dual-axis-skill-reviewer — скилл для Claude Code: обзор и установка

---
name: dual-axis-skill-reviewer
description: "Review skills in any project using a dual-axis method: (1) deterministic code-based checks (structure, scripts, tests, execution safety) and (2) LLM deep review findings. Use when you need reproducible quality scoring for `skills/*/SKILL.md`, want to gate merges with a score threshold (for example 90+), or need concrete improvement items for low-scoring skills. Works across projects via --project-root."
---

# Dual Axis Skill Reviewer

Run the dual-axis reviewer script and save reports to `reports/`.

The script supports:
- Random or fixed skill selection
- Auto-axis scoring with optional test execution
- LLM prompt generation
- LLM JSON review merge with weighted final score
- Cross-project review via `--project-root`
- Non-scoring display of `skills-index.yaml` production verification declarations

## When to Use

- Need reproducible scoring for one skill in `skills/*/SKILL.md`.
- Need improvement items when final score is below 90.
- Need both deterministic checks and qualitative LLM code/content review.
- Need to review skills in a **different project** from the command line.

## Prerequisites

- Python 3.9+
- `uv` (recommended — auto-resolves `pyyaml` dependency via inline metadata)
- For tests: `uv sync --extra dev` or equivalent in the target project
- For LLM-axis merge: JSON file that follows the LLM review schema (see Resources)

## Workflow

Determine the correct script path based on your context:

- **Same project**: `skills/dual-axis-skill-reviewer/scripts/run_dual_axis_review.py`
- **Global install**: `~/.claude/skills/dual-axis-skill-reviewer/scripts/run_dual_axis_review.py`

The examples below use `REVIEWER` as a placeholder. Set it once:

```bash
# If reviewing from the same project:
REVIEWER=skills/dual-axis-skill-reviewer/scripts/run_dual_axis_review.py

# If reviewing another project (global install):
REVIEWER=~/.claude/skills/dual-axis-skill-reviewer/scripts/run_dual_axis_review.py
```

### Step 1: Run Auto Axis + Generate LLM Prompt

```bash
uv run "$REVIEWER" \
  --project-root . \
  --emit-llm-prompt \
  --output-dir reports/
```

When reviewing a different project, point `--project-root` to it:

```bash
uv run "$REVIEWER" \
  --project-root /path/to/other/project \
  --emit-llm-prompt \
  --output-dir reports/
```

### Step 2: Run LLM Review
- Use the generated prompt file in `reports/skill_review_prompt_<skill>_<timestamp>.md`.
- Ask the LLM to return strict JSON output.
- When running inside Claude Code, let Claude act as orchestrator: read the generated prompt, produce the LLM review JSON, and save it for the merge step.

### Step 3: Merge Auto + LLM Axes

```bash
uv run "$REVIEWER" \
  --project-root . \
  --skill <skill-name> \
  --llm-review-json <path-to-llm-review.json> \
  --auto-weight 0.5 \
  --llm-weight 0.5 \
  --output-dir reports/
```

### Step 4: Optional Controls

- Fix selection for reproducibility: `--skill <name>` or `--seed <int>`
- Review all skills at once: `--all`
- Skip tests for quick triage: `--skip-tests`
- Change report location: `--output-dir <dir>`
- Increase `--auto-weight` for stricter deterministic gating.
- Increase `--llm-weight` when qualitative/code-review depth is prioritized.

## Output

- `reports/skill_review_<skill>_<timestamp>.json`
- `reports/skill_review_<skill>_<timestamp>.md`
- `reports/skill_review_prompt_<skill>_<timestamp>.md` (when `--emit-llm-prompt` is enabled)

When the target project has a `skills-index.yaml` entry with a complete `verification` block, JSON
and Markdown reports also show its declared axes, `not_verified` gaps, non-applicable axes, and
`all_applicable_axes_passed`. This section is informational only. It is excluded from auto/LLM/final
scores and does not replace a live high-severity issue check.

## Installation (Global)

To use this skill from any project, symlink it into `~/.claude/skills/`:

```bash
ln -sfn /path/to/claude-trading-skills/skills/dual-axis-skill-reviewer \
  ~/.claude/skills/dual-axis-skill-reviewer
```

After this, Claude Code will discover the skill in all projects, and the script is accessible at `~/.claude/skills/dual-axis-skill-reviewer/scripts/run_dual_axis_review.py`.

## Resources

- Auto axis scores metadata, workflow coverage, execution safety, artifact presence, and test health.
- Auto axis detects `knowledge_only` skills and adjusts script/test expectations to avoid unfair penalties.
- LLM axis scores deep content quality (correctness, risk, missing logic, maintainability).
- Final score is weighted average.
- If final score is below 90, improvement items are required and listed in the markdown report.
- Script: `skills/dual-axis-skill-reviewer/scripts/run_dual_axis_review.py`
- LLM schema: `references/llm_review_schema.md`
- Rubric detail: `references/scoring_rubric.md`