Prompt Chain
1marcelserrano/prompt-chain
Quebre tarefas longas em prompts que se autopropagam entre chats
Ver no GitHubSkill para dividir tarefas complexas em cadeias de prompts autopropaláveis que atravessam múltiplas sessões de chat. Cada fase gera a próxima com contexto completo, eliminando limitações de contexto e permitindo trabalho multi-etapa sem colapso de estado.
Instalação
npx convez-ai 1marcelserrano/prompt-chain1 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
- Projetos multi-fase que excedem limite de contexto
- Trabalhos que precisam ser retomados em novas sessões
- Isolar falhas em diferentes estágios de processamento
- Handoff de prompts entre usuários ou contextos
Veredito

Solução pragmática e sem dependências para trabalhos longos em Claude. Elimina friction real em tarefas multi-etapa, mas exige disciplina do usuário para estruturar e copiar prompts entre chats — não automatiza execução, apenas organiza estado.
README
A tradução para o português deste README ainda está na fila.
prompt-chain
English · Português
Break complex multi-step tasks into self-propagating prompts that survive context limits.
The whole skill is one 307-line markdown file. No dependencies, no server, no external state store. Built for anyone running long multi-phase work in Claude — developers and non-developers alike.
Install • How it works • When to use • FAQ
A Claude skill for chaining work across fresh chat sessions. Each prompt generates the next, carrying full context forward. Cold-start safe. Context-limit immune.
How it works
flowchart LR
A[You paste<br>STAGE 1] --> B[Chat 1 executes<br>+ emits STAGE 2 prompt]
B -->|copy-paste| C[Chat 2 executes<br>+ emits STAGE 3 prompt]
C -->|copy-paste| D[Chat N executes]
D --> E[CHAIN COMPLETE<br>all deliverables listed]
B -.blocker.-> P[CHAIN PAUSED<br>asks, never guesses]
| Without prompt-chain | With prompt-chain |
|---|---|
| One mega-prompt, context fills up, output degrades, you lose state when the chat dies. | A first prompt sets the chain. Each chat finishes its phase and outputs the next self-contained prompt. You paste it into a new chat. State carries. |
Same work. Half the friction. No mid-session collapse.
Install
# Via skills CLI (recommended)
npx skills add 1marcelserrano/prompt-chain
# Or manually — back up first if the folder already exists:
# mv ~/.claude/skills/prompt-chain ~/.claude/skills/prompt-chain.backup
git clone https://github.com/1marcelserrano/prompt-chain.git
cp -r prompt-chain/skills/prompt-chain ~/.claude/skills/
Verify: open a new Claude session and run /skills (or ask "what skills do you have?"). prompt-chain should be listed. If it isn't, check that ~/.claude/skills/prompt-chain/SKILL.md exists and restart the session.
Update: git pull in the cloned repo, then re-copy. New versions are logged in CHANGELOG.md.
When to use
Trigger phrases (EN + PT-BR):
- "execute in stages across separate chats"
- "split this across sessions"
- "self-propagating prompt"
- "prompt that generates the next"
- "chain of prompts"
- "cold start between chats"
- "executar por etapas em chats separados"
- "cada etapa em um novo chat"
- "prompt autocontido autopropagante"
Natural language works too. Just describe the multi-phase task and ask for it to be split.
Why
- Context limits stop being a wall. Each phase runs in a fresh session with only what it needs.
- Phases isolate failures. A broken phase 3 doesn't poison phases 4-7.
- Resumable. Lost the chat? Open the last good prompt, continue.
- Hand-offable. Each prompt is self-contained — you or someone else can pick up any link.
What it does / doesn't
| Does | Doesn't |
|---|---|
| Break P0 → P1 → P2 → ... structure | Execute the phases for you |
| Pass context forward via prompt text | Use external state stores |
| Auto-detect chainable patterns | Force-chain trivial single-step tasks |
| Work in any chat-based Claude interface | Require a specific platform |
How it compares
Two neighboring families solve related problems. Different jobs:
| Session-handoff tools | Workflow engines | prompt-chain | |
|---|---|---|---|
| When it acts | Reacts when a session fills up — save state, resume later | Runs a coded pipeline you build first | Plans the whole crossing upfront — N stages, each with its own Definition of Done |
| Scope | One link at a time | Full orchestration: retries, validation, branching | The full chain is designed before stage 1 runs |
| Where state lives | Files, hooks, hidden dirs | Databases, APIs, app state | Inside the prompt text itself |
| Requires | Usually Claude Code (plugin + hooks) | Code, a runtime, often a server | A chat window |
| Examples | handoff, claude-handoff | LangGraph, workflow engines | this repo |
Sourcegraph's Amp Handoff also generates the next thread's prompt — reactive, one link at a time, inside Amp's paid stack. prompt-chain does the planned, multi-stage version of that idea in plain markdown.
Use handoff when a session surprises you. Use a workflow engine when the orchestration deserves code. Use prompt-chain when you can see the phases coming and want zero infrastructure.
Worked examples: examples/ — full chains showing Stage 1, the emitted Stage 2, and CHAIN COMPLETE.
Evidence: benchmark/ — the same dead task continued three ways (no carry · handoff snapshot · chain prompt), raw transcripts included.
Ready-to-fill seeds: templates/ — research, refactor, launch, audit, and content chains.
FAQ
Does it send my data anywhere? No. The skill is a markdown file Claude reads locally. Your prompts and context stay inside your normal Claude sessions — nothing extra is stored or transmitted.
How do I uninstall?
Delete the folder: rm -rf ~/.claude/skills/prompt-chain. Done.
What does it work with? Any chat-based Claude interface: Claude Code (CLI/desktop), Claude.ai, Claude Cowork. macOS, Linux, Windows — it's just markdown.
Will it be maintained?
Yes. Versions follow CHANGELOG.md. Update with git pull or re-run npx skills add.
About the author
Built by Marcel Serrano, founder of MSCREATIVE.SYSTEMS™ — Barcelona. This skill is the propagation backbone behind the MSCS editorial and skill ecosystem.
More on working with AI without losing your own voice: Fronteirista — the free newsletter where these systems get built in public.
Contributing
PRs welcome. See CONTRIBUTING.md.
License
MIT — see LICENSE.
Forged at MSCREATIVE.SYSTEMS™ — Barcelona