Workflow Quick Reference

Workflow Quick Reference

**Model:** Contractor (you direct, Claude orchestrates)

---

The Loop

Your instruction
    ↓
[PLAN] (if multi-file or unclear) → Show plan → Your approval
    ↓
[EXECUTE] Implement, verify, done
    ↓
[REPORT] Summary + what's ready
    ↓
Repeat

---

I Ask You When

• Design forks: "Option A (fast) vs. Option B (robust). Which?"
• Code ambiguity: "Spec unclear on X. Assume Y?"
• Replication edge case: "Just missed tolerance. Investigate?"
• Scope question: "Also refactor Y while here, or focus on X?"

---

I Just Execute When

• Code fix is obvious (bug, pattern application)
• Verification (tolerance checks, tests, compilation)
• Documentation (logs, commits)
• Plotting (per established standards)
• Deployment (after you approve, I ship automatically)

---

Quality Gates (No Exceptions)

Score	Action
>= 80	Ready to commit
< 80	Fix blocking issues

---

Non-Negotiables

• Path convention: pathlib.Path for Python; here::here() for R; always relative paths
• Seed convention: random.seed(YYYYMMDD) + numpy.random.seed(YYYYMMDD) at top of every stochastic Python script; set.seed(YYYYMMDD) at top of every stochastic R script
• Figure standards: 300 DPI, white or transparent background, journal-quality axes, colorblind-safe palette
• Color palette: viridis / colorblind R palette; Okabe-Ito palette for Python (matplotlib)
• Tolerance thresholds: integers exact; point estimates ±0.01; SEs ±0.05; p-values same significance bracket; percentages ±0.1pp

---

Preferences

**Visual:** 300 DPI, Okabe-Ito palette, `bbox_inches='tight'` for Python figures **Reporting:** Structured Markdown report with comparison tables (paper value vs. ours vs. diff) **Session logs:** Always (post-plan, incremental, end-of-session) **Replication:** Flag all near-misses (within 2× tolerance); investigate before proceeding to extensions

---

Exploration Mode

For experimental work, use the **Fast-Track** workflow:

• Work in explorations/ folder
• 60/100 quality threshold (vs. 80/100 for production)
• No plan needed — just a research value check (2 min)
• See .claude/rules/exploration-fast-track.md

---

Next Step

You provide task → I plan (if needed) → Your approval → Execute → Done.