For helper loading, follow package references. Write in the shared technical English; name the evidence behind the placement, as evidence says.

What's next

Give one answer: where the user is in the workflow, the next step, the exact command, and the session to run it in. Read the state first; ask only when two positions fit. Write nothing to the project, and stop after the answer: the user runs the next step.

1. Read the state

Run python3 <package>/skills/whats-next/scripts/workflow_state.py <project>. It prints the branch and its distance from the default branch, the uncommitted files, the investigation location, every note and report our skills left (understanding note, plan, diagnosis note, PR review guide, acceptance report) with its summary line and date, local tickets, visual-pr tasks, Matt's tracker setup, and whether a local tool record exists. It reads only and creates nothing.

Add what the scan cannot see: the step this conversation just finished and the path or URL it reported, the decisions the user stated, and the argument when it names a goal. When the branch has an upstream, check for an open PR with the host's Git tooling, such as gh pr view. Print the tool record with python3 <package>/skills/tool_setup_path.py --show <project> to learn which originals are installed and how the host registers them; none means no setup ran, so take install commands from original skills.

2. Place the user

Take the first row whose evidence holds, top to bottom. "The same session" is the one that holds the planning conversation. "A new session" is a fresh context window: a new session in the host, or its clear command.

An idea with no repository under it takes Matt's grill-me. A huge, foggy effort takes his wayfinder, which merges at to-spec. When two rows fit and the conversation does not settle it, ask one question with the recommended row first, as asking says for a round and for a host without the question tool.

3. Answer

Reply in this shape, with the real paths and URLs from the scan:

**Where you are:** <position, with its evidence: branch, file, PR, report line>
**Next:** <skill> — <what it settles, one line>
**Run:** <the command for this host, with the paths and the choices already settled>
**Session:** <same | new> — <reason>
**Carry:** <the paths or URLs the next step needs, which a fresh session cannot see>
**Also possible:** <one alternative row, and when it is the better one>

Write the command as the host registers it: /aantoniadis-skills:<name> in Claude Code, $aantoniadis-skills:<name> in Codex, /<name> or opencode run --command <name> in OpenCode; an original as the tool record lists it, such as /mattpocock-skills:implement or /implement. Put a kickoff choice the conversation already settled inside the command, so the next skill asks less. When an original is not installed, give its install command and the local route instead. Name what was not checked, such as a tracker that could not be reached. Done when the user can paste the command.