AI Steering Files
A practical catalog of repository steering files for AI coding agents. Use it to decide which context files your repo needs, what each one should contain, and which examples are worth studying.
Browse the web version: https://ai-steering.lvtd.dev.
Use this repo when:
- agents keep missing repo-specific commands, constraints, or product context
- maintainers need one place for tool-neutral instructions
- a team wants Claude, Gemini, Copilot, or Cursor guidance without duplicating shared rules
Install the Reusable Skill
Install ai-steering-files when you want an agent to reuse this guidance in another project.
With the Skills CLI:
npx skills add LVTD-LLC/ai-steering
With GitHub CLI:
gh skill install LVTD-LLC/ai-steering ai-steering-files --agent codex --scope user
Source: skills/ai-steering-files/SKILL.md
Agent-Agnostic Files
Start here for rules and context that should help every coding assistant work safely in the repo.
The shared operating manual every coding agent should read before changing the repo.
Use it for:
- repo-specific behavior that applies across tools
- safe defaults, guardrails, and approval boundaries
- exact commands agents can run with confidence
- review, branch, and merge expectations
What it should contain:
- concise project summary
- preferred workflow
- exact install, test, lint, build, and deploy commands
- branch and PR policy
- actions that need approval
- rules agents should not bypass
Useful resources:
Great examples:
Product context that keeps implementation tradeoffs tied to users and outcomes.
Use it for:
- target users and use cases
- the problem the product solves
- major workflows and user journeys
- business objectives and success criteria
What it should contain:
- why the project exists
- who it serves
- what good looks like
- what is in scope and out of scope
- the outcomes that matter most
Useful resources:
Great examples:
A checked-in measurement contract for product events, metrics, attribution, and privacy boundaries.
Use it for:
- canonical product and marketing metrics
- event names, properties, identity, and versioning rules
- funnel stages, attribution, and source-of-truth decisions
- consent, privacy, retention, and excluded data
- analytics validation and governance
What it should contain:
- the outcomes and funnels the product measures
- canonical event and property contracts
- identity, attribution, and metric-source rules
- data that must never be collected
- consent and privacy requirements
- how to validate changes and evolve the schema safely
Useful resources:
Great examples:
The technical source of truth for stack, commands, and implementation constraints.
Use it for:
- framework and library choices
- runtime and package manager versions
- build, check, test, and deploy commands
- deployment targets, integrations, and APIs
What it should contain:
- the canonical stack
- supported commands
- preferred libraries and patterns
- external services and APIs
- the constraints agents must respect
Useful resources:
Great examples:
A repo map that helps agents put code, docs, and tests in the right place.
Use it for:
- top-level directory layout
- naming conventions
- import patterns
- module boundaries and ownership lines
What it should contain:
- a directory map
- ownership boundaries
- import rules
- placement rules for new files
- special-case folders that matter for the agent
Useful resources:
Great examples:
Long-term direction, non-goals, and taste principles that should outlast individual specs.
Use it for:
- long-term product direction
- non-goals
- design and product taste
- the future state the project is trying to create
What it should contain:
- vision and durable constraints
- what the project should become
- what should not drift
- success criteria at the level of outcomes, not implementation
Useful resources:
Great examples:
A design-system source of truth for AI-generated UI changes.
Use it for:
- design tokens and theme choices
- typography and spacing
- layout rules
- component patterns
- accessibility expectations
What it should contain:
- the visual language
- spacing and type rules
- layout constraints
- component style preferences
- UI patterns to avoid
Useful resources:
Great examples:
Agent-Specific Files
Add these when a tool needs its own syntax, memory model, or narrower workflow rules.
The Claude Code memory file for repo-specific instructions and workflows.
Use it for:
- Claude-specific guidance
- repo conventions that should persist across Claude sessions
- build/test/verify expectations for Claude
- workflow notes that are useful in Claude Code
What it should contain:
- project-specific instructions
- code style and review rules
- build and test commands
- escalation boundaries
- paths or workflows Claude should handle carefully
Useful resources:
Great examples:
The Gemini CLI context file for repo instructions, commands, and response preferences.
Use it for:
- Gemini-specific instructions
- repo conventions that should persist across Gemini sessions
- project context and response preferences
- custom commands or tool usage hints
What it should contain:
- project-specific context
- coding conventions
- build/test/verify notes
- files or folders Gemini should prioritize
Useful resources:
Great examples:
.github/copilot-instructions.md
Repository-wide instructions GitHub Copilot should apply when suggesting changes.
Use it for:
- Copilot instructions that apply across the repo
- build/test/validation guidance
- repo-specific coding standards
- review and safety expectations
What it should contain:
- how to build and test the repo
- conventions and patterns to follow
- constraints that should always be respected
- any special validation steps
Useful resources:
Great examples:
.cursor/rules/*.mdc
Path-scoped Cursor rules for workflows that need narrower guidance than a repo-wide file.
Use it for:
- file-specific or workflow-specific rules
- path-scoped instructions
- rule sets for certain parts of the repo
- agent behavior that should not apply globally
What it should contain:
- path or task scope
- the conventions to follow in that scope
- examples of good output
- what should auto-apply vs stay manual
Useful resources:
Great examples:
Keeping It In Sync
Catalog guidance lives in src/data/steering-data.js. Edit that file, run npm run sync:readme, then run npm run build before opening a PR.