What's next

Advise the user which installed skill or agent to use next, and run nothing. Done when the user has one to three suggestions, each with the exact way to start it, and the choice is left to them.

Procedure

1. Take the goal from the request or the argument. When there is none, read the situation, using read-only commands only:
   • in a git repo: git status, the current branch, and git log --oneline -10
   • the lifecycle handoff files, each of which marks a stage reached: .agents_workspace/product/business/BUSINESS_PLAN.md (Shape done), product/plans/*.md (a plan not built is in flight), product/ARCHITECTURE.md (built), product/EVIDENCE.md (proven)
   • the newest file in .agents_workspace/handoff/, if any, for work a past session left open
2. Collect the candidates: every skill and agent in this session's listings, any plugin's, plus each user-only skill in the table below whose plugin claude plugin list shows as enabled.
3. If the request and the repo still leave the goal open, ask one round with AskUserQuestion: at most four questions, each with a recommended answer. With no human present, skip the question, assume the most likely goal, and say so in the answer.
4. If no installed candidate fits, run claude plugin list --available --json. Its available array holds only plugins not yet installed, each with pluginId, description, and marketplaceName. Match the descriptions against the goal and keep at most two, looking first at plugins whose name starts with ceh-, then at other marketplaces the user added, and at claude-plugins-official last. Mark each "not installed", with claude plugin install <pluginId> --scope <user|project|local> and then /reload-plugins as the way to start it. Leave the scope to the user, recommending project: user turns the plugin on in every folder, and a committed project scope turns it on for everyone who clones the repo.
5. Answer in the Output format and stop.

Rules

• Run nothing: no Skill tool call, no subagent dispatch, no install, no file write. The user chooses what runs, so a suggestion that starts itself takes that choice away.
• Suggest an installed skill or agent first, and a plugin to install only through step 4. Never name a plugin from memory, which may be stale or not in any marketplace the user added. When step 4's command fails, point to the install guide, https://raw.githubusercontent.com/cheneeheng/ceh-claude-code-library/main/docs/GETTING_STARTED.md.
• Rank by fit to the moment, not by breadth: the skill whose trigger matches what the user is about to do comes first.
• Give the reason for each suggestion from what was read ("product/plans/auth.md exists and no code matches it yet"), never a generic one ("planning is useful").

User-only skills

These set disable-model-invocation: true, so they never appear in the session's skill listing. Suggest one only when its plugin is enabled.

Skill	Start it	When it fits
ceh-coding-conduct:refactor-repo	/ceh-coding-conduct:refactor-repo [module-or-path]	A whole codebase or module has grown complex and should shrink, not one branch's diff
ceh-orchestration-lab:orchestrate	/ceh-orchestration-lab:orchestrate [model] [task]	Experimental: run a coding task as a supervised loop, with cheaper implementer subagents doing every edit
ceh-orchestration-lab:plan-then-implement	/ceh-orchestration-lab:plan-then-implement [model] [task]	Experimental: run a coding task as one complete plan handed to one cheaper implementer subagent

Output

| #   | Suggestion                  | Why now                               | Start it                          |
| --- | --------------------------- | ------------------------------------- | --------------------------------- |
| 1   | `<plugin>:<skill or agent>` | <the fact read that makes it fit now> | `/<plugin>:<skill> <args>` or ask |

Read: <what the advice rests on, one line>. Assumed: <the goal, if it was assumed>.

For an agent, "Start it" is the request that makes Claude dispatch it, such as "ask the bulk-reader agent which files call X". For a plugin from step 4, "Suggestion" is <plugin> (not installed) and "Start it" is the install command plus /reload-plugins.