YourAKShaw/handoff

Pick up every AI coding session exactly where the last one stopped — with any agent.
One small file carries the goal, the state, and the next steps from session to session.
Explore the skill · Handoff format · Contribute · Get help · Security
## Every new session starts from zero
Close a session and the context goes with it. The next one — tomorrow, in another agent, or after the credits run out — has to rediscover the goal, re-read the repository, and re-learn the constraints you already stated: the branch never to touch, the design that must not change, the command that only works with a special flag. Work gets repeated, and decisions get quietly reversed.
Handoff is an agent skill that fixes this with one file, HANDOFF.md, at the root of your repository. At the end of a session the agent rewrites it; at the start of the next, any agent reads it first and continues from the first open step. It works the same in Claude Code, Codex, Gemini CLI, Antigravity, Grok, or any agent that can read a Markdown file.
> Hold only what the next session needs, point to everything else, and never carry secrets.
## ⚙️ What the skill brings to your agent
| Capability | What it means in practice |
| Instant continuation | A new session reads one short file and starts the first next step, without re-reading the whole repository. |
| Constraints that stick | Decisions and standing rules you gave in one session are recorded, so the next agent keeps them. |
| Agent-neutral | Plain Markdown at the repository root, plus a rule for AGENTS.md, CLAUDE.md or GEMINI.md. No lock-in to one tool. |
| Always current | The file is rewritten, never appended. Finished work drops out; the length stays under about 80 lines. |
| Safe by default | Secrets, tokens, one-time codes and personal data are never recorded. |
| Version-controlled | The handoff is committed with the work when commits are allowed, so its history travels with the repository. |
## 🧩 How it works
mermaid
flowchart LR
A[Session ends] -->|"handoff" or /handoff| B[Agent rewrites HANDOFF.md]
B --> C[(HANDOFF.md in the repo)]
C --> D[New session, any agent]
D -->|reads first| E[Continues from Next steps]
E --> A
HANDOFF.md always has the same five sections, defined in the handoff format:
| Section | Holds |
| Goal right now | The current aim and the standing constraints. |
| Where things stand | Branch, what is finished, the last checks actually run. |
| Next steps | A numbered list, each naming the file or command to start from. |
| Open questions | Decisions waiting on someone. |
| Gotchas | Traps a fresh session would otherwise hit. |
## Three workflows
| Workflow | When | Result |
| Save a handoff | End of a session, when you ask | HANDOFF.md rewritten and committed when allowed |
| Resume a handoff | Start of a session | Work continues from the first open step |
| Bootstrap a repository | A repository without a handoff | HANDOFF.md plus a rule in its agent instructions |
## 📥 Get started
Keep SKILL.md, references/, workflows/, assets/ and scripts/ together so the skill can find its resources.
### Install once for every agent
Clone one shared copy and link it into each agent. Full details, including where each agent looks, are in agent setup.