Context Optimizer Skill

Ghost

Run lean. Stay sharp. Go longer.

Ghost is a context optimization skill for Claude that eliminates token waste without touching model capability. Sessions last longer. Outputs stay sharp. Nothing is compromised.

Two versions. One repo

---

Versions

Version	For	How it works
Ghost.md	Claude.ai (browser/app)	Upload via + → Skills. Runs silently in every chat.
context-optimizer-skill/SKILL.md	Claude Code (terminal)	Drop in .claude/skills/. Uses file-based memory system.

---

The Problem

Claude sessions die fast — not because the model is weak, but because context fills up with garbage:

• Full conversation history replaying every single turn
• File reads sitting idle after they've been acted on
• Repeated project context that should live in one file
• One mega-session bleeding across 6 unrelated tasks
• Connectors and tools enabled all session for a step that took 2 minutes

Result: bloated context → slower reasoning → session limits in 30–60 minutes.

Ghost fixes all of it.

---

Version 1 — Claude.ai (Ghost.md)

Install: + → Skills → upload Ghost.md

Ghost activates automatically in every Claude.ai conversation and:

• Scopes tasks before acting — loads only what's needed, ignores everything else
• Compresses silently when history grows — keeps decisions, drops noise
• Switches context cleanly when you change topics — no cross-task residue
• Disciplines tool use — no unnecessary web searches or raw output carryover
• Answers exactly what's asked — no padding, no re-explaining

Commands

Command	What happens
/scope	Restate task in 3 bullets, identify exact info needed
/compress	Summarize session — decisions, state, next step. Drop noise.
/snapshot	Generate a copyable block to resume this session later
/resume	Paste a prior snapshot — Ghost picks up exactly where you left off
/clear	Acknowledge topic switch, drop prior task context
/audit	Show what's in context, what can be dropped

Session Snapshot

The /snapshot command generates a resume block you can copy and paste into a new session:

GHOST SESSION SNAPSHOT
─────────────────────
Task: [current goal]
Done: [completed items]
Open: [unresolved items]
Key decisions: [list]
Next step: [exact action]
─────────────────────

Paste it with /resume in any new chat. Zero context loss across sessions.

---

Version 2 — Claude Code (context-optimizer-skill/)

Install: Drop context-optimizer-skill/SKILL.md into .claude/skills/ in your project.

Copy the templates into your project:

your-project/
├── CLAUDE.md                        ← permanent project memory
├── docs/
│   ├── session-handoff.md           ← resume point between sessions
│   ├── decision-log.md              ← architecture decisions, out of chat
│   └── context-map.md               ← module map, navigate without reading files

Ghost enforces a 3-tier memory system:

Tier 1 — CLAUDE.md          Permanent rules. Read once per session. Never repeated in chat.
Tier 2 — docs/              Session state, decisions, module map. Survives /clear.
Tier 3 — Live chat          Temporary scratchpad. Expendable.

Session Flow

START    → Read CLAUDE.md + session-handoff.md → scope task
WORK     → Touch only needed files
DONE     → /compress → update handoff + decision-log
SWITCH   → /handoff → /clear → /resume

Commands

Command	Action
/scope	Restate task, list exact files, ignore everything else
/compress	Summarize history → decisions + state + next step
/handoff	Write full state to session-handoff.md
/resume	Load CLAUDE.md + handoff, continue from next exact step
/audit	Flag what can be cleared or moved to files
/init	Generate all memory files for a new project

Full command prompts in context-optimizer-skill/commands/commands.md.

---

What Changes

Before Ghost	After Ghost
Session limits in 30–60 min	Sessions that actually last
Full project context every prompt	Lives in CLAUDE.md / memory, referenced once
File reads sitting in context	Cleared after use, re-fetched on demand
One session for everything	Task-scoped, clean switches
Vague prompts → broad exploration	Scoped prompts → targeted execution
Cross-task context pollution	Clean slate on topic switch

---

Repo Structure

Ghost/
├── Ghost.md                              ← Claude.ai skill (upload this)
├── README.md
├── LICENSE
└── context-optimizer-skill/             ← Claude Code skill
    ├── SKILL.md                          ← upload/drop this
    ├── commands/
    │   └── commands.md                   ← slash command prompts
    └── templates/
        ├── CLAUDE.md
        ├── session-handoff.md
        ├── decision-log.md
        └── context-map.md

---

Important

Ghost reduces wasted tokens — not useful ones. It does not bypass Anthropic's service limits. What it does: make every token count, so the same budget goes dramatically further.

A clean 20k context beats a bloated 180k context. Every time.

---

License

MIT — free to use, modify, and build on.

---

Built by Nistal — because hitting session limits every 30 minutes gets old fast.

context-optimizer-skill v1.0.0