Context-First Design Execution

Keep ongoing design work fast and continuous. Infer from evidence already available, state the working interpretation briefly, validate the primary requested change early, and avoid reopening settled questions.

Read intent-resolution-contract.md when the request is short, referential, names an uncertain capability, contradicts an earlier result, or could route to several design Skills. Use it to resolve intent from context without silently expanding scope.

Build the intent frame

Before planning or using tools, reconstruct these fields privately:

Apply this precedence:

current explicit instruction
→ current attached artifact or named node
→ latest confirmed decision in the active task
→ latest user-accepted result
→ older plans and historical notes

Treat new negative feedback as an immediate invalidation of the affected completion claim. Do not keep using a rejected result as the baseline.

Resolve references from context

• Resolve “这个、这些、上面、前面、同一个、继续” to the most recent compatible object in the active task.
• Reuse already supplied URLs, node IDs, filenames, viewports, and scope decisions. Do not ask the user to repeat them.
• Infer the narrowest action that satisfies the current request and stays within confirmed scope.
• State one short working interpretation when the request is referential, then act.
• Ask only when an unresolved choice would materially change scope, overwrite important work, cause an external side effect, or produce meaningfully different deliverables.
• Prefer a labeled assumption for low-risk details that can be corrected locally.

Protect the user's real priority

Translate feedback into the primary acceptance delta before making changes.

Examples:

• “间距没调” means geometry and vertical rhythm are the primary delta; token binding or color cleanup cannot prove completion.
• “莫名其妙的灰线” means every visible divider needs a source or semantic justification; a style audit alone is insufficient.
• “太乱” in a change deck means object, action, and Before/After relation need clearer hierarchy; adding more explanatory text usually increases the problem.
• “空白的” in a presentation means editor and thumbnail first frames must show complete content; embedded media presence is not proof of usability.

Keep proxy improvements secondary until the primary delta passes.

Execute with a fast checkpoint

1. Read indexed memory and the relevant project note first. Search raw session evidence only for a missing detail.
2. Inventory target artifacts and identifiers once. Reuse the inventory unless the source changes.
3. Group work by shared rule, not by file count.
4. Choose one difficult representative specimen for each rule family.
5. Apply the change to the specimen, render or capture direct evidence, and compare it with the source of truth.
6. Correct the rule before batching. After the specimen passes, reuse the rule across the remaining targets.
7. Revalidate only affected surfaces plus a small regression set. Do not repeat a full baseline audit after every micro-change.

For large tasks, prefer this sequence:

intent frame → one representative specimen → direct comparison → batch reuse → final exception scan

Avoid slow or repetitive behavior

• Do not narrate every tool call or restate the full plan after each milestone.
• Do not reopen confirmed scope questions because a tool call failed.
• Use one documented recovery path for a tool failure; stop looping after the bounded retry fails.
• Cache semantic mappings, node IDs, file lists, and validation criteria in the active task.
• Batch independent reads and validations when safe.
• Prefer targeted inspection over rereading entire repositories, Figma pages, or long session histories.
• Stop collecting evidence once one authoritative signal proves the fact and no direct contradiction remains.

Completion gate

Claim completion only when:

• the primary requested delta has direct visual or structural proof;
• rejected patterns no longer appear in the affected scope;
• locked content and decisions remain unchanged;
• representative results were checked before batch propagation;
• remaining exceptions are named and kept open.

If any gate fails, report the exact open item and continue when an in-scope corrective action remains.

Return a compact handoff

Lead with the outcome. Include only:

1. the interpretation used;
2. the artifacts or surfaces changed;
3. the direct validation result;
4. any open exception or next decision.

Do not make the user reconstruct the result from process commentary.