Claude, Do It Like Last Time: Steady Design, Front-End, and Skills
1. A structure Claude can verify on its own
Consistent results come not from the model’s tendencies, but from clear completion criteria and verification loops.
- Completion criteria
- Provide acceptance criteria that Claude can run directly, such as tests, builds, and screenshot comparisons.
- Rather than vague requests for improvement, asking it to find differences from the original and fix them produces more stable repetition.
- Verification strength
- For simple tasks, put the criteria in the prompt; for important tasks, fix the end condition with `/goal` or a Stop hook.
- You can also have a separate reviewer subagent review only the results in a new context.
2. Placing rules in the right memory layer
Drift often happens not because there are no rules, but because they are in the wrong place or buried in too much context.
- Separation by role
- Keep only the core rules needed for all tasks in `CLAUDE.md`, briefly.
- Separate domain knowledge and long procedures that are needed only sometimes into skills.
- Fix lint and verification that must run in Hooks rather than in descriptions.
- Context management
- Use `/clear` to clean up previous context between unrelated tasks.
- Hand research and review to separate subagents so the implementation context does not become unnecessarily long.
3. Connecting Claude Design and Figma MCP through implementation
Reusable code comes out only when the design system is used as the basis and the Figma file is organized.
- Figma preparation
- Turn repeated elements into components and connect them to real code components through Code Connect.
- Manage colors, spacing, and typography with variables and tokens.
- Use meaningful layer names, Auto Layout, and comments to convey not only what is visible on screen but also the underlying structure and behavior.
- Front-end loop
- Divide the work so that Claude Code explores the code structure, writes a plan, and then implements it.
- Repeat the process of taking screenshots in the browser and comparing them with the original design to fix differences.
4. Skills that work the same way every time
Skills do not replace new judgment; they keep the agent from drifting away from the already defined task criteria.
- Triggers and structure
- Write in `description` both what the skill does and when to use it so it is called precisely.
- Keep the body short and split long reference materials and execution scripts into separate files.
- For stages that can easily break, apply exact values, commands, and low freedom.
- Verification design
- Include a checklist and a verification → fix → reverify loop inside the skill.
- For repetitive work, provide a verified script rather than generating code every time.
- In a new session, compare before and after applying the skill to confirm whether the results are actually more stable.
