Context & Token Efficiency

Shorter instructions leave room for the task. These examples do not claim a token-savings percentage. Load the file that matches the task, and leave the rest out.

Project map

AGENTS.mdProject identity, the Next.js version, and the paths of scoped rules and skills.

.cursor/rules/UI, security, and performance notes. Request the file that matches the task.

.agents/skills/Longer procedures. Read the skill whose description matches the task, and leave the others out.

Standards & Best PracticesComponent restrictions, accessibility, responsive behavior, and code snippets.

UI implementation rules are in Standards & Best Practices. Agent workflows are in Skills & Workflows.

Focused context

Keep the root file shortBefore

AGENTS.md contains every button rule, every token, and every component API.

After

AGENTS.md names the project, the Next.js version, and the paths of scoped rules and skills.

Codex stops at its configured instruction size. A short root file leaves room for a nested file when a task is inside one folder.

Load a skill for the taskBefore

One always-on rule mixes UI tokens, authorization, and rendering advice into every prompt.

After

Three skills each have a name and description. The longer note stays in references/ and is read when that subject is in the task.

Cursor, Codex, and OpenCode can show the description first and load the body later.

Point at the installed versionBefore

The prompt tells the agent to read node_modules/next/dist/docs/ and call /_next/mcp.

After

The prompt says this install is Next.js 15.5.27 and those paths are not present.

An agent that opens a missing path spends the turn on a file that this version does not ship.

Leave MCP out until it is approvedBefore

A generated .cursor/mcp.json enables GitHub, filesystem, and a browser with a copied token.

After

The generator writes a review note under examples/ and does not create .cursor/mcp.json.

An enabled server is a permission. A note can be read without granting that permission.

Reusable prompts

Each prompt is generated from the published rules for one component or one UI guideline. Proposed CSC standards are not approved policy, and the prompt does not claim a WCAG audit.

Loading prompts…