# Kiro Review
> Get an INDEPENDENT second opinion from the Kiro CLI run strictly READ-ONLY — a separate reviewer runtime over the real diff, with no write access to the repo. Reads the real diff (branch, a commit, or uncommitted) to build a focused prompt, optionally injects the stack's convention skills (`--skill`, since kiro can't load them itself), then launches kiro via the bundled helper with trust scoped to `fs_read,execute_bash` so it can grep/git to verify findings but never edit. Never uses `--trust-all-tools` (the auto-mode classifier dead-ends read-only tasks on it) or hand-rolls the prompt through the shell; keeps secrets out, treats findings as candidates to verify, and reports an explicit clean result. Use only on explicit user intent or when a user-invoked workflow composes it as the kiro reviewer.
- Skill: `ulpi-io/kiro-review` (Agent Skill, multi-file: 2 files)
- Install (CLI): `npx skillmds add ulpi-io/kiro-review`
- Raw SKILL.md: https://api.skillmd.com/api/skills/ulpi-io/kiro-review/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: ulpi-io (https://skillmd.com/u/ulpi-io)
- Updated: 2026-09-09
- Page: https://skillmd.com/skills/ulpi-io/kiro-review
---
This skill orchestrates an external reviewer and must stay disciplined.
Non-negotiable rules:
1. Read the real diff before writing the Kiro prompt.
2. Make the prompt specific to the changed areas and likely risks.
3. Never put secrets or credentials in the prompt.
4. Carry forward exclusion lists on later rounds.
5. Verify returned findings before acting on them.
6. A review is READ-ONLY: launch kiro via `helpers/run-kiro.sh --mode review` (scoped trust
`fs_read,execute_bash`). NEVER use `--trust-all-tools` / `-a` — the auto-mode classifier blocks
unrestricted trust on a non-mutating task, so all-tools just dead-ends.
# Kiro Review
## Inputs
- `$request`: Optional scope hint such as `last commit`, `uncommitted`, `auth focus`, or `round 2`
## Goal
Use `kiro-cli` to get an external review pass that:
- uses the right diff scope
- focuses on the actual change surface
- returns structured findings instead of generic commentary
## Step 0: Verify Kiro availability
Check:
- `which kiro-cli`
- `kiro-cli whoami` or the minimal auth check needed in this environment
If the CLI is unavailable or not authenticated, explain the blocker and stop.
**Success criteria**: Kiro can run successfully from the current repository.
## Step 1: Resolve review scope
Determine whether to review:
- the full branch
- uncommitted changes
- a specific commit
Read the diff summary and changed-file list first.
If there is nothing to review, stop and say so explicitly.
**Success criteria**: The review target is explicit and backed by a real diff.
## Step 2: Build the focused Kiro prompt
Create a compact prompt that includes:
- what changed
- the major risk areas
- any previously fixed issues to exclude on later rounds
- an instruction to verify findings against the actual code
- the expected compact output format
- if a stack/convention skill is installed for kiro (`.kiro/skills//`), pass it to the helper as
`--skill ` so kiro reviews against those conventions — kiro has no `Skill` tool to load it
itself, and one-shot `--no-interactive` runs can't rely on auto-activation
Avoid generic prompts. They produce weak results.
**Success criteria**: The prompt is specific to the change set rather than reusable boilerplate.
## Step 3: Run Kiro read-only via the helper
Do NOT hand-roll the kiro command (mktemp/heredoc/trust flags broke before). Instead:
1. **Write the prompt to a file with the Write tool** (literal bytes — no shell, no heredoc, no mktemp).
Use an absolute path, e.g. `/tmp/kiro-review-prompt.txt`, and write a FRESH file each round.
2. **Run the helper in review mode:**
```bash
bash /helpers/run-kiro.sh --mode review [--skill ]... [--agent --model auto] --prompt-file /tmp/kiro-review-prompt.txt 2>&1
```
Optional: `--agent ` runs kiro as a kiro-native reviewer agent (`.kiro/agents/.json` with a
valid `model` + native `fs_read`/`execute_bash` tools — see the model list via `kiro-cli chat --list-models`);
`--model auto` overrides an agent that pins an invalid model.
`--mode review` scopes trust to **`--trust-tools=fs_read,execute_bash`** — kiro can read files and run
`grep`/`find`/`git` to verify findings against source, with NO write access. Do NOT use
`--trust-all-tools` / `-a`: a review is read-only and the auto-mode classifier blocks unrestricted
trust on non-mutating tasks. The helper feeds the prompt via kiro's stdin, refuses to launch on an
empty prompt, and captures a git baseline.
**Models** — pass `--model `. Valid values: `auto` (default), `claude-opus-4.8` / `4.7` / `4.6`,
`claude-sonnet-4.6` / `4.5` / `4`, `claude-haiku-4.5` (run `kiro-cli chat --list-models` to refresh).
Claude-Code names like `opus` / `sonnet` are NOT valid — kiro errors `model 'opus' is not available`.
Use a Bash timeout of 600000 ms for a large review. If the helper exits with "prompt file is EMPTY",
re-write the prompt and retry — never run kiro on an empty prompt.
**Success criteria**: Kiro runs read-only on the intended scope and returns parseable findings.
## Step 4: Summarize findings
Report:
- review scope
- findings by priority
- file and line references when available
- explicit clean result when no material findings are returned
If the user wants fixes, verify each finding locally before changing code.
**Success criteria**: The user gets a readable review summary instead of raw CLI logs.
## Step 5: Iterate only with exclusions
On later rounds:
- list prior fixed findings in the exclusion block
- narrow the scope to newly changed files when possible
- avoid repeated full-branch reviews unless the code changed broadly again
**Success criteria**: Follow-up rounds target new issues instead of recycling old ones.
## Guardrails
- Do not run this skill PROACTIVELY on your own initiative — only on explicit user intent (e.g.
`/kiro-review`) or when an explicit user-invoked workflow composes it as the kiro reviewer (e.g.
`/ship-playbook` with a kiro review role). It is no longer `disable-model-invocation`, so workflows
can call it; that is not license to run it unprompted.
- Do not put secrets, tokens, or private config into the prompt.
- Do not trust findings blindly without local verification.
- Do not skip diff reading before prompt construction.
- Launch kiro only via `helpers/run-kiro.sh --mode review` (scoped read-only trust); never
`--trust-all-tools`/`-a` for a review, and never hand-roll mktemp/heredoc to pass the prompt.
## Output Contract
Report:
1. the review scope
2. the main focus areas given to Kiro
3. findings by priority with locations when available
4. explicit clean result if nothing material was found
5. whether a next round should exclude previously fixed issues