guidelines-init

Guidelines Init

Generate a PROJECT_GUIDELINES.md file that becomes the single source of truth for project conventions. This file is later consumed by /guidelines-review.

Arguments

$ARGUMENTS determines the mode:

Argument	Mode
(empty)	Simple — flat list of rules, minimal interview
--simple	Simple — same as empty
--full	Full — categorized guidelines with severity tags, comprehensive interview
--upgrade	Upgrade — converts an existing simple-mode file to full mode

---

Simple Mode Workflow

Default when no flag is provided. Optimized for getting rules on paper fast.

Phase 1 — Quick scan

Identify the project's tech stack and shape. This is for context only — not an exhaustive convention analysis.

• Read package.json, go.mod, Cargo.toml, or pyproject.toml (whichever exists) to identify the primary language and framework.
• Use Glob("src/*/") or Glob("*/") to understand directory structure.
• Note the test framework if obvious from config files.
• Check for monorepo signals: Glob("packages/*/package.json"), Glob("apps/*/"), lerna.json, pnpm-workspace.yaml, turbo.json.

Do NOT run the full signal file table or pattern detection. Keep this under 30 seconds.

Phase 2 — Ask for rules

Present a 2-3 sentence summary of what you detected:

"This looks like a TypeScript/React project using Jest for testing, organized by feature under src/features/."

Then ask a single open-ended question:

"What rules matter to you? List whatever comes to mind — coding conventions, architectural boundaries, things that annoy you in code review. I'll turn them into a guidelines file."

Accept the user's response as-is. Do NOT:

• Force answers into categories
• Ask follow-up questions for each category
• Suggest rules the user didn't mention
• Run a structured 8-category interview

If the user's list is very short (< 3 rules), it's fine to ask: "Anything else, or is that the full list?" — but only once.

Phase 3 — Generate

Write PROJECT_GUIDELINES.md with this format:

# Project Guidelines

<!-- mode: simple -->
<!-- Generated by guidelines-init | Last updated: YYYY-MM-DD -->

## Overview
[One paragraph: project description and tech stack]

## Rules

- [Rule 1 — clean up the user's wording to be actionable, but preserve their intent]
- [Rule 2]
- [Rule 3]
- ...

Rules for the generated file:

• Include the <!-- mode: simple --> marker on the line after the heading. This tells /guidelines-review which mode to use.
• Put all rules under a single ## Rules heading as a flat bullet list.
• Clean up grammar and make each bullet actionable, but do not change the meaning.
• Do NOT add rules the user did not mention.
• Do NOT add severity tags. A rule is a rule — if it's in the file, it matters.
• Do NOT add category headings (Code Style, Architecture, etc.).

After writing, tell the user the file is ready and offer: "Want to add, remove, or reword anything?" — but skip the summary table. The file is short enough to read.

Recommend committing PROJECT_GUIDELINES.md to the repo.

---

Full Mode Workflow

Used when --full is provided. Comprehensive auto-detection, structured interview, categorized output with severity tags.

Phase 1 — Auto-detect

Scan the project to infer as much as possible before asking the user anything. Read files, don't just check existence.

Signal files to scan:

Signal file	What to infer
package.json	Language (TS/JS), framework, test runner, linter, formatter
tsconfig.json / jsconfig.json	Strict mode, module system, path aliases
.eslintrc* / eslint.config.*	Lint rules already codified
.prettierrc*	Formatting conventions
pyproject.toml / setup.cfg / requirements.txt	Python tooling, linter config
ruff.toml / .flake8	Python linting conventions
go.mod	Go module conventions, Go version
Cargo.toml	Rust edition, features, dependencies
Dockerfile / docker-compose.yml	Containerization conventions
.github/workflows/*.yml	CI/CD patterns, checks, test commands
.editorconfig	Indentation, line endings, charset
CLAUDE.md / .github/copilot-instructions.md	Prior AI-assistant conventions
.gitignore	What the project considers generated/private

Pattern detection (use Glob and Grep, not shell commands):

• Use Glob("**/*.test.*") and Glob("**/*.spec.*") to detect test file naming
• Use Glob("**/*_test.go") or Glob("**/test_*.py") for language-specific test patterns
• Use Grep on a sample of source files (first 10 found) to detect:
  • Import style (named vs default exports, absolute vs relative paths)
  • Indentation (tabs vs spaces, width)
  • Quote style (single vs double)
  • Semicolon usage
• Use Glob("*/") and Glob("src/*/") to detect directory structure organization
• Check for monorepo signals: Glob("packages/*/package.json"), Glob("apps/*/"), presence of lerna.json, pnpm-workspace.yaml, or turbo.json

Phase 2 — Interview

Present auto-detected findings as a summary, then ask targeted questions to fill gaps. Organize questions by category. Skip categories already fully answered by auto-detection.

Ask these in order:

1. Code Style — Naming conventions, max line length, import ordering, quote style
2. Architecture — Preferred patterns, directory structure rules, module boundaries
3. Error Handling — Strategy (exceptions, Result types), logging conventions
4. Testing — Coverage areas, naming conventions, organization, mocking strategy
5. Git Workflow — Branch naming, commit message format, PR size expectations
6. Documentation — Docstrings/JSDoc policy, README standards, inline comments
7. Security — Secret management, input validation, dependency update policy
8. Performance — Caching strategy, N+1 awareness, bundle size, lazy loading

Don't ask about categories that clearly don't apply (e.g., skip "bundle size" for a CLI tool, skip "N+1 queries" for a project without a database).

If references/guideline-categories.md is available in the plugin directory, read the sections relevant to the detected tech stack for example rules to reference during generation. Do not copy examples verbatim — adapt them to the project's own code.

Phase 3 — Generate

Create PROJECT_GUIDELINES.md in the repo root.

Critical: every section or subsection must have a severity comment on the line immediately after the heading. Format: <!-- severity: error|warning|info -->

Default severity mapping:

Category	Default Severity
Security	error
Error Handling	error
Architecture	error
Code Style	warning
Testing	warning
Git Workflow	warning
Documentation	warning
Performance	info

Template:

# Project Guidelines

<!-- mode: full -->
<!-- Generated by guidelines-init | Last updated: YYYY-MM-DD -->

## Overview
<!-- One paragraph describing the project and its primary tech stack -->

## Code Style

### Naming
<!-- severity: warning -->
- ...

### Formatting
<!-- severity: warning -->
- ...

### Imports
<!-- severity: warning -->
- ...

## Architecture

### Directory Structure
<!-- severity: error -->
- ...

### Module Boundaries
<!-- severity: error -->
- ...

## Error Handling
<!-- severity: error -->
- ...

## Testing

### Requirements
<!-- severity: warning -->
- ...

### Conventions
<!-- severity: warning -->
- ...

## Git Workflow
<!-- severity: warning -->
- ...

## Documentation
<!-- severity: warning -->
- ...

## Security
<!-- severity: error -->
- ...

## Performance
<!-- severity: info -->
- ...

Formatting rules for the generated file:

• The <!-- severity: ... --> comment is the contract that /guidelines-review depends on. It must appear on the line directly after the section heading, before any content.
• Use concrete examples with code blocks. Don't say "use descriptive names" — say use descriptive names like fetchUserById, not getData.
• Each bullet should be independently actionable.
• Use the project's own code as examples wherever possible.
• Omit sections that don't apply to the project entirely, don't leave them blank.

Phase 4 — Confirm

Show the user a summary table:

| Category       | Rules | Severity |
|----------------|-------|----------|
| Code Style     | 8     | warning  |
| Architecture   | 4     | error    |
| Error Handling | 3     | error    |
| Testing        | 5     | warning  |
| ...            | ...   | ...      |

Ask if they want to:

• Adjust severity levels for any category
• Add, remove, or modify any specific rules
• Change the location of the file

Once confirmed, recommend committing PROJECT_GUIDELINES.md to the repo.

---

Upgrade Workflow

Used when --upgrade is provided. Converts a simple-mode file to full mode.

1. Verify PROJECT_GUIDELINES.md exists. If not, tell the user to run /guidelines-init first and stop.
2. Check for <!-- mode: simple -->. If the file is already full mode, tell the user it's already in full mode and stop.
3. Read the existing simple rules into memory.
4. Run full-mode Phase 1 (auto-detect) to gather project context.
5. Run full-mode Phase 2 (interview), but pre-populate with the existing simple rules. Present a proposed mapping of existing rules to categories:

   "Here's how I'd categorize your existing rules:

   • 'Don't add code to handler.ts' → Architecture > Module Boundaries
   • 'Everything needs test cases' → Testing > Requirements Agree with this mapping, or want to move anything?"

6. Run full-mode Phase 3 (generate) with the <!-- mode: full --> marker. Incorporate the existing simple rules into the appropriate category sections.
7. Run full-mode Phase 4 (confirm).

---

Edge Cases

Monorepo: If you detect a monorepo (multiple package.json files, packages/ or apps/ directory, or lerna.json / pnpm-workspace.yaml / turbo.json), ask whether guidelines should be global or per-package. In both modes, generate a root-level file. In simple mode, keep it as a flat list. In full mode, note where per-package overrides would go.

Existing guidelines with mode mismatch:

• Running with no flag or --simple when a full-mode file exists: warn the user that replacing it will lose severity tags and category structure. Suggest using the existing full-mode file or running /guidelines-review against it instead.
• Running with --full when a simple-mode file exists: suggest --upgrade instead of replacing, since upgrade preserves existing rules.

Existing guidelines, same mode: Ask whether to update (merge new detections) or replace entirely. Default to merging.

Empty project: In simple mode, generate a starter with 3-5 universal rules marked as <!-- TODO: add your project-specific rules -->. In full mode, generate a starter template with sensible defaults and <!-- TODO: customize --> markers.

Rules

• Do NOT invent conventions that aren't supported by evidence from the scan or user input. If something is unknown and the user declines to specify, mark it as "Not specified" rather than guessing.
• Keep the output actionable and scannable — bullet points over paragraphs.
• In full mode, the severity comment format must be exactly <!-- severity: error -->, <!-- severity: warning -->, or <!-- severity: info -->.
• In simple mode, do NOT add severity tags or category structure.
• The mode marker (<!-- mode: simple --> or <!-- mode: full -->) must always be present in generated files.