How to use the CloudAnalyzer Command-Line Harness agent skill

What the skill lets the agent do

Once the skill is installed, the agent can operate the cli-anything-cloudanalyzer command-line harness. The harness exposes 27 commands across 8 groups: point cloud evaluation, trajectory evaluation, ground segmentation QA, config-driven quality gates, baseline evolution, processing, visualization, and an interactive REPL.

The source describes the evaluations as Chamfer distance, F1, AUC and Hausdorff for point clouds, and ATE, RPE, drift, lateral and longitudinal error for trajectories. The skill also asks for JSON output so the agent can read results.

• Evaluate a point cloud against a reference (evaluate run)
• Compare or diff two point clouds, with optional ICP or GICP registration (evaluate compare, evaluate diff)
• Batch-evaluate a folder of point clouds against one reference (evaluate batch)
• Evaluate trajectories and apply limits such as --max-ate (trajectory evaluate)
• Run a config-driven quality gate from a YAML file (check run, check init)
• Save QA summaries to a history folder and decide whether to promote, keep or reject a baseline (baseline save, baseline decision)
• Downsample, sample, filter, split, merge and convert point cloud files (process group)

Install the skill (ZIP or Claude apps upload)

Before installing the skill, make sure you have Python 3.10 or newer. The skill also needs the CloudAnalyzer tool itself, which the source lists as the pip package cloudanalyzer. The harness is the pip package cli-anything-cloudanalyzer.

To add the skill to Claude Code, download the ZIP from this page and unzip it into ~/.claude/skills/ for a personal install, or into a project's .claude/skills/ folder to share it with one project. Alternatively, in the Claude apps, open Settings, go to Capabilities, then Skills, and upload the ZIP there.

• Requirement: Python 3.10+
• Requirement: pip package cloudanalyzer (CloudAnalyzer)
• Requirement: pip package cli-anything-cloudanalyzer (the harness the skill calls)
• Skill file in the ZIP: SKILL.md

How to trigger the skill with natural requests

You do not need to type command names. Describe the QA task in plain words that name your files and the check you want, such as the reference cloud, the trajectory files, or a pass/fail threshold. The agent should pick the matching command group from the skill.

Requests work best when they name the input files and the limit that matters. Examples below are phrasings you can try; the commands the agent runs come from the skill's documented workflows.

• Evaluate output.pcd against reference.pcd and tell me the Chamfer distance and F1.
• Check whether est.csv meets a max ATE of 0.5 m and max lateral error of 0.3 against gt.csv.
• Run the QA config cloudanalyzer.yaml and tell me if it passes.
• Compare the ground segmentation in est_ground.pcd with ref_ground.pcd and fail if F1 is below 0.9.
• Should the candidate summary in qa/summary.json be promoted as the new baseline?

Worked examples from the skill

The skill documents four typical workflows. The first evaluates and gates one point cloud with JSON output. The second generates a starter config and runs it. The third runs QA, saves the summary to a history folder, and asks for a baseline decision. The fourth checks ground segmentation against a minimum F1.

The --json flag goes before the command name. Options such as --min-f1 are added after the command's file arguments.

What you need before running it

The commands need local point cloud or trajectory files to do anything useful. The source names .pcd and .las point cloud files and CSV trajectory files. The check init command accepts the profiles mapping, localization, perception and integrated.

The source lists the skill as running scripts and not using the network. Use the skill only on data you are allowed to process locally.

• Python 3.10 or newer
• CloudAnalyzer installed via pip (cloudanalyzer)
• Local .pcd or .las point cloud files and CSV trajectory files
• Reference data to compare against

Safety notes: scripts, local processing and file writes

The skill runs scripts on your machine. The tool executes local processing, and some commands open viewers or write files. Commands such as process split write output folders and metadata, and evaluate pipeline and process commands write output files named with -o.

Review the data and the command an agent plans to run, especially on sensitive datasets. Keep the global --project option pointed at a project file you trust.

• Some commands open a point cloud viewer (inspect view)
• Output options (-o) and project files write to disk
• Review sensitive datasets before letting an agent run batch jobs
• The source lists no network use for this skill

Tips for reliable results

Ask for JSON output so the agent can parse results. Use check run with a config file when you want one pass/fail decision across several metrics. Use baseline save with --keep to limit how many history entries are kept.

• Use --json before the command name when an agent must read the result
• Set thresholds explicitly, for example --min-f1 0.9 or --max-ate 0.5
• Use session new to create a project file and session history to review recent operations
• Run info show on a point cloud first to check its metadata