diff --git a/README.md b/README.md
index 7dec7e8..7ab4f81 100644
--- a/README.md
+++ b/README.md
@@ -15,18 +15,29 @@ It assumes analysis repos made from [analysis_template](https://github.com/quadb
 (pixi-managed) and code repos from the [scverse cookiecutter](https://github.com/scverse/cookiecutter-scverse)
 (uv-managed). Nothing in it is specific to a dataset, an analysis or a compute environment.
 
-## Install
+## Use
 
-**The plugin**, once per machine:
+Start each analysis task as its own session, from the analysis repo's main checkout. Pull it first: the worktree
+branches from `origin`'s default branch but runs on the main checkout's environment.
 
 ```bash
-claude plugin marketplace add quadbio/claude-plugins
-claude plugin install analysis-workflow@quadbio
+claude --worktree <task>
 ```
 
-The skill is then available everywhere, so it can set up a new project. The hooks and session-start rules act only
-in repos that enable the plugin in their committed `.claude/settings.json`, as
-[analysis_template](https://github.com/quadbio/analysis_template) does:
+1. The agent works on that worktree's branch, in a new task directory `analysis/<topic>/<task>_vN/`; a later
+   answer to the same question gets the next version, in a new session.
+2. It asks you which working object is current, then works in scripts; outputs go to the task directory.
+3. Results reach a shared working object only after you sign off the dry run.
+4. It opens a PR. Once it is merged, leave with `/exit` and remove the worktree.
+
+Notebooks and environment changes (`pixi install`, `pixi add`) stay with you, in the main checkout. With no
+analysis repo yet, ask Claude to start an analysis project: the skill sets one up.
+
+## Install
+
+**The plugin**, once per machine, from the [quadbio marketplace](https://github.com/quadbio/claude-plugins). The
+skill is then available everywhere; the hooks and session-start rules act only in repos that enable the plugin in
+their committed `.claude/settings.json`, as [analysis_template](https://github.com/quadbio/analysis_template) does:
 
 ```json
 {
diff --git a/hooks/session_context.md b/hooks/session_context.md
index 13eb658..a40253e 100644
--- a/hooks/session_context.md
+++ b/hooks/session_context.md
@@ -2,7 +2,7 @@ This repository follows the analysis-workflow conventions. Load the `analysis-wo
 analysis code, outputs or data. Always:
 
 1. One task = one session = one git worktree = one directory `analysis/<topic>/.../<name>_vN/`. A new task
-   directory needs a new session.
+   directory needs a new session, which the human starts with `claude --worktree <name>` in the main checkout.
 2. Every output path comes from `task_paths(__file__)` (package `analysis_workflow`); never a bare relative path.
 3. Ask the human which working object (and which label keys) is current; never infer it from a script or the
    newest file name.
diff --git a/skills/analysis-workflow/SKILL.md b/skills/analysis-workflow/SKILL.md
index fd5c327..7449ef8 100644
--- a/skills/analysis-workflow/SKILL.md
+++ b/skills/analysis-workflow/SKILL.md
@@ -41,7 +41,8 @@ into the human lane.
 
 **Start**
 1. One task = one session = one git worktree = one branch = one directory `analysis/<topic>/.../<name>_vN/`, merged
-   through one or more PRs. A new task directory needs a new session; a later answer to the same question is `_v2`.
+   through one or more PRs. A new task directory needs a new session, which the human starts with
+   `claude --worktree <name>` in the main checkout; a later answer to the same question is `_v2`.
 2. Copy `assets/task_template/`.
 3. Ask which working object and which label keys are current, and record them under Inputs. Never infer them from a
    script or the newest file name.