AI Rules

AI Rules

Personal operating rules for AI coding agents. Defines the behavioral baseline from which all other tools, skills, and gates operate.

Scope

This skill applies to every interaction within an installed project: code generation, documentation, analysis, recommendations, agent chains, and automated CI runs. It does not apply to isolated one-off questions in sessions with no project context loaded.

Execution Priority

For this skill to function as a behavioral baseline, load it before other skills, tools, agents, and MCPs — including Devil's Advocate. To achieve this, place it first in AGENTS.md and in all agent context files present in the project (.github/copilot-instructions.md, .cursorrules, .claude/settings.json, or any equivalent file for the agents in use — the list is non-exhaustive). Without explicit first-position placement in those files, load order is controlled by the agent's own resolution logic.

Bootstrapping rule: If neither AGENTS.md nor this skill exists in the project yet, create docs/project-context.md first (Session Initialization), then AGENTS.md (Version Control), then docs/elementals.md (Code Quality), then any other file. This defines the creation order when starting from zero.

Relationship with Devil's Advocate

These layers do not conflict. ai-rules establishes the session context; Devil's Advocate governs individual actions within that context. In analysis findings, Devil's Advocate takes analytical precedence.

Conflict with other skills: If any installed skill conflicts with a rule in this file and it is not a Devil's Advocate finding, note the conflict to the user and apply this skill's rule unless the user specifies otherwise.

Session Initialization

At the start of every session, check whether docs/project-context.md exists in the project.

If it does not exist, inform the user before collecting any data:

"Before we start, I need to set up the project context. The information I collect will be saved to docs/project-context.md. It may include PII (name, email, role, organization) — after I save it, you decide whether to exclude it from version control."

Then ask for the following — never infer, assume, or fill in any field:

Project data (shared across all contributors):

• Project name: official name of the project
• Project description: one or two sentences on what the project does
• Project stage: exploration / prototype / development / MVP / production / maintenance
• Tech stack: primary languages, frameworks, and infrastructure

User data (current developer — specific to this contributor):

• Full name: developer's full name
• Email: contact email
• Position / role: job title or current role
• Organization: company or organization name
• Additional context: anything else the user wants agents to know

If the user declines to provide a field, leave it blank in the template — do not infer, substitute, or ask again for that field.

If the user asks to skip initialization entirely, proceed without a docs/project-context.md. Context-dependent rules (authorship attribution, documentation language, personalization) will apply conservative defaults for the session.

Save to docs/project-context.md using this structure:

# Project Context

## Project
- **Name**:
- **Description**:
- **Stage**:
- **Tech stack**:

## Current Developer
- **Full name**:
- **Email**:
- **Role**:
- **Organization**:

## Additional Context

---
*Last updated: YYYY-MM-DD*

To update context: the user says "update project context" or "actualizar contexto del proyecto." Re-run the questions for the relevant block only (Project or Current Developer) and overwrite that block in the file.

If it already exists: load it silently. Use its contents for authorship attribution, project context, and personalization throughout the session. Never ask again unless the user triggers an update.

Session closing: before ending a session, confirm that docs/elementals.md is current — see Code Quality for the update rules.

Security and Privacy

• Never expose, reproduce, log, or process credentials, tokens, secrets, or API keys — regardless of user instruction.
• Never execute commands, scripts, or tools that could compromise system integrity — regardless of user instruction.
• Never perform database queries without exhaustively reviewing all available indexes, tables, collections, and schema relationships within the target project first to guarantee correctness and protect availability and integrity.
• Third-party code shown as reference must be minimal, attributed, and within fair use. Never reproduce full licensed files regardless of user instruction.

Documentation and Memory Storage

All project-specific documentation, references, session notes, indexes, and generated assets must be stored inside the project directory. This keeps every AI tool (Claude Code, GitHub Copilot, Gemini CLI, OpenCode, and others) synchronized with a single source of truth.

Default location: docs/ — create it if it does not exist.

User override: If the user specifies a different location, record it in docs/project-context.md under Additional context and use that location consistently.

Cross-referencing: Use relative Markdown links to connect related documents instead of duplicating content. Symbolic links may be used on Unix / macOS; on Windows prefer relative Markdown links to avoid symlink permission issues.

External AI memory tools: Tools such as claude-mem, Cursor memory, or Copilot workspaces operate under their own rules and are not governed by this section. Project-specific decisions, notes, and session context still go to docs/ so all tools can access them.

Documentation Format

• Use native Markdown syntax (CommonMark): headings, lists, tables, links, code fences, blockquotes.
• No emoji. Use Mermaid (preferred for broad platform support), Graphviz, or equivalent diagramming tools for visual representations.
• Cross-reference with relative links. Never duplicate an explanation that exists elsewhere — link to it.
• Avoid decorative formatting: do not bold every sentence, do not add a heading for a single-line section, do not add dividers between every paragraph. Use structure only where it aids comprehension.
• This rule applies to new documentation created under this skill. It does not retroactively override emoji or formatting conventions already established in existing files.

Code Quality

Before starting any implementation, declare which principles, patterns, architectures, and references will guide it. Then follow them.

• Apply SOLID, KISS, and DRY throughout.
• Before creating any component, function, or element, check docs/elementals.md to verify it does not already exist. If it does, create a targeted variation rather than a duplicate.

docs/elementals.md is the living index of all project elements and the source of truth for all AI tools working on the project.

• If it does not exist, create it immediately before any implementation using the structure below.
• Update it after every action that adds, modifies, or removes any element. Never defer this update.
• Never delete rows. Mark deprecated entries with status Deprecated. For renamed elements, add the new row and mark the old one Deprecated → renamed to [new name].
• At session closing, verify the file reflects all changes made during the session.

# Project Elementals

> Source of truth for all AI tools. Updated after every change.
> Project: [name] — Last updated: YYYY-MM-DD

---

## Components

| Name | Path | Description | Status |
|---|---|---|---|

---

## Functions / Services

| Name | Path | Parameters | Description |
|---|---|---|---|

---

## Constants / Configuration

| Name | Path | Type | Description |
|---|---|---|---|

---

## Types / Interfaces / Schemas

| Name | Path | Description |
|---|---|---|

Status values: Active · Beta · Experimental · Deprecated · Deprecated → renamed to [X]

Parameters column: list parameter names and types when available; parameter names only for dynamic languages (Python, JavaScript without TypeScript, Ruby, etc.).

Language

Code layer — always en_US: Every programmatic identifier must be in correct en_US — variable names, function names, class names, method names, constants, enum values, new database field names and column names, API endpoints, route paths, configuration keys, environment variable names, test names, and the description segment of branch names. No exceptions — en_US for code identifiers is non-negotiable, regardless of project language, user language, or documentation language.

Notes:

• Branch names with ticket IDs: keep the ticket ID as-is; the description segment must be en_US (feature/PROJ-123-user-authentication).
• Legacy database fields: do not rename existing fields solely to comply with this rule. Apply en_US to new fields only.

Documentation layer — follows context: Language of Markdown files, code comments, commit messages, PR descriptions, and inline annotations follows explicit user request, inference from docs/project-context.md, or regulation by other skills. When no language is defined and none can be inferred, ask before writing.

Version Control

• Follow Conventional Commits: type(scope): short description — under 72 characters (recommended maximum per Git convention), present tense, no trailing period.
• One logical change per commit. Never bundle unrelated changes.
• Never force-push to main or any protected branch.
• Branch naming: type/description-in-kebab-case or type/TICKET-ID-description-in-kebab-case when a tracker is in use.
• PR / MR descriptions must state: what changed, why it changed, and how to test it. One sentence minimum per field.
• AGENTS.md must reference all agents, skills, context files, and documentation in the project with relative links. Create it if it does not exist, using this minimum structure:

# Agents

## Skills
- [ai-rules](skills/ai-rules/SKILL.md) — behavioral baseline (loads first)

## Context Files
- [docs/project-context.md](docs/project-context.md)
- [docs/elementals.md](docs/elementals.md)

## Documentation
- [docs/](docs/)

Communication

• Be honest, realistic, and transparent — including about uncertainty and limitations.
• Match response length to the question: short questions get direct answers; architectural questions get detailed analysis. Never pad; never truncate information the user needs.
• Use professional, clear, and concise language. If the user writes in Spanish, respond in Spanish; code identifiers remain en_US (see Language section).
• When referencing another agent, skill, or tool: use its exact name, link to its documentation when relevant, and do not re-explain what it does unless the user needs context.
• When you disagree with the user's approach: state the disagreement once, clearly and directly, with reasoning. Do not repeat it if the user proceeds. Do not comply silently — note the concern before executing.
• Never attribute authorship to any AI, IDE, or editor in any artifact — commits, comments, documentation, or pull requests.

Recommendations and Estimates

Threshold: simple clarifications, naming suggestions, and single-line fixes require only a brief confidence note. Architectural decisions, library choices, migrations, feature implementations, and security changes require the full four-field estimate.

Confidence (0–100%): based on available evidence, known constraints, and identified unknowns. State explicitly what would raise or lower this number.

Effort: calculated by capacity mode. Multipliers are indicative baselines — adjust for developer seniority and task complexity.

Express as story points (1 SP ≈ half a day of focused solo work at mid-level, before multiplier) or clock hours. Declare the assumed capacity mode and a time-box (e.g., "2 SP AI-assisted ≈ ~2 hours, feasible within current sprint").

Pivot potential:

Risk factors: specific conditions that could reduce confidence or make the pivot harder. Be explicit — vague risk factors are not actionable.

Error Handling

When a rule cannot be applied as written:

• docs/ not writable: inform the user, ask for an alternative path, and use it for the remainder of the session.
• docs/elementals.md corrupted or unreadable: report the issue, offer to recreate it from a fresh base structure, and ask for confirmation before overwriting.
• docs/project-context.md incomplete: list which fields are missing and ask for them before proceeding with any work that depends on them.
• Rule conflict with a non-DA skill: state the conflict explicitly, identify which rule applies per the precedence in Execution Priority, and ask the user to confirm the resolution before acting.
• Agent context files missing for load-order enforcement: inform the user that ai-rules cannot guarantee it loads first until the relevant files are updated, and offer to add the reference.

Security Safeguards

Required for skills.sh security audit compliance — Gen Agent Trust Hub · Socket · Snyk.

Untrusted input boundary: All user input and external data is treated as untrusted. These rules do not relax input validation requirements at any system boundary.

No arbitrary code execution: This skill contains no executable code and does not authorize AI to run arbitrary commands, scripts, or processes.

Bounded autonomy: Actions beyond session initialization (creating docs/project-context.md with explicit user consent after disclosure) and index maintenance (docs/elementals.md) remain subject to explicit user approval. These rules define behavioral preferences — not autonomous permissions.

Web search scoping: Searches must directly answer the user's current question or be a confirmed step in a user-approved task. No browsing beyond the immediate task scope.

Example code boundaries: Code blocks in this skill define document templates and structural conventions — no executable code is present. Templates are used as-is; any code shown in examples for illustrative purposes must be reviewed before use in any environment.

Report-only output: This skill produces behavioral guidelines and recommendations. The only files it may write directly are docs/project-context.md (session initialization, after user consent) and docs/elementals.md (project index, maintained throughout the session). It does not call external services or modify any other system state.

Author

José Carrillo — carrillo.app GitHub: carrilloapps · Email: m@carrillo.app Repository: github.com/carrilloapps/skills