goallisted
Install: claude install-skill millerlai/claude-all-in-one
> `<cai>` is the cai-codex command line that `$setup` wrote into your instructions (the cai-codex block in AGENTS.md). `<cai-root>` is what `<cai> --root` prints.
Take this design/plan doc from requirement to verified implementation: $ARGUMENTS
Step 1 always runs, and can end the run two ways — its gate fails, sending
you to Step 2, or its findings outlast three rounds. Otherwise Step 3 picks
a lane and Step 4 closes out, at tiers the skills set, not anything here.
## Step 1 — Review the design document
Decide the skeleton first: a document carrying a `## Work breakdown` or
`## Implementation spec` heading is `detail`, otherwise `hld`. Invoke the
`plan-review` skill against the doc, passing both the doc and that skeleton.
The second argument is not optional — `plan-review` runs lens 8 first for a
detail design (`plugins/cai/skills/plan-review/SKILL.md:71-74`), and without
it that rule never fires.
- Blocker/Major findings that are objective errors in the doc → fix the doc
directly, then re-run `plan-review`'s Step 1 on the fixed version, per its
own "Folding it back" rule.
- Section 4 ("requirement decisions to confirm") → surface these with
the question tool and wait for an answer. Never resolve them yourself —
that section exists specifically so scope decisions aren't made silently.
- **At most three rounds of fix-and-recheck.** Findings open after the third
end the run — report them rather than implementing against them.
- If `plan-review`'s Step 0 gate fai