compact-guard
Compact Guard preserves critical context before Claude Code's automatic or manual compaction cycles by identifying the five most important files, documenting current task progress, architectural decisions, and blockers. Use this skill immediately when an auto-compact warning appears or before manually invoking the `/compact` command to ensure essential work state survives the context window reduction.
git clone --depth 1 https://github.com/rohitg00/pro-workflow /tmp/compact-guard && cp -r /tmp/compact-guard/skills/compact-guard ~/.claude/skills/compact-guardSKILL.md
# Compact Guard Protect important context through compaction cycles. Based on Claude Code internals: compaction restores max 5 files with 5K tokens each, within a 50K total budget. ## Trigger Use before `/compact` or when auto-compact warning appears. ## Key Constants (from Claude Code source) - `POST_COMPACT_MAX_FILES_TO_RESTORE = 5` — only 5 files survive - `POST_COMPACT_TOKEN_BUDGET = 50K` — total restore budget - `POST_COMPACT_MAX_TOKENS_PER_FILE = 5K` — per-file limit - Auto-compact fires at `context_window - 13K` buffer ## Pre-Compact Checklist Before compacting, save these to memory or a scratch file: 1. **Current task** — What are you working on? One sentence. 2. **Files in progress** — Which files are being edited? (max 5 — compaction only restores 5) 3. **Decisions made** — Any architectural choices made this session 4. **Blockers** — What's preventing progress? 5. **Next steps** — What to do immediately after compact ## Strategy: Microcompact First Before full compaction, try microcompact: - Large tool results (test output, grep results) can be trimmed - File reads that are no longer relevant can be dropped - Use subagents for heavy exploration to keep main context clean ## Post-Compact Recovery After compaction, immediately: 1. Re-read the top-priority file (the one you're actively editing) 2. Check task list for current progress 3. Review any scratch notes saved pre-compact 4. Resume from next steps ## Prevention Strategies | Strategy | Token Savings | When | |----------|--------------|------| | Delegate grep/search to subagent | 30-60% per search | Always for broad searches | | Read only needed lines (`offset`/`limit`) | 50-90% per read | Large files | | Compact at task boundaries | Preserves coherence | Between logical steps | | Use `/resume` for fresh start | 100% | Unrelated new task | ## Output After running compact-guard: ```text COMPACT GUARD Files to preserve: [list top 5] Task state: [one sentence] Decisions: [key choices] Next step: [immediate action after compact] Ready to compact. Run /compact now. ``` ## Rules - Always save state before compacting — never compact blind - Prioritize the file you're actively editing as #1 to restore - If auto-compact fires unexpectedly, immediately re-read your working file - Keep CLAUDE.md under 60 lines to leave room for actual context
Analyzes and optimizes context window usage across sessions. Use when context feels bloated, sessions run slow, or approaching compaction limits.
Analyze session token usage and cost patterns. Identify expensive operations and recommend optimizations. Use to understand and reduce session costs.
Specialized debugging agent. Use when facing hard bugs, test failures, or runtime errors that need systematic investigation.
Multi-phase development agent. Research > Plan > Implement with validation gates. Use PROACTIVELY when building features that touch >5 files or require architecture decisions.
Analyze permission denial patterns and generate optimized alwaysAllow/alwaysDeny rules. Use when permission prompts slow down workflow.
Break down complex tasks into implementation plans before writing code. Use when task touches >5 files, requires architecture decisions, or has unclear requirements.
Code review specialist that verifies every finding against actual code before reporting. Use before committing, for PR reviews, or after major changes.
Confidence-gated exploration that assesses readiness before implementation. Scores 0-100 across five dimensions and gives GO/HOLD verdict.