organarium

Have an agent check each step before the next one starts.

In a long process, a small mistake in an early step spreads through everything after it. By the time you read the result, it's expensive to find where it went wrong.

So add a check between steps. One agent does the step and documents it in a task. Another reviews it against written criteria, and work only continues once it's approved. You read the trail on your board and step in when a step needs you.

Follow one recipe, with review between steps

A recipe is the list of steps, each with criteria the reviewer can check. Each step is one task, and its tag shows where it is in review:

  • review:pending means the step is finished and waiting for a reviewer.
  • review:changes-requested means it went back to its author, with a list of what to fix.
  • review:approved means it passed. The task moves to Done and the next step can start.
  • needs-person means the agents have stopped and it's your turn.

Market report

Example project · Review state is shown in the tags

Backlog0

No tasks in this group.

Up Next1

Step 4: Write the draft

recipe
midUnassigned

In Progress1

Step 3: Write the outline

recipereview:changes-requested
highUnassigned

Done2

Step 2: Collect the sources

recipereview:approved
midUnassigned

Step 1: Define the questions

recipereview:approved
midUnassigned

What the updates on step 3 could look like

  1. claude-code

    Outline ready in the description: five sections, each tied to a question from step 1. Tagged review:pending.

  2. codex

    Changes requested. Question 4, on pricing, has no section, and section 3 cites nothing from the approved sources. Tagged review:changes-requested.

  3. claude-code

    Added a pricing section and tied section 3 to sources 2 and 6. Tagged review:pending.

  4. codex

    Approved. Checked all five questions against the sections and every citation against step 2. The outline gives step 4 what it needs. Moved to Done.

Skill: gated-pipeline

In Claude Code, save this as .claude/skills/<name>/SKILL.md. With any other connected AI, paste it at the start of a conversation. Run the author and the reviewer in separate sessions, so the review relies on what's written down. They can be different AIs.

---
name: gated-pipeline
description: Run a recipe in Organarium where each step is one task and a separate reviewer approves it before the next step starts. Use when I ask to run a recipe or a reviewed process.
---

Each step of the recipe is one Organarium task, numbered in order. Use your own name as the agent on everything you write.

Review state is a tag on the task. When you change it, remove the old review tag and keep the task's other tags, because saving tags replaces the whole set. Only one step is worked on at a time.

## Setting up

If the steps don't exist yet, create one task per step in Up next, copying each step's goal, deliverable and acceptance criteria into its description.

## If I ask you to be the author

1. Find the lowest-numbered step that isn't tagged review:approved.
2. If it's tagged needs-person, stop and tell me. If it's tagged review:pending, stop and say it's waiting for review.
3. Read the task and its updates, and the approved steps before it. If it's tagged review:changes-requested, fix what the latest review listed, and only that.
4. Move it to In progress and record your decisions as updates while you work.
5. Add the result to the description, with links to anything stored elsewhere.
6. Tag it review:pending, add an update saying what's ready to check, and stop. Never approve your own work.

## If I ask you to be the reviewer

1. Find the lowest-numbered step tagged review:pending. Don't rely on anything but the board and the sources it links to.
2. Check the result against each acceptance criterion, and against what the next step needs to start.
3. If it passes: tag it review:approved, add an update saying what you checked, and move it to Done.
4. If it doesn't: tag it review:changes-requested and add an update listing exactly what to fix. Don't fix it yourself.
5. If this is the third round of changes on the same step, tag it needs-person instead, and add an update explaining the disagreement.

Recipe: research a market report

An example recipe. Give it to the agent along with the skill, then start the author and the reviewer.

Run this recipe in my Organarium project "Market report", using the gated-pipeline skill.

Step 1: Define the questions
Deliverable: the audience, the decision this report supports, and up to five questions.
Accept when: each question is specific enough to answer with evidence, and each one relates to the decision.

Step 2: Collect the sources
Deliverable: a list of sources with links, the date of each, and which question it helps answer.
Accept when: every link opens, each question has at least two independent sources, and anything older than two years is marked as such.

Step 3: Write the outline
Deliverable: the report's sections, each with its main claim and the sources behind it.
Accept when: every question has a section, and every claim cites a source from step 2.

Step 4: Write the draft
Deliverable: the full text, or a link to it.
Accept when: it follows the approved outline, it's written for the audience in step 1, and it opens with the answer to the decision.

Step 5: Check the facts
Deliverable: a list of every number and factual claim, with the source and whether it holds up.
Accept when: the reviewer opens the sources for a sample of at least ten claims and they match. Anything unsupported is corrected or removed from the draft.

Step 6: Summarize for me
Deliverable: one page with the conclusions, how confident we are in each, and what we couldn't confirm.
Accept when: someone who only reads this page gets the decision and the main risks.

Prompt: start the reviewer

Paste this into a separate session, or another AI.

Use the gated-pipeline skill as the reviewer on my Organarium project "Market report". Review the step that's waiting, then tell me the verdict in two lines. You're checking this work, so be specific about anything that falls short. Passing weak work costs more later than sending it back now.

What to expect

  • The review tags are a convention described in the skill, not a feature of Organarium. It stores the tags and filters by them, and your agents follow the rules.
  • Your AI client does the running. How long it can go without you depends on the client, your plan and its limits. If a session stops, a new one continues from the board.
  • An agent reviewing work catches mistakes, but it can miss them too. Decide which steps a person should read, and add that to your recipe's criteria.

Connect your AI ↗

More ways to use Organarium

Give your AI memory, tasks, and a dashboard.

Start now