MemStack™ v3.5.4: Skill Framework for Claude Code

You are running with MemStack™ enabled. Skills use the official Anthropic SKILL.md format: each skill lives in skills/{name}/SKILL.md with YAML frontmatter (name + description). Hooks in .claude/hooks/ fire deterministically on CC lifecycle events. Rules in .claude/rules/ are always loaded at session start.

Current skill count: 131 total (87 free + 44 Pro-exclusive). Verify it live with memstack-skill-loader/scripts/check_skill_drift.py rather than trusting this line.

v3.5.4 changes: Documentation alignment pass: corrected skill counts to 128 total (85 free + 43 Pro-exclusive) as of that release, added the free git-guard skill, rewrote ADDING-SKILLS.md as the canonical skill-change checklist (16 count locations, 3-channel update path), and added automated skill-count drift enforcement. TokenStack™ is now the sole context-compression proxy. New skills still default to Pro-exclusive, graduating to free after 90 days.

Branch Strategy

• All new work happens on the dev branch. Never commit directly to master.
• Merge to master only after Reviewer confirms working.
• Merge commit format: merge: [description] dev->master
• Setup: git checkout -b dev && git push -u origin dev
• Merge: git checkout master && git merge dev && git push && git checkout dev

Global Rules

See .claude/rules/memstack.md for the full rule set. Summary:

1. Read the project's CLAUDE.md first if one exists
2. Commit format: [ProjectName] description or type(scope): description, with no Co-Authored-By trailer
3. Always build before push (enforced by hook)
4. Document decisions in CLAUDE.md
5. Skill chain: Work → Seal (hook) → Diary → Monitor (hook)

Architecture (v3.2.1)

MemStack™ v3.2.1 uses three layers:

Layer	What	How	Examples
Hooks	Deterministic safety gates	Shell scripts fired by CC lifecycle events	Seal (pre-push), Deploy (post-commit), Monitor + TokenStack™ + CLAUDE.md indexer (session start/end)
Rules	Always-on behavioral guidance	Markdown files loaded every session	Echo recall, Diary logging, Work planning, global conventions
Skills	Context-aware workflows	skills/{name}/SKILL.md: official Anthropic format	Echo, Diary, Work, Project, Scan, Quill, Forge, Sight, Shard

Hooks always fire: deterministic. Rules always load: persistent behavioral layer. Skills fire when CC detects matching triggers.

Hook Configuration

Hooks are wired in .claude/settings.json:

Hook Script	CC Event	Behavior
pre-push.sh	PreToolUse (git push)	Build check, secrets scan, commit format: blocks push on failure
post-commit.sh	PostToolUse (git commit)	Debug artifact scan, secrets check: warns after commit
session-start.sh	SessionStart	TokenStack™ auto-start + CLAUDE.md auto-index + reports "working" to API
session-end.sh	Stop	Reports "completed" status to monitoring API

Rules Configuration

Rules in .claude/rules/ are loaded automatically every session:

Rule File	Skill Enhanced	Behavior
memstack.md	Global	Commit format (standard + conventional), build safety, no secrets, deprecated skill guard
echo.md	Echo (Lv.5)	Always-on memory recall protocol, vector search first, SQLite fallback
diary.md	Diary (Lv.5)	Always-on session logging awareness: log with structured handoff
work.md	Work (Lv.5)	Always-on task planning protocol, activate on plan/todo/task
tokenstack.md	TokenStack™	Compression proxy awareness: troubleshooting, stats check

Slash Commands

Command	File	Behavior
/memstack-search <query>	.claude/commands/memstack-search.md	Quick memory search, runs memstack-db.py search
/memstack-headroom	.claude/commands/memstack-headroom.md	TokenStack™ proxy status and token savings

Hook Exit Codes

Code	Meaning
0	Success: continue
1	Error (logged, continues)
2	Block the operation

Trigger Types

• Keyword: fires when specific phrases appear in prompt
• Passive: always-on background behavior (now hooks in v3.0)
• Contextual: fires when conditions are detected (file size, session state)

Skill Index

#	Skill	Emoji	Type	Level	Function	Key Triggers
1	Familiar	👻	Keyword	Lv.2	Multi-agent dispatch	"dispatch", "send familiar", "split task"
2	Echo	🔊	Keyword	Lv.5	Semantic memory recall (LanceDB + SQLite)	"recall", "last session", "do you remember" + rule
3	Seal	🔒	Passive	Hook	Git commit guardian →.claude/hooks/pre-push.sh	Deterministic on git push
4	Work	📋	Keyword	Lv.5	Plan execution (copy/append/resume)	"copy plan", "append plan", "resume plan", "todo" + rule
5	Project	💾	Contextual	Lv.3	Session handoff & lifecycle	"save project", "handoff", "context running low"
6	Grimoire	📖	Keyword	Lv.2	CLAUDE.md management	"update context", "update claude", "save library"
7	Scan	🔍	Keyword	Lv.2	Project analysis & pricing	"scan project", "estimate", "how much to charge"
8	Quill	✒️	Keyword	Lv.2	Client quotation generator	"create quotation", "generate quote", "proposal"
9	Forge	🔨	Keyword	Lv.2	New skill creation	"forge this", "new skill", "create enchantment"
10	Diary	📓	Contextual	Lv.5	Session documentation + structured handoff	"save diary", "log session", end of session + rule
11	Shard	💎	Contextual	Lv.2	Large file refactoring (1000+ LOC)	"shard this", "split file", files over 1K lines
12	Sight	👁️	Keyword	Lv.2	Architecture visualization	"draw", "diagram", "visualize", "architecture"
13	Monitor	📡	Passive	Hook	CC Monitor self-reporting →.claude/hooks/session-*.sh	Deterministic on session start/end
14	Deploy	🚀	Passive	Hook	Build & deployment guardian →.claude/hooks/post-commit.sh	Deterministic on git commit
16	Humanize	✍️	Keyword	Lv.1	Remove AI writing patterns from text	"humanize", "make it sound natural", "clean up writing"
17	State	📍	Contextual	Lv.1	Living STATE.md: current task/blockers/next steps	"update state", "project state", "where was I"
18	Verify	✅	Keyword	Lv.1	Pre-commit work verification report	"verify", "check this work", "does it pass"
19	Governor	🏛️	Contextual	Lv.1	Portfolio governance (tier/phase constraints)	"new project", "what tier", "scope", "project init"
20	Compress	⚙️	Keyword	Lv.2	TokenStack™ proxy management & stats	"tokenstack", "compression", "token savings", "proxy status"

Skill Deconfliction

When multiple skills could activate on the same prompt, use these ownership rules:

• "commit" →post-commit hook fires automatically
• "push" / "ship it" / "deploy" →pre-push hook blocks if checks fail
• "build" →Neither, just run the build command directly
• "recall" / "remember" →Echo only (not Diary or Project)
• "save diary" / "log session" →Diary only (not Project)
• "save project" / "handoff" →Project only (not Diary)
• "todo" / "plan" →Work only
• "/memstack-search" →Slash command (quick search, no full Echo activation)
• "where was I" / "update state" →State only (not Echo or Project)
• "verify" / "check this work" →Verify only (not Seal hook)
• "humanize" / "rewrite" →Humanize only
• "tier" / "scope" / "what's allowed" →Governor only

Storage

• Database (primary): db/memstack.db: SQLite with WAL mode
• DB Helper: python db/memstack-db.py <command>: repository pattern CLI
• Commands: init, add-session, add-insight, search, get-sessions, get-insights, get-context, set-context, add-plan-task, get-plan, update-task, export-md, stats

Paths

• Skills: C:\Projects\memstack\skills\{name}\SKILL.md | Deprecated: skills\_deprecated\ | Hooks: .claude/hooks/ | Rules: .claude/rules/ | Commands: .claude/commands/ | DB: C:\Projects\memstack\db\ | Config: config.json

Architecture inspired by Developer Kaki's MemoryCore (github.com/Kiyoraka/Project-AI-MemoryCore)