Agent Skills: Claude Code Session Priming

Claude Code Session Priming

You have several powerful plugins and tools installed. Follow these protocols throughout the session to make full use of them.

1. Claude-mem (Persistent Memory)

Claude-mem provides semantic memory across sessions via MCP tools. A context index is delivered automatically at session start in a system reminder.

Protocol: Search Before Re-Investigating

Before reading files or exploring code to understand something, check memory first. Past sessions likely already recorded the answer.

1. search(query) -> scan the index for relevant observation IDs
2. timeline(anchor=ID) -> get surrounding context
3. get_observations([IDs]) -> fetch full details only for filtered IDs

Never fetch full details without filtering first. The 3-layer workflow provides 10x token savings.

Protocol: Save After Significant Work

After completing any of the following, call save_memory to record it:

• Discoveries: codebase structure, how a system works, where key code lives
• Decisions: architectural choices, approach trade-offs, why option A over B
• Completed work: what was built/changed, the final state, key details
• Bug findings: root cause, fix applied, symptoms vs actual problem
• Learnings: gotchas, undocumented behavior, things that surprised you

Write memory entries as self-contained observations. Future sessions will see the title and token cost in the context index, then decide whether to fetch the full record. A good title and enough detail to be useful standalone are key.

Protocol: Use the Context Index

The session-start context index shows past observations with:

• ID, timestamp, type (bugfix/feature/decision/discovery/etc.)
• Title, token cost to read, tokens of work that produced it
• File associations

Trust this index for past decisions and learnings. Only fetch full observations when you need implementation details, rationale, or debugging context. Critical types (bugfix, decision) often merit detailed fetching.

Skills: /claude-mem:make-plan and /claude-mem:do

These skills create implementation plans with documentation discovery and execute plans using subagents. Use them for structured multi-step work.

2. git-ai-search (Conversation Context from Git)

git-ai tracks AI-generated code and the conversations that produced it.

When to Use

• Resuming work on a git repo: Search for AI context on recent commits to understand what was done and why
• Investigating unfamiliar code: Check if AI sessions contributed to specific files or line ranges
• Picking up a teammate's work: Restore their conversation context
• PR reviews: Understand AI involvement in changes

Key Commands

git-ai search --commit <sha>              # AI context for a commit
git-ai search --file <path> --lines 50-75 # AI context for specific lines
git-ai search --pattern "keyword"         # Search prompt content
git-ai continue --commit <sha>            # Restore session context

Use /git-ai-search to invoke the full skill when deeper investigation is needed.

3. Cozempic (Context Weight Management)

Cozempic prevents context bloat, which causes degraded performance and lost state (especially agent teams).

Automatic Protection

The Cozempic guard daemon starts automatically at session init. It monitors session size and can auto-prune before compaction kills agent teams.

When to Use Proactively

• Long sessions: When you've been working for a while and context feels heavy, run /cozempic diagnose to check
• Before agent teams: Ensure guard mode is active before spawning teams with TeamCreate. Agent team state is lost when auto-compaction triggers.
• After large file reads: If you've read many large files, context may be bloated with stale content

Quick Reference

| Situation | Action | |-----------|--------| | Check session size | cozempic current | | Diagnose bloat | /cozempic diagnose | | Prune and reload | /cozempic treat | | Protect agent teams | Guard daemon (auto-started) |

Prescriptions

• gentle (under 5MB): progress collapse, file dedup, metadata strip
• standard (5-20MB): + thinking blocks, tool trim, stale reads
• aggressive (over 20MB): + error collapse, document dedup, mega-block trim

4. PAL MCP (Multi-Model Collaboration)

PAL provides access to external models for second opinions, deep analysis, and consensus building.

When to Use

• Complex debugging: mcp__pal__debug for systematic root cause analysis
• Architecture decisions: mcp__pal__consensus to consult multiple models
• Code review: mcp__pal__codereview for structured review with expert validation
• Before commits: mcp__pal__precommit to validate changes
• Deep analysis: mcp__pal__thinkdeep for multi-step investigation

Protocol: Choose the Right Tool

| Need | PAL Tool | |------|----------| | Second opinion on approach | chat | | Systematic debugging | debug | | Architecture/code analysis | analyze | | Multi-model decision making | consensus | | Code review | codereview | | Pre-commit validation | precommit | | Security audit | secaudit | | Refactoring opportunities | refactor | | Test generation | testgen |

5. Session Workflow Summary

At Session Start

1. Read the Claude-mem context index (delivered automatically)
2. If resuming work in a git repo, consider git-ai search on recent commits
3. Search Claude-mem for relevant past work before starting new investigation

During Work

1. Search memory before re-reading files or re-exploring code
2. Save significant findings, decisions, and completions to memory
3. Use PAL tools for complex analysis, debugging, and decisions
4. Monitor context health; use Cozempic if sessions run long

Before Agent Teams

1. Verify Cozempic guard is running (check session-start logs)
2. If not running: cozempic guard --threshold 50 -rx standard --interval 30

At Session End

1. Save any unsaved important findings to Claude-mem
2. For git repos, work will be captured by git-ai automatically on commit