Split long work into stages that survive a new session.
Long work rarely fits in one session. The conversation fills up, the tokens run out or the client restarts, and whatever only existed in the chat is gone.
So keep the work out of the chat. Each stage is a task with a clear deliverable, and the agent records its progress as it goes. Whoever continues starts from the board, whether that's the same AI or a different one.
Run the work in stages
Agree on the stages before any work starts. Each one says what it needs, what it produces and how to tell it's finished, so a new session can take it over without asking you.
- Describe the goal and have your agent propose the stages.
- Approve them. Each stage becomes a task, in order.
- The agent works through one stage at a time, saving its progress and the deliverable to the task.
- When a session ends, start a new one with the resume prompt.
Move the website to a new CMS
Example project · A new session can continue stage 3 from its updates
Backlog0
No tasks in this group.
Up Next1
Stage 4: Check every link and redirect
In Progress1
Stage 3: Move the content
Done2
Stage 2: Map old URLs to new ones
Stage 1: List every page on the site
Skill: staged-work
In Claude Code, save this as .claude/skills/<name>/SKILL.md. With any other connected AI, paste it at the start of a conversation.
---
name: staged-work
description: Run long work as stages saved in Organarium, so any session or agent can continue it. Use when work won't fit in one session.
---
Organarium is the record of this work. This session can end at any moment, and whoever continues will only have the board. Use your own name as the agent on everything you write.
## Setting up
1. Create the project. Put the goal, the constraints and what "finished" means in its description.
2. Propose the stages as a short list and wait for my approval.
3. Create one task per stage, titled "Stage 1: ...", "Stage 2: ..." and so on, in Up next. Each description says:
- Goal: what this stage is for.
- Input: what it needs from earlier stages.
- Deliverable: what it produces.
- Done when: how to tell it's finished.
## Working
1. Read the project description and the list of stages. Take the lowest-numbered stage that isn't done.
2. If it's already in progress, read its updates first and continue from there. Don't redo work that's recorded as finished.
3. Move it to In progress. If it has several parts, add them as subtasks and check them off as you go.
4. Add an update after each meaningful piece of work: what you did, what you decided and why, and what comes next. Do this as you go. You may not get the chance to do it at the end.
5. Add the deliverable to the task description. If it's a file or lives somewhere else, save it there and add a link to the task.
6. When "Done when" is met, add a final update and move the task to Done. Then continue with the next stage.
## If you're running out of room
Stop starting new work. Add an update with exactly where things stand and the next action to take.Continue from another session or another AI
This is all it takes to pick the work back up. The prompt is the same whether you're opening a new session or giving the work to another connected AI.
Prompt: pick up where the last agent stopped
Continue the work in my Organarium project "Move the website to a new CMS". Read the project description and its tasks, find the first stage that isn't done, and read that stage's updates. Tell me in two lines where things stand, then continue following the same rules: record your progress as updates while you work.
Rely on what's on the board. If something you need isn't there, tell me what's missing instead of guessing.What to expect
- A new session only knows what was recorded, so the habit that matters most is adding updates during the work. An agent that runs out of tokens can't write a goodbye.
- Organarium doesn't store files. Deliverables are text in the task, or links to wherever the files live.
- Reading from the board can mean less repeated context than reloading a long chat, but how many tokens you use depends on your client and your work.
More ways to use Organarium
Capture ideas and everyday tasks
Pull action items out of a brainstorm, save an idea in one sentence, and keep personal lists up to date, down to the groceries on your receipt.
Run several agents from one board
Let one agent plan and others do the work, even across different AIs. Or keep an agent working through the tasks you approve from another session.
Check each step before the next one
Build long processes where every step is documented in a task, and a reviewer approves it or requests changes before work moves on.