Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/bitwikiorg/continuity/llms.txt

Use this file to discover all available pages before exploring further.

Claude Code reads CLAUDE.md automatically from the project root on every turn, and also discovers nested CLAUDE.md files in subdirectories as the agent moves through a project. This makes it the natural entry point for Continuity: place your boot instructions, state chain declarations, and load order in CLAUDE.md and Claude Code enforces them by self-reference — the file declares what state exists, and the agent follows its own specification as it runs.

​
Setup

1

Copy Continuity templates to your workspace

Copy the files matching your chosen tier from the Continuity repo into your project root. The Minimum tier is enough to start:
cp continuity/{AGENTS,IDENTITY,SELF,SOUL,STATE,INIT,TODO,SNAPSHOT}.md /your/workspace/
Replace all {{UPPER_CASE}} placeholders with your agent’s actual values before booting.
2

Create or configure CLAUDE.md as the entry file

Claude Code does not auto-load AGENTS.md — it only loads CLAUDE.md. You have two options:Option A — Rename: Rename the pack’s AGENTS.md to CLAUDE.md. This makes CLAUDE.md the single authoritative entry file.Option B — Delegate: Keep AGENTS.md and create a minimal CLAUDE.md that delegates to it:
# CLAUDE.md
Read AGENTS.md and follow its boot instructions.
Either approach works. Option A is simpler for new workspaces. Option B is useful when you share a workspace with runtimes that auto-load AGENTS.md (Codex, Hermes, OpenClaw).
3

Declare the load order in CLAUDE.md

CLAUDE.md should declare which state files to load on boot and in what order. A minimal declaration looks like this:
# CLAUDE.md

## Boot sequence
On every session, load in order:
1. IDENTITY.md — who I am
2. STATE.md — current project truth
3. TODO.md — active task queue
4. SNAPSHOT.md — last known good state

Read STATE.md before acting.
After meaningful work, update STATE.md, TODO.md, and SNAPSHOT.md.
You can extend the load order as your agent’s loop grows — add MEMORY.md, TOOLS.md, PLAN.md, or any file your loop actually needs. Load only what the current task requires; do not load everything on every turn.
4

Replace placeholders and run INIT.md on first boot

Search for any remaining {{UPPER_CASE}} markers across all copied files and fill them in. Then prompt Claude Code to run the INIT.md reconstitution procedure:
Run the INIT.md protocol. Read STATE.md and TODO.md. If READY, proceed with the next valid action.
INIT.md runs nine stages: preflight, inspect, scan, load, reconcile, discover, validate, declare, and snapshot. It creates any missing files and emits READY when the state chain is valid.

​
Writeback contract

Claude Code can read and write files directly. After completing meaningful work, the agent must update the state files whose claims are no longer accurate. The minimum writeback set is:
FileWhat to update
STATE.mdNew operational truth — what IS now
TODO.mdMark completed items; add new tasks
SNAPSHOT.mdRefresh the restoration anchor
completed/Move done tasks here with evidence
For larger changes, also update:
  • LEARNINGS.md — if new insights emerged during the session
  • MEMORY.md / memory/ — if durable facts changed
  • PLAN.md — if the intended sequence shifted
  • snapshots/ — if STATE.md was overwritten (preserve old value first)
Do not preserve stale state for politeness. If reality changed during a session, the state files must reflect that change before the session ends. Stale state causes the next session to boot with incorrect context.

​
Auto-injection and AGENTS.promptinclude.md

Claude Code auto-loads CLAUDE.md natively — it does not use AGENTS.promptinclude.md as an auto-injection mechanism. AGENTS.promptinclude.md is for runtimes such as Agent Zero and Cursor that support a prompt-include protocol. For Claude Code, the equivalent is declaring your load order directly inside CLAUDE.md so the agent loads the right state files on every turn by following its own specification. If you also use runtimes that do read AGENTS.promptinclude.md (Agent Zero, Cursor), you can keep the file in your workspace for those runtimes. The full required-tier list from the Continuity repository: 
# AGENTS.promptinclude.md

# Required tier (auto-loaded)
@AGENTS.md
@IDENTITY.md
@SOUL.md
@SELF.md
@USER.md
@INIT.md
@MEMORY.md
@TOOLS.md
@STATE.md
@CONTEXT.md
@PLAN.md
@TODO.md
@SNAPSHOT.md

# Recommended tier (uncomment to enable)
# @KNOWLEDGE.md
# @INFRASTRUCTURE.md
# @POTENTIAL.md
# @CONTEXT_BUDGET.md
# @MISSIONS.md
# @TASKS.md
# @LEARNINGS.md
# @WARNING.md
# @HEARTBEAT.md
# @EVENTS.md
# @BOOTSTRAP.md
# @BOOT.md
Keep the load order in CLAUDE.md small. Only include files the agent genuinely needs on every turn. Files that are only relevant for specific tasks should be loaded on demand, not listed in the boot sequence — your context budget is finite.

​
Nested workspace example

All Continuity files can be nested into project subdirectories. Claude Code discovers nested CLAUDE.md files as the agent descends into subdirectories, building a scoped state chain: 
/my-project
  CLAUDE.md          ← global identity and load order
  STATE.md           ← "Current: Building auth system"
  TODO.md            ← high-level task queue

  /backend
    CLAUDE.md        ← backend-scoped boot instructions
    STATE.md         ← "Current: Refactoring JWT handler"
    TODO.md          ← backend tasks

    /auth
      STATE.md       ← "Current: Writing token rotation tests"
When Claude Code works inside /backend/auth/, it loads the full chain from root to the active scope:
  1. /CLAUDE.md — global identity
  2. /STATE.md — project status
  3. /backend/CLAUDE.md — backend-scoped instructions
  4. /backend/STATE.md — subsystem status
  5. /backend/auth/STATE.md — current task context
Child state files refine scope and add local detail. They cannot weaken safety rules or authority boundaries declared by parent files.

​
CLAUDE.md vs AGENTS.md

Either CLAUDE.md or AGENTS.md can serve as the Continuity entry file — the system does not prescribe which one. The only constraint is that the file must be the one your runtime auto-loads or is pointed to explicitly. If you share a workspace across multiple runtimes (e.g., Claude Code and Codex), the Option B delegation pattern — a CLAUDE.md that reads AGENTS.md — lets you maintain a single authoritative entry file while satisfying both runtimes’ auto-load conventions.

Build docs developers (and LLMs) love

Get started for freeTalk to us