brycewang-stanford/new-project

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.

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

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:

   • Research specification (quality_reports/research_spec_*.md)
   • Domain profile (.claude/references/domain-profile.md) — if still template

2. Run /discover lit with the research topic:

   • Librarian collects literature
   • librarian-critic reviews coverage
   • Output: literature synthesis + BibTeX entries

3. Run /discover data to find datasets:

   • Explorer searches for data sources
   • explorer-critic assesses data quality

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

Step 2: Strategy Phase

4. Run /strategize to design the empirical strategy:
   • Strategist proposes identification strategy
   • strategist-critic validates the design

Gate: Strategy memo must pass strategist-critic review (score >= 80).

Step 3: Execution Phase

5. Run /analyze to implement the strategy:

   • Data-engineer cleans data and creates figures
   • Coder writes analysis scripts
   • coder-critic reviews code

6. Run /write to draft the paper:

   • Writer drafts sections
   • Humanizer pass strips AI patterns

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

Step 4: Peer Review Phase

7. Run /review --all for comprehensive review:

   • strategist-critic + coder-critic + writer-critic + Verifier in parallel
   • Weighted aggregate score computed

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

   • domain-referee (subject expertise) + methods-referee (econometrics)
   • Independent, blind reports
   • Orchestrator synthesizes editorial decision

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

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

9. Run /submit target for journal recommendations
10. Run /submit package for replication package
11. 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.

---

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.