Skip to content

new-project

Category: drafting
Field: —
License: none declared
Updated: 2026-05-11
Stages: paper-drafting

New Project

Launch a full research pipeline from idea to paper, orchestrated through the dependency graph.

Input: $ARGUMENTS — a research topic or interactive for a guided start via /discover interview.

Pipeline Overview

This skill orchestrates the full dependency graph. Each phase activates when its dependencies are met. The orchestrator manages agent dispatch, three-strikes escalation, and quality gates.

Text Only
Phase 1: Discovery
  ├── /discover interview → Research Spec + Domain Profile
  ├── /discover lit → Literature Synthesis + BibTeX
  └── /discover data → Data Assessment

Phase 2: Strategy (depends on Phase 1)
  ├── /strategize → Strategy Memo + Robustness Plan
  └── /strategize theory → Theory Section (conditional — econometric methods, theory+empirics, structural, methodological reduced-form)

Phase 3: Execution (depends on Phase 2)
  ├── /analyze → Scripts + Tables + Figures
  └── /write → Paper Sections

Phase 4: Peer Review (depends on Phase 3)
  ├── /review --all → Comprehensive Quality Score
  └── /review --peer → domain-referee + methods-referee Reports

Phase 5: Submission (depends on Phase 4, score >= 95)
  ├── /submit target → Journal Recommendations
  ├── /submit package → Replication Package
  └── /submit final → Final Verification

Workflow

Step 0: Enter Plan Mode

Before any work begins: 1. Enter plan mode — use EnterPlanMode 2. Create the project folder structure — data/raw/, data/cleaned/, scripts/R/, paper/sections/, paper/figures/, paper/tables/, etc. 3. Draft a high-level plan — what phases are needed, estimated scope 4. Save to disk — quality_reports/plans/YYYY-MM-DD_new-project.md 5. Present to user — wait for approval before proceeding 6. Exit plan mode — only after approval

Step 1: Discovery Phase

  1. If interactive or no research spec exists: Run /discover interview to produce:
  2. Research specification (quality_reports/research_spec_*.md)

  3. Domain profile (.claude/references/domain-profile.md) — if still template

  4. Run /discover lit with the research topic:

  5. Librarian collects literature
  6. librarian-critic reviews coverage

  7. Output: literature synthesis + BibTeX entries

  8. Run /discover data to find datasets:

  9. Explorer searches for data sources
  10. explorer-critic assesses data quality

Gate: Research spec and literature review must exist before proceeding.

Step 2: Strategy Phase

  1. Run /strategize to design the empirical strategy:
  2. Strategist proposes identification strategy
  3. strategist-critic validates the design

4b. If paper type is econometric methods, theory+empirics, structural, or methodological reduced-form: Run /strategize theory to produce the formal theory section: - Theorist drafts assumptions, theorems, proofs - theorist-critic audits proof validity (4 phases, early-stop on critical gaps) - Theorist-critic score contributes 20% to the weighted aggregate when present (see quality.md)

Skip this step for applied papers using off-the-shelf estimators.

Gate: Strategy memo must pass strategist-critic review (score >= 80). If theory section exists, theorist-critic must also pass (score >= 80).

Step 3: Execution Phase

  1. Run /analyze to implement the strategy:
  2. Data-engineer cleans data and creates figures
  3. Coder writes analysis scripts

  4. coder-critic reviews code

  5. Run /write to draft the paper:

  6. Writer drafts sections
  7. Humanizer pass strips AI patterns

Gate: Code must pass coder-critic review. Paper sections must exist.

Step 4: Peer Review Phase

  1. Run /review --all for comprehensive review:
  2. strategist-critic + coder-critic + writer-critic + Verifier in parallel

  3. Weighted aggregate score computed

  4. Run /review --peer for simulated peer review:

  5. domain-referee (subject expertise) + methods-referee (econometrics)
  6. Independent, blind reports
  7. Orchestrator synthesizes editorial decision

Gate: Aggregate score >= 80 (commit-ready). Score >= 90 for submission.

Step 5: Submission Phase (optional, user-triggered)

  1. Run /submit target for journal recommendations
  2. Run /submit package for replication package
  3. Run /submit final for final verification

User Interaction Points

The pipeline pauses for user input at these points: - After interview (approve research spec) - After strategy memo (approve identification strategy) - After data analysis (review results before paper drafting) - After peer review (review feedback before revision) - Before submission (approve journal choice)

Between pauses, the orchestrator runs autonomously per workflow.md.

Bundled Resources (Level 3)

Resource Path When
Quality gates config/quality-gates.json Phase transitions — score thresholds
Gotchas gotchas.md Always — known failure points

Principles

  • This is always orchestrated. Unlike other skills, /new-project always runs through the full pipeline.
  • Dependency-driven. Phases activate by dependency, not forced sequence.
  • Quality-gated. Each phase transition requires passing quality checks.
  • User retains control. Pipeline pauses at key decision points.
  • Resumable. If interrupted, the pipeline resumes from the last completed phase.