYY-MM-DD]")"
STEM="$(basename "scholar-idea-[topic-slug]-[YYYY-MM-DD]")"mkdir -p "$OUTDIR"
bash "${SCHOLAR_SKILL_DIR:-.}/scripts/gates/version-check.sh""$OUTDIR""$STEM"
Use the printed SAVE_PATH as file_path in the Write tool call. Re-run this block (with the appropriate BASE) for each additional file. The same version suffix must be used for all related output files (.md, .docx, .tex, .pdf).
Derive [topic-slug] from the first 4-6 significant words of the user's topic, lowercased, spaces replaced with hyphens (e.g., redlining-activity-space-segregation)
Use today's date for [YYYY-MM-DD]
Save to the current working directory
File header to prepend:
# Scholar Idea Exploration: [original topic as provided by user]
*Generated by /scholar-idea on [YYYY-MM-DD]*
---
Then write the full output (all 12 sections) exactly as displayed on screen.
Close Process Log:
Run the following to finalize the process log:
OUTPUT_ROOT="${OUTPUT_ROOT:-output}"
SKILL_NAME="scholar-idea"
LOG_DATE=$(date +%Y-%m-%d)
LOG_FILE="${OUTPUT_ROOT}/logs/process-log-${SKILL_NAME}-${LOG_DATE}.md"if [ ! -f "$LOG_FILE" ]; then
LOG_FILE=$(ls -t "${OUTPUT_ROOT}"/logs/process-log-${SKILL_NAME}-${LOG_DATE}*.md 2>/dev/null | head -1)
ficat >> "$LOG_FILE" << LOGFOOTER
## Output Files
[list each output file path as a bullet]
## Summary
- **Steps completed**: [N completed]/[N total]
- **Files produced**: [count]
- **Errors**: [count, or 0]
- **Time finished**: $(date +%H:%M:%S)
LOGFOOTERecho"Process log saved to $LOG_FILE"
REFINED RESEARCH QUESTIONS — original vs. refined RQs side-by-side with [REFINED: reason] markers (Step 8c)
FINAL VERDICT — post-refinement rating table + PROCEED/REVISE/ABANDON per RQ (Step 8d)
RECOMMENDED QUESTION — single best option with justification citing panel consensus
NEXT COMMANDS — exact skill invocations with arguments pre-filled
(file save confirmation line) — Output saved to scholar-idea-[slug]-[date].md
Quality Rules
Before finalizing, verify:
Literature scan followed tiered protocol: local library (Tier 1) searched FIRST for all angles, then external APIs (Tier 2), then WebSearch (Tier 3) only for gaps — do not skip to WebSearch without checking local library
Novelty claims cite specific papers identified in Step 3, not generic statements like "this is understudied"
Data inventory identifies at least one publicly accessible dataset for each PROCEED RQ
Questions are specific enough to operationalize with the identified data
Mechanisms and confounders are explicit
Claims do not exceed plausible design strength of the recommended dataset
The recommended RQ has a clear gap, a named dataset, and a concrete next-step path
Multi-agent panel ran: all 5 evaluator agents (Theorist, Methodologist, Domain Expert, Journal Editor, Devil's Advocate) were spawned in parallel via Task tool
RQs refined: all ★★ suggestions and Devil's Advocate mitigations applied; original vs. refined versions shown side-by-side
No FATAL FLAW RQ recommended: any RQ rated FATAL FLAW by the Devil's Advocate was not selected as the recommended question
Panel consensus cited in recommendation: Step 9 justification references specific agent ratings and cross-agent agreement, not just the literature scan
Output has been saved to a .md file in the current working directory using the Write tool