Install KLOD

Choose the setup that fits your project. KLOD works with Spec Kit and OpenSpec, and the two portable skills can also be installed with APM.

Using Spec Kit or OpenSpec?

• Set up the Spec Kit extension to add KLOD planning and review commands to your existing workflow.
• Set up the OpenSpec custom schema to include human operation, implementation tasks and verification in your changes.

These integrations are maintained by KLOD, independently of both projects. Each setup guide includes installation, first-use instructions and the scope of compatibility checks.

Use the portable skills

A skill is a set of Markdown instructions for a coding assistant such as Claude Code, Codex or Cursor. KLOD provides two. Installing them adds files to your project's skills directory. You then ask your assistant to use a skill to review the project or change code. The full text of each skill is published on this site, so you can read its instructions before using it.

Use the skills in the project you want to build or review. To apply KLOD without an assistant, start with the pattern and examples.

1. Install with APM

Install APM, the Agent Package Manager, with one of the commands below. If APM is already installed, continue to installing KLOD.

macOS or Linux with Homebrew

brew install apm

macOS or Linux without Homebrew

curl -sSL https://aka.ms/apm-unix | sh

Windows PowerShell

irm https://aka.ms/apm-windows | iex

Any system with Python

python3 -m pip install apm-cli

Install KLOD in your project

Open a terminal in the project folder and run:

apm install AlexandruGirlea/keepthelightson#v0.1.2

APM detects your assistant from the project files. To choose it explicitly, use --target:

apm install AlexandruGirlea/keepthelightson#v0.1.2 --target claude

See APM's targets matrix for assistant detection. Keep the release tag in the command to install a specific version.

For a shared project, add the dependency to apm.yml, run apm install, and commit apm.yml with apm.lock.yaml:

dependencies:
  apm:
    - AlexandruGirlea/keepthelightson#v0.1.2

2. Run the repository check

In your coding assistant, run /klod-check in Claude Code or $klod-check in Codex. You can also ask:

Use klod-check to review this project for AI dependencies and missing
manual paths. Write the findings and next actions in klod_report.md.

Open klod_report.md at the project root. The check reads the repository without changing application code or interrupting live services.

Build or change an AI feature

Run /klod in Claude Code or $klod in Codex, then describe the work:

Add invoice extraction from uploaded PDFs, with a queue people can review
and complete when AI is unavailable.

The skill helps you save the inputs, build the manual controls and record the capability in LIGHTS.md. For an off switch, ask:

Add an authorised AI off switch and manual takeover. Handle queued actions
and late results, and require a person to approve restart.

See the human-control guide for self-run AI and connected equipment.

3. Before you rely on a check

Any assistant that can read project files, follow the skill references, trace callers and write a Markdown report can run klod-check. For large repositories, work in stages and record the files and line numbers that support each finding. Before you act on a report, check that the assistant:

• loads the installed skill and its references;
• finds a known AI integration and traces it to the affected business capability;
• separates missing evidence from a confirmed defect;
• doesn't invent drill results or mark an accepted risk as fixed;
• writes klod_report.md with evidence you can open and actions you can assign.

Review the findings against your code and operating records. Try the skill on a small project you understand before relying on it for unfamiliar code, and check the evidence behind each finding.

4. After the report

1. Verify the cited files and collect any operating records held elsewhere.
2. Assign an owner and a date to each significant gap.
3. Ask klod to address one confirmed finding, using its ID from klod_report.md.
4. Arrange an authorised real-work drill with AI unavailable. Measure Time to Manual (how long until the first correct result without AI) and manual capacity.
5. Update LIGHTS.md and run klod-check again.

Read the example report and what the four levels mean.

Not using a coding assistant?

Use the self-check to review one capability in your browser and download an action list. It uses your answers; verify the records before declaring a level.