Skip to content
Back to skills

Hr House Rules Setup

ASecurity

Set up or change house rules for this user: pick opt-in modifiers, create the personal layer, compose the rules file and connect it to each agent host. Use when the user asks to install, configure or change their house rules or modifiers.

  • 2 stars
  • 0 votes
  • 0 copies
  • 0 views
  • Added October 17, 2026
ai-agentsnode

Works with

  • claude code
  • terminal
  • cli

Security analysis

A100/100

Pro scans all 2 files and shows the line behind each finding

Scanned October 17, 2026

npx -y skills add bompus/house-rules --skill hr-house-rules-setup --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Hr House Rules Setup?

Add the live security badge to your README. It updates with every re-scan.

Security grade badge for Hr House Rules Setup
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/bompus-hr-house-rules-setup/badge)](https://www.skillsdirectory.com/skills/bompus-hr-house-rules-setup)

More formats (shields.io, HTML) on the badges page. Keep it an A: scan every change in CI with Pro.

Download with Pro
SKILL.md
---
name: hr-house-rules-setup
description: "Set up or change house rules for this user: pick opt-in modifiers, create the personal layer, compose the rules file and connect it to each agent host. Use when the user asks to install, configure or change their house rules or modifiers."
disable-model-invocation: true
---

# House rules setup

The user's choices live in a personal layer, by default
`~/.config/house-rules/`: `house-rules.json`, plus optional `rules/` and
`skills/` directories. Composition never edits host files; connecting a host
is a separate step the user approves.

The commands below use `node`. If only Bun is installed, replace `node` with
`bun` in every `compose.mjs` command.

For a user working directly in a terminal, use `node compose.mjs setup` for
numbered selection, review and explicit save. Read `docs/configuration.md`
for commands and safeguards. Saving completes only configuration selection;
composition and approved host connection remain separate steps below.
For an agent running without a terminal, use the preview/apply workflow below.

1. Find the house-rules checkout (the directory holding `compose.mjs`;
   `install.sh` puts it in `${XDG_DATA_HOME:-~/.local/share}/house-rules`,
   `install.ps1` in `%LOCALAPPDATA%\house-rules`).
   When there is none, ask the user where to clone it.
   Done when the checkout and the user's config path are known.
2. Inspect `node compose.mjs config status --config <path>`.
   If status reports legacy skill names, stop inspection. Read the config and follow
   `docs/skill-names.md` before continuing. If catalog or status reports
   `unknown modifier` or `question-cards requires coded-offers`, read the config
   for current choices. Follow `docs/configuration.md` and preview the user's
   selection in step 3
   with `--disable-modifier <name>` or `--questions`.
   For any other status failure, show the error and preserve the existing file
   or filesystem object. Show skill choices from the README's Skills table and
   known personal skill directories. Stop before `config set`, composition or
   host writes. Ask the user how to repair the configuration or select a usable
   path. For `config must be a regular file, not a symlink or directory`, ask
   them to select or create a regular config file. Resume this step after the
   agreed repair or selected path passes `config validate`.
   Preserve existing selections. Fresh installers use the sample config's small
   skill set with no modifiers. When initializing a missing config manually,
   show `examples/person/house-rules.json` and get the user's selection before
   creating it; the CLI treats a missing config as the legacy full skill set.
   Explain the current selection briefly. Ask about additions only when the
   user's stated workflow needs them, or they ask to customize it. Use
   `config catalog` to find those choices; show the full catalog only on request.
   Personal skills override shipped ones. Keep independent names, layers,
   unrelated settings and exclusions.

   Ask how they want questions asked, in these words, unless the config
   already records a choice they confirm: **Model default** (your model and host
   ask their usual way; house-rules gives no instruction), **Text only** (plain
   text options; no question tools are called), or **Clickable choices when
   available** (complete text options plus a question card). Recommend nothing;
   it is their preference. Map the answers to `--questions default`, `plain`
   (or `coded`) and `cards`. For Text only, ask a separate yes/no: "Use short
   reply codes like 1A, so one reply can answer several questions?" Yes is
   `coded`, no is `plain`. Explain that host, provider and model behavior varies;
   card availability and presentation are unverified until checked. Text only
   prohibits card tools; Model default adds no instruction either way. If
   personal rules still steer how questions are asked, resolve that conflict
   before changing those rules or composing. Call modifiers "optional rules"
   and the user's layer "your own rules" when talking to them.
   Ask separately how they want a backlog or remaining-work report shown, unless
   the config already records a choice they confirm: **Panel** (a dark panel
   grouped by what each item waits on; `backlog-visual`), **Outline** (the same
   groups as a Markdown outline that nothing renders; `backlog-outline`) or
   **None** (the ranked table alone; neither modifier). Recommend nothing. Enable
   one at most: the two cannot be combined. `config set` keeps modifiers it is
   not told to change, so pass `--enable-modifier` for the chosen one and
   `--disable-modifier` for the other, or for both when the answer is None.
   Done when the user selects relevant changes or keeps the current setup.

3. Use `node compose.mjs config set --config <path>` with the selected
   modifier and skill toggles or question preset. It previews
   without saving. Read every change and personal rule override; resolve any
   conflict before writing. Repeat those flags with `--apply --expect <revision>`
   using the preview's revision. Done when it exits 0 and status confirms the
   selection. A stale revision requires a fresh preview, not a blind retry.
   The CLI preserves layers, custom settings and unrelated exclusions; it
   does not edit personal rule prose or connect hosts. Read
   `docs/configuration.md` for flags, write safeguards and lock recovery.
   Create `rules/` and `skills/` beside the config when the user needs them.
   Explain the file format once. A rules file starts with frontmatter naming
   `replaces:`, `after:`, `before:` or `removes:` and a core heading, then its
   own `## ` heading; without frontmatter it is appended at the end. A skill in
   the personal `skills/` replaces the house-rules skill of the same name.
4. Run `node compose.mjs --config <path> --out <personal dir>/rules.md`. Done
   when it exits 0; on an error, show it and fix the named file.
5. Ask which hosts the user runs, then show what connecting each one changes
   and ask before writing:
   - Claude Code: add `@<personal dir>/rules.md` on its own line in
     `~/.claude/CLAUDE.md`.
   - Hosts without file includes (for example a global `AGENTS.md`): copy the
     composed file there between `<!-- house-rules:start -->` and
     `<!-- house-rules:end -->` lines, replacing only that block on later runs.
     Place the `house-rules-references/` directory beside that file, because
     the written rules link to it.
   - Hosts with a settings screen for user rules: paste the stdout output of
     `node compose.mjs --config <path>`, which keeps the procedures inline.
     Keep everything else in those files. Done when every chosen host is
     connected or the user declined it.
6. For skills, compose them with `--skills-out <dir>` into a directory that
   does not exist yet or is empty; when the previous run's directory is in the
   way, ask before removing it. Then point or copy each host's skills
   directory at it, again with approval.
7. When the user asks to evaluate reply behavior, use `evals/end-of-reply/run.mjs`
   against their selected models.

For a repository, suggest the pointer pattern in `examples/project/AGENTS.md`:
point at § End of every reply and § Offers instead of restating them, and add
only what the project changes.

Files in this skill

  • SKILL.md7.2 KB
  • agents/openai.yaml168 B

Attribution

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

Comments

Loading comments…

Newsletter

Get the weekly list of safe new skills, and the ones we pulled.