Workflow improvements: leverage new Claude Code capabilities

[evansenter]

Workflow Improvement Suggestions

Audit of current workflows against recently added Claude Code features. Organized by impact.

---

High Impact

1. Add PostCompact hook to complement PreCompact

Current: pre-compact.sh checkpoints WIP state before compaction. But after compaction, the session resumes with a compressed context and no confirmation that the checkpoint was restored correctly.

Improvement: A PostCompact hook could:

• Verify the WIP checkpoint was included in the compacted context
• Re-inject critical state if the compaction dropped it (e.g., current todo list, work ID)
• Publish a context_compacted event so other sessions know this one lost detail

Effort: Small — mirrors pre-compact.sh pattern.

---

2. Add WorktreeCreate / WorktreeRemove hooks for event bus integration

Current: /parallel-work start manually broadcasts parallel_work_started. But when Claude's built-in worktree system creates/removes worktrees (via --worktree flag or agent isolation), there's no event bus registration.

Improvement:

• WorktreeCreate hook → auto-register worktree session on event bus, publish parallel_work_started
• WorktreeRemove hook → auto-unregister, publish parallel_work_completed
• This means any worktree (not just /parallel-work) gets event bus coordination for free

Effort: Medium — need to handle the case where worktree sessions are agent-spawned (short-lived) vs user-initiated (long-lived).

---

3. Use /effort levels in workflows

Current: All work runs at the model's default effort level regardless of task complexity.

Improvement: Integrate effort levels into /work phases:

• Scope/explore phase: effort medium (reading code, understanding structure)
• Architecture phase: effort high or max (design decisions, trade-offs)
• Implementation: effort medium (straightforward coding)
• Review/reflection: effort high (catching subtle bugs, security issues)

Could also add effort hints to /pr-review remote — reviewer feedback that disagrees with your opinion deserves effort max analysis.

Effort: Small — add effort level switches at phase transitions in work.md.

---

4. Use /rewind as a safety net in /work

Current: If you go down a wrong path during implementation, the only recovery is manual git stash / git checkout and re-explaining context.

Improvement: /work could create named checkpoints at phase boundaries:

• Before architecture decisions
• Before implementation starts
• Before each major file change

Then suggest /rewind when the user says "that's wrong" or "go back" instead of trying to undo changes manually.

Effort: Medium — need to understand /rewind checkpoint API.

---

Medium Impact

5. Use /fork for architecture exploration

Current: /work phase 3 (Architect) launches parallel agents to propose approaches, then presents them sequentially. The user picks one and the others are discarded.

Improvement: Instead of discarding, /fork the conversation at the decision point. This lets the user:

• Pick approach A now
• Come back later and /resume the fork to try approach B
• Compare real implementations, not just proposals

Effort: Small — add a fork before committing to architecture choice.

---

6. /context awareness in /work

Current: No visibility into context pressure during long /work sessions. Compaction happens unexpectedly.

Improvement:

• Check context usage at phase boundaries
• If >70% full before implementation starts, proactively compact with focused instructions
• Show context % in /im-lost output

Effort: Small — add context checks at transitions.

---

7. /security-review as a quality gate

Current: Quality gates are: lint, format, test, /pr-review local. No security-specific check.

Improvement: Add /security-review to the pre-push checklist in /work, especially for:

• Changes touching auth, crypto, user input, API endpoints
• New dependencies
• Config file changes (could leak secrets)

Could be conditional — only run when changed files match security-sensitive patterns.

Effort: Small — add to /pr-review local flow.

---

8. /diff integration in /pr-review local

Current: /pr-review local runs git diff and analyzes it programmatically.

Improvement: Suggest /diff to the user at the end of local review so they can visually inspect per-turn changes before pushing. Especially useful for large changesets where the summary might miss something.

Effort: Trivial — add a suggestion line.

---

Lower Impact (Nice to Have)

9. /stats and /insights in reflection phase

Current: /work reflection phase asks "what was hard?" and publishes to event bus.

Improvement: Pull /insights data during reflection to ground it in actual metrics — how many turns did this take? How much context was used? Were there failed tool calls? This makes reflections more actionable.

Effort: Medium — need to parse insights output.

---

10. Session naming for better /event-bus-status

Current: Sessions show as ./main or directory-based names. When running multiple sessions in the same repo, they're hard to distinguish.

Improvement: /work should set a session name with /rename based on the work ID: e.g., issue-42: Add auth middleware. This makes /event-bus-status immediately readable.

Effort: Trivial — add /rename call at /work start.

---

11. Auto-memory for recurring friction

Current: Reflection publishes gotcha_discovered and pattern_found to event bus, but these are ephemeral — they scroll off the event list.

Improvement: When reflection surfaces a gotcha or pattern that's repo-specific, also save it to Claude's auto-memory. This way it persists across sessions without relying on event bus history.

Effort: Small — add memory write step to reflection.

---

Summary

#	Improvement	Impact	Effort
1	PostCompact hook	High	Small
2	WorktreeCreate/Remove hooks	High	Medium
3	Effort levels in /work	High	Small
4	/rewind checkpoints in /work	High	Medium
5	/fork for architecture	Medium	Small
6	/context awareness	Medium	Small
7	/security-review gate	Medium	Small
8	/diff in pr-review	Medium	Trivial
9	/stats in reflection	Low	Medium
10	Session naming	Low	Trivial
11	Auto-memory for gotchas	Low	Small