---

name: context-transfer description: "Generate a state document (context transfer) that allows any fresh AI instance to resume work without conversation history. Use this skill whenever the user asks to checkpoint, create a state transfer, save state, generate a handoff document, create compaction insurance, or says anything like 'snapshot this', 'create a context transfer', 'checkpoint', 'save state for handoff', 'write a recovery doc', or 'transfer context'. Also trigger when the user mentions compaction concerns, session handoff, multi-session continuity, or wants to preserve the current conversation's state. Works for any conversation type: projects, research, debugging, investigations, design sessions, or long-running chats. This skill produces a structured markdown document optimized for AI cold-start recovery, not human notes."

Context Transfer

Generate structured state documents for AI cold-start recovery across any conversation type.

When To Use

• User asks to checkpoint, snapshot, or save state
• Session is getting long and compaction insurance is needed
• Work will continue in a new session (handoff)
• Multi-session work needs persistent state
• User mentions compaction, context loss, or continuity concerns
• Any conversation where accumulated decisions, findings, or state would be costly to lose

Before Starting

Read the full protocol:

Read references/protocol_v2.4s.md

This contains the complete section definitions, templates, rules, and quality checklist.

Quick Reference

The context transfer has 11 sections. Scale to complexity — not every conversation needs all sections.

Section	Purpose	Required?
I. What This Is	Identity + "It is NOT" correction	Always
II. Domain Context	Problem-space briefing	Multi-session+
III. Architecture	System structure	Build/engineering work
IV. Decisions	Numbered with narrative WHY	When decisions exist
V. Current State	Step sequence or phase summary	Always
VI. Artifacts	File inventory with status	When files exist
VII. Anti-Drift Warnings	Observed/predicted failure modes	Multi-session+
VIII. What NOT To Do	Catastrophic prohibitions	When applicable
IX. Glossary	Context-specific terms	Domain-specific
X. Open Questions	Unresolved issues	When they exist
XI. Recovery Pointer	Exact restart location + path	Always

Minimum viable: Sections I, V, XI.

Workflow

1. Read the protocol (references/protocol_v2.4s.md) — understand all section rules
2. Assess scope — use Scaling Guide to determine which sections are needed
3. Gather state from conversation — extract decisions, current step, artifacts, open questions
4. Write each section following the protocol templates exactly
5. Apply critical rules:
   • §I must include "It is NOT..." correction sentence
   • §IV decisions must have narrative WHY (mark [inferred] if rationale not explicitly stated)
   • §VII every warning tagged [observed] or [predicted]
   • §XI must include Work lives in: (write UNKNOWN — ask user if path not known)
6. Run Quality Checklist from protocol
7. Save to /mnt/user-data/outputs/ and present to user

Output

Save as: [topic_or_project]_context_transfer.md in /mnt/user-data/outputs/

The document header should include:

# Context Transfer — [Topic or Project Name]

**Version:** [N]
**Updated:** [date]
**Purpose:** Compaction insurance. If you just got compacted, read this first.

Key Rules (from protocol)

These are the rules most likely to be missed. Full rules are in the protocol reference.

• Single source of truth: One context transfer per conversation/project. Never maintain parallel state docs.
• Correction sentence: §I must state what it IS then what it is NOT.
• Narrative WHY: §IV uses prose, not tables. Tables compress reasoning too much for cold start.
• Inferred rationale: If WHY isn't explicitly stated, mark as [inferred] — never present inference as ground truth.
• Tag all warnings: §VII requires [observed] or [predicted] on every warning. No untagged.
• Never fabricate paths: §XI Work lives in: must be real or UNKNOWN — ask user.
• Update immediately: Write to file after each decision/state change. Don't batch.

Notes

• This skill is validated across Claude, GPT, and Gemini cold instances (n=5 tests, zero operational failures)
• The protocol is model-agnostic — works without cognitive scaffolding
• Optimized for AI pattern recognition, not human readability
• The context transfer IS the conversation's memory — treat it as critical infrastructure