Claude Code workflow improvements from gemicro session

[evansenter]

Summary

Improvements identified during a gemicro development session (PRs #135, #136, issue #137 coordination).

Global Improvements

1. Add make and cargo permissions to settings.json

What it enables: Auto-approve common build commands across all projects with Makefiles or Rust codebases.

Change to ~/.claude/settings.json:

"Bash(make:*)",
"Bash(cargo check:*)",
"Bash(cargo build:*)",
"Bash(cargo test:*)",
"Bash(cargo clippy:*)",
"Bash(cargo fmt:*)",
"Bash(cargo run:*)",
"Bash(cargo doc:*)"

Why: These are safe, read-only or local-only operations that don't need human approval.

---

2. Add /rfc-response command

What it enables: Structured responses to RFC-style issues with:

• Context/learnings section
• Assumptions table (Assumption | Confidence | Impact if Wrong)
• Questions with rationale and proposed solutions
• Requirements needing human input (separate section)
• Actionable requirements table (# | Requirement | Owner | Blocked By)

Why: Issue #137 on gemicro showed the value of structured RFC responses. A command ensures consistency across projects.

Create ~/.claude/commands/rfc-response.md:

Generate a structured RFC response for the given issue.

## Structure

1. **Context/Learnings** - What we learned that's relevant
2. **Assumptions** - Table with Assumption | Confidence | Impact if Wrong
3. **Questions** - Each with:
   - Why it matters
   - Proposed solutions (in preference order)
   - What we need confirmed
4. **Requirements Needing Human Input** - Decisions that block progress
5. **Actionable Requirements** - Table with # | Requirement | Owner | Blocked By

## Instructions

- Read the issue carefully
- Reference specific code/PRs where relevant
- Propose solutions, don't just ask open-ended questions
- Identify blockers explicitly

$ARGUMENTS

---

3. Add TodoWrite guidance to CLAUDE.md

Current gap: The global CLAUDE.md doesn't mention when to use TodoWrite.

Suggested addition:

## Task Tracking

Use TodoWrite for:
- Multi-phase implementations (like gemicro issue #84's phases)
- Tasks with 3+ distinct steps
- When resuming from context limits (helps re-orient)

Don't use TodoWrite for:
- Single-step tasks
- Pure research/exploration
- Conversational exchanges

---

4. Clarify self-review timing in CLAUDE.md

Current text:

After completing implementation work: summarize changes and run /pr-feedback --local

Issue: Ambiguous about whether this is before or after pushing.

Suggested rewrite:

- After completing implementation work: summarize what was done, then run `/pr-feedback --local` for self-review before pushing

---

Checklist

• Add make and cargo permissions to ~/.claude/settings.json
• Create ~/.claude/commands/rfc-response.md
• Add TodoWrite guidance to ~/.claude/CLAUDE.md
• Clarify self-review timing in ~/.claude/CLAUDE.md