Context Graph Compressor
Adityapal67/context-graph-compressor
Comprima conversas em grafos JSON portáteis para handoffs entre chats e modelos
Ver no GitHubSkill que comprime conversas longas com IA em grafos de estado portáteis e estruturados em JSON, otimizando handoffs entre sessões e modelos diferentes. Reduz até 97% do consumo de tokens ao transferir contexto, com suporte a múltiplas LLMs como GPT, Gemini e Mistral.
Instalação
npx convez-ai Adityapal67/context-graph-compressor1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Transferir contexto de projeto entre sessões de chat sem reenviar histórico completo
- Compartilhar estado de conversa entre diferentes modelos de IA (Claude, GPT, Gemini)
- Arquivar e reutilizar contexto estruturado com tags de importância (high/medium/low)
- Otimizar custo de tokens em fluxos de trabalho com múltiplas conversas
Veredito

Solução bem pensada para um problema real: compressão de contexto estruturada e portável, com metadados (importância, status, dependências) que outras ferramentas não oferecem. Formato JSON máquina-legível é a principal vantagem sobre markdown/bullets, e o suporte cross-LLM amplia seu alcance.
README
A tradução para o português deste README ainda está na fila.
context-graph-compressor
A Claude skill that compresses any chat conversation into a minimal, portable JSON context graph — so you can continue the same conversation in a new chat or a different LLM at a fraction of the token cost.
The Problem
Every message in a long chat re-sends the entire history:
Turn 50: [turn1 + turn2 + ... + turn49 + your message] → ~15,000 tokens
The Solution
Compress the chat once, paste the output into a new session:
New chat turn 1: [compressed graph ~400 tokens + your message] → done
Same context. 97% fewer tokens.
Why This Is Different
Several context-handling solutions already exist. Here's how this one compares:
| This skill | Claude's built-in compaction | Handoff skills (Smithery / ClaudSkills) | |
|---|---|---|---|
| Works across sessions | ✅ Export & paste anywhere | ❌ Same chat only | ✅ |
| Works with other LLMs | ✅ GPT, Gemini, Mistral, etc. | ❌ Claude only | ❌ Claude Code only |
| Output format | JSON graph | Internal (not exportable) | Plain text / markdown bullets |
| Machine-readable | ✅ Structured, parseable | ❌ | ❌ |
| Importance tagging | ✅ h/m/l per node | ❌ | ❌ |
| Status tracking | ✅ open/resolved/deferred/blocked | ❌ | ❌ |
| Graph relationships | ✅ depends_on, resolves, supersedes... | ❌ | ❌ |
| Assumptions explicit | ✅ A node type with confidence | ❌ | ❌ |
| Downloadable file | ✅ .json | ❌ | ❌ |
| Token target | ~400–800 tokens | Varies, not optimized | Varies |
In plain terms
Claude's built-in compaction kicks in automatically when a chat gets too long — but it works silently inside the same conversation. You can't export it, restart clean, or take it to another LLM. It's a life raft, not a handoff.
Existing handoff skills produce a markdown bullet list — readable by humans, but not structured enough for a machine. There's no importance ranking, no node graph, no way to programmatically filter what a new LLM should prioritize. And they're all built for Claude Code, not Claude.ai.
This skill treats context as data, not prose. The JSON graph is structured so any LLM can parse it, importance tags tell the receiving model what it must know vs. what's optional, and compact mode is explicitly optimized to hit the lowest possible token count for the handoff — not just "shorter than the original."
What It Outputs
A structured JSON graph with two modes:
compact (default) — for new chat handoff
~300–500 tokens. Abbreviated keys, telegraphic summaries, no fluff.
{
"v": 1,
"mode": "compact",
"desc": "Building a JWT-secured Spring Boot API with MySQL.",
"n": [
{
"id": "n1",
"t": "F",
"i": "h",
"s": "Spring Boot 3.2, Java 21, MySQL. Maven. Constructor injection throughout.",
"c": [
{ "id": "n1.1", "t": "F", "i": "m", "s": "DB: app_db. ddl-auto=update. Config in application.yml." }
]
},
{
"id": "n2",
"t": "P",
"i": "h",
"s": "403 on all endpoints fixed: added .requestMatchers('/auth/**').permitAll() before anyRequest().authenticated()."
}
],
"handoff": "Continue a Spring Boot JWT API project. Stack: Java 21, MySQL, JPA. Auth working (403 fixed). N+1 query fixed with @EntityGraph. Refresh tokens deferred."
}
readable — for humans / archiving
Full field names, prose summaries, complete handoff prompt. Good for sharing with teammates or saving a project's history.
How to Use
Step 1 — Compress your chat
In any Claude chat, say:
compress this chat
or
save context / handoff to new chat / too many tokens, start fresh
The skill outputs the compact JSON and saves a downloadable .json file.
Step 2 — Paste into new chat
Start a new chat and paste:
Context from previous session:
<paste JSON here>
Continue from this state.
That's it. The new LLM picks up exactly where you left off.
Node Types
| Short | Full | Meaning |
|---|---|---|
F | fact | Technical facts, stack, config, versions |
D | decision | Choices made, things ruled out |
P | problem | Bugs, errors, and their fixes |
G | goal | What the user is building / trying to achieve |
C | code | Snippets, method names, annotations |
X | context | Open/unresolved threads |
Importance Levels
| Short | Meaning |
|---|---|
h | High — losing this breaks continuity |
m | Medium — useful but not critical |
l | Low — dropped in compact mode |
Installation
- Download
context-graph-compressor.skill - Go to Claude.ai → Settings → Skills
- Upload the
.skillfile - Done — Claude will use it automatically when you ask to compress a chat
Works With
- Claude (claude.ai)
- Any LLM that accepts a system/user prompt — GPT-4, Gemini, Mistral, etc.
- Just paste the JSON +
"Continue from this state."as your first message
Files
context-graph-compressor/
├── SKILL.md — Core skill instructions for Claude
└── README.md — This file
License
MIT — free to use, modify, and share.