Anthropic / Claude agent skills and Continuous-Claude workflow

Anthropic-related projects propose persistent workflows and reusable skills for Claude agents. - Continuous-Claude solves context loss and forgetting in long coding sessions - Claude Code compresses prior dialogue into lossy summaries - Multiple compressions drop detailed facts like function parameters - Continuous-Claude provides a Continuity System replacing fuzzy memory with documents - Ledger logs goals, completed work, and next steps in-session - /clear reloads the ledger to restore state after context reset - Handoff generates markdown handover docs for cross-session transitions - Workflow splits fast "skills" actions and full Agents for heavy tasks - Agent pipeline: Plan -> Validate -> Implement in new contexts - MCP code execution runs tools externally, returning only essential outputs - Hooks: SessionStart, PreCompact, StatusLine automate loading and preservation - Onboarding (/onboard) generates initial cognitive docs from a repository - TDD Workflow enforces failing test, implement, refactor cycle - Recall uses Reasoning History to report prior attempts and commits - Open-source repository link provided for Continuous-Claude - Anthropic Agent Skills package domain knowledge into SKILL.md with YAML metadata - Agent Skills load in three levels: metadata, main instructions, resources/code - Level 1 metadata is always loaded; Level 2 loads on trigger (<5k tokens typical) - Level 3 resources/scripts load on demand via shell commands - Skills support file-system access, bash, code execution in Agent VM - Anthropic published an Agent Skills standard and provided prebuilt Skills links - Best practices: clear metadata, stepwise instructions, scripts for deterministic tasks
Key insights
Continuous-Claude addresses Claude Code session forgetting with ledger and handoff
Continuous-Claude implements a Continuity System to address Claude Code’s loss of context in long sessions by storing structured documents instead of relying on repeated compressed summaries. It logs goals, completed work, and next steps in a real-time ledger; supports a /clear action that reloads the ledger to restore state; and generates markdown handoff documents at day’s end or task completion to enable cross-session or cross-agent continuity. The system uses lifecycle hooks (SessionStart, PreCompact, StatusLine) to auto-load and preserve state and to intercept compaction, preventing information loss caused by successive summary compression.
Agent orchestration separates lightweight skills and heavy agents
The Continuous-Claude architecture distinguishes between 'skills'—quick in-session actions like making a commit or checking docs—and 'agents' that run in fresh contexts for resource-heavy tasks (e.g., designing backend architecture). Typical agent flow defined: Plan Agent -> Validate Agent -> Implement Agent. This split reduces token consumption in the main session and allows complex workflows to execute in isolated, clean contexts while preserving an overall documented state.
External tool execution and MCP reduce token pollution
Continuous-Claude runs external tools and scripts outside the agent context (MCP code execution) and only returns essential results to the agent, minimizing token usage and preventing tool definitions or verbose outputs from polluting the main context. The project recommends executable scripts for deterministic operations and uses hooks to save state before automatic compaction events.
Anthropic’s Agent Skills standard enables progressive, on-demand loading
Anthropic published an Agent Skills standard that packages domain knowledge and workflows into filesystem-backed skill directories. Each Skill contains SKILL.md with YAML metadata and supports three progressive loading levels: Level 1 metadata always loaded (low token cost), Level 2 main instructions loaded on trigger (typically <5k tokens), and Level 3 resources/scripts loaded on demand via shell commands. Skills run in an Agent VM and can access files, bash, and code to keep large reference material out of the default context until needed.
Skill design best practices emphasize metadata, modularity, and scripts
Recommended best practices for Agent Skills include clear metadata describing what the skill does and when to trigger it, concise SKILL.md with step-by-step instructions and examples, splitting large content into subfiles, and using executable scripts for deterministic tasks to avoid bloating the agent context. Anthropic provides prebuilt Skills and examples to accelerate adoption and testing.