Next-step advisor

Read the project state and produce a short ranked recommendation of 1-3 actions. This skill never executes the recommended commands — it only suggests.

Phase 1: Detect state

Read in this order, skipping silently any source that does not exist:

1. .ezproject.json for mode and stack
2. INITIAL.md — exists? when was it last modified? does it still contain placeholders?
3. PRPs/*.md — active (un-archived) PRPs
4. PRPs/done/*.md — recently completed (last 30 days)
5. git status --short — uncommitted files
6. git rev-parse --abbrev-ref HEAD — current branch
7. git log --since="2.weeks.ago" --oneline — recent activity
8. git log -1 --format="%ar — %s" — last commit summary
9. CLAUDE.md last modification date — drift signal
10. grep -l "ai-patchlab\|2026-05-12\|\[REMPLIR\]" *.md — unfinished kickoff?

Phase 2: Decision tree (top-down, first match wins)

1. Placeholders remain in INITIAL/CLAUDE/ROADMAP/README → run kickoff to finish project setup
2. No INITIAL.md, fresh repo, on main → run kickoff to start a new feature interview
3. INITIAL.md exists, no PRP for it → run generate-prp INITIAL.md
4. Active PRP, never executed → run execute-prp PRPs/<name>.md in a NEW conversation (fresh context per ADR-013)
5. Active PRP with partial progress → resume execute-prp PRPs/<name>.md
6. PRP moved to PRPs/done/ within last 24h, no retrospective committed since → run retrospective last in a NEW conversation
7. Uncommitted changes on a feature branch → run review-code before committing
8. Feature branch ahead of main, tests passing, no active PRP → open a PR or run the release pipeline if available
9. 3+ weeks since CLAUDE.md modified AND 10+ commits since → run audit-project
10. 30+ days since pyproject.toml change, never ran a dependency check → run a dependency audit
11. Many TODO/FIXME comments accumulating → run a cleanup pass
12. On main, nothing in progress → list top 3 unchecked ROADMAP items and ask the user to pick, then run kickoff

Phase 3: Output

Produce a short report with two sections:

## Where you are
- Mode/Stack: ...
- Branch: ...
- Active PRP: ...
- Uncommitted: ...
- Last commit: ...
- Recent activity: ...

## Recommended next steps
1. [PRIMARY] <command with exact args> — <one-sentence reason tied to detected state>
2. [SECONDARY] <command> — <reason>
3. [LATER] <command> — <reason>

Rules

• Always print the literal command to run, with exact arguments.
• Tag each recommendation [PRIMARY] / [SECONDARY] / [LATER].
• Always remind that execute-prp and retrospective should run in fresh sessions (ADR-013).
• If state is genuinely ambiguous (e.g., 5 PRPs active), ASK the user to pick rather than guess.
• Never fabricate state — if a marker is missing or unreadable, say so and skip that branch.
• Keep output under 30 lines.
• This skill MUST NOT execute the recommended commands. It only suggests.