Review 360
Review a plan, a pull request, or an existing system from the person's job through the architecture to the code.
Review 360 connects three questions:
- Can someone finish the job, recognize success, and recover when something goes wrong?
- Is the proposed mechanism necessary, and what can be removed while preserving its guarantees?
- Can a new reader explain each abstraction, its ownership, and its behavior?
Use all three perspectives for a system review, or select one. Plans receive prospective assessments. Existing behavior is checked against execution or source, with those forms of evidence kept distinct.
Install
With the skills CLI:
npx skills add pierrebrunelle/review-360 --skill review-360
The installer helps you choose a supported agent and installation scope. Install one copy per agent to avoid duplicate discovery. To inspect the package without installing:
npx skills add pierrebrunelle/review-360 --list
For an agent with filesystem-based skill discovery, copy skills/review-360/ into its documented skill directory. Keep the folder and its references together. The skill needs no other installed skill.
The root plugin.json uses the Agent Plugins format. Clients supporting that format can discover the same skill without a separate content tree. Installation and permissions remain specific to the client.
In Claude Code:
/plugin marketplace add pierrebrunelle/review-360
/plugin install review-360@pierrebrunelle-plugins
The Claude manifest and marketplace are metadata adapters. They use the same skills/ directory as the portable package.
Use
Ask your agent to use review-360 with a target and, optionally, a mode:
Use review-360 all on this proposal before we commit to it.
Use review-360 all on PR 123, including the affected product flow.
Use review-360 code on the storage package; examine ownership and cleanup.
Use review-360 design on this PR stack. What can disappear safely?
Use review-360 product walk on the CLI's first export.
Use review-360 product trouble on the empty and return states.
Use your agent's native invocation syntax when it has one. A Claude plugin skill is /review-360:review-360. Ordinary text works when the agent can read and follow a skill folder.
| Mode | Purpose |
|---|---|
all | Connect product usefulness, architecture, and code in one assessment |
product | Walk a job, establish a card, test trouble, research, critique, or write a decision brief |
design | Challenge requirements and mechanisms while preserving behavior and guarantees |
code | Review correctness, abstractions, ownership, names, and maintainability |
Product commands are cast, walk, trouble, research, critic, and brief. Notes stay in the response unless you request persistent artifacts or continue an existing file workflow. The default file location, when needed, is lifecycle/ in the target project.
A focused request reads only its guidance. A full review gathers evidence once and connects the perspectives where the evidence supports it. A whole-codebase request reports inspected paths and coverage limits. Review alone does not edit the product or publish comments. An explicit request to fix findings authorizes those fixes within its stated scope.
Portability and cost
The installed skill contains Markdown instructions and local references. It has no executable, model selection, required subagent, hook, MCP server, or runtime service. It follows the target project's conventions and uses the tools available in the current session. Missing access is reported as a verification limit.
One discovery entry points to guidance loaded by mode. Repository documentation, validation tools, and evaluation fixtures are outside the skill folder. They are not required for a review.
Validation records packaging checks, context estimates, behavioral checks, and their limits. Smaller files do not establish faster execution or equal review quality. Native loading and permissions vary by harness; the portable format is not a claim of identical behavior everywhere.
Maintain
Run the dependency-free package checks with Python 3.9 or newer:
python3 tools/validate.py
For an external structural audit, use AgentForge:
agentforge audit --root skills
agentforge lint skills/review-360 --fail-under 80
AgentForge is optional development tooling. Its scores and routing similarities are heuristics. Use the evaluation cases to check whether a revision preserves the actual judgments before relying on a budget reduction.
Origins
Created by Pierre Brunelle from walk-the-job, challenge-the-design, and marcel-review. The code guidance retains Marcel Kornacker's emphasis on coherent abstractions and the cost of reading code. The package contains newly written general guidance, without the original Pixeltable-specific rules or review-comment archive.
Pstack, by Lauren Tan, informed the packaging of a coherent engineering practice. AgentForge informed selective loading, structural inspection, and validation boundaries. Neither is an installed dependency.
Licensed under MIT.