一键导入
context-hygiene
Use when a long or complex task needs a safe session boundary, durable resume state, or separation into focused follow-up sessions.
用 Codex 或 Claude 帮你安装 复制这段 Prompt,粘贴到 Codex、Claude 或其他助手里,让它检查 Skill 页面并帮你完成安装。
菜单
Use when a long or complex task needs a safe session boundary, durable resume state, or separation into focused follow-up sessions.
用 Codex 或 Claude 帮你安装 复制这段 Prompt,粘贴到 Codex、Claude 或其他助手里,让它检查 Skill 页面并帮你完成安装。
Use this skill when the user wants to test, inspect, verify, or screenshot a page in the browser (e.g., "/browser-test", "/browser", "test in browser", "take screenshots", "verify in Chrome", "check devtools"). On Windows environments, this skill bypasses browser_subagent and uses Playwright MCP tools to navigate, evaluate JS/scroll, and capture screenshots reliably.
Use this skill when finding, scoring, or profiling prospects for the Cinematic Digital Showroom offer targeting luxury residential interior designers and boutique interior design studios. Activated by "find prospects", "research interior designers", "find Lagos prospects", "find Dubai prospects", "who should I target this week", "score this designer", "prospect research", "find designers with weak websites", or any task that involves identifying, qualifying, or prioritizing outreach targets for the BEVAMPED studio offer. Do NOT use for general web research, competitor profiling, or sales outreach copy (use sales-enablement or copywriting skills for those).
Run an after-action learning and capability-impact audit after a substantial completed task, milestone, investigation, launch, or creative project. Activate on `/learn`, "what did we learn", "audit what just happened", "which skill should improve", "update our workflow from this", or requests to convert recent work into reusable system improvements. Do not use for ordinary status summaries, unfinished work, generic code review, or automatic skill/memory mutation without explicit approval.
Use this skill when the user asks for senior creative direction, luxury web design judgment, spatial-editorial critique, or taste calibration for high-end interior, architecture-adjacent, spatial storytelling, portfolio, gallery-like, residential design, premium studio, or editorial brand websites.
Analyze and translate design references, award-site videos, screen recordings, screenshots, transcripts, AI observation reports, precedent collections, and inspiration links into brand-specific decisions. Use for requests such as "analyze these references", "what should we borrow", "compare these award websites", "turn these video reports into a plan", or "what should we keep, adapt, reject, or defer". Do not use for competitor pricing/market dossiers, approved motion-library lookup, brand diagnosis without references, or implementation from an already approved plan.
Use this skill when designing, auditing, or planning high-end interior decorator, interior design, spatial design, architecture, styling, furniture, decor, showroom, gallery, or luxury home websites. Activated by requests like "build an interior decorator portfolio", "make this spatial design site feel premium", "create a high-end interiors website", "design a decorator portfolio", "make this feel like a 20k website", "use the design audit grammar", or any visual task where rooms, materials, light, objects, before and after transformations, or curated spaces must drive the experience.
基于 SOC 职业分类
| name | context-hygiene |
| description | Use when a long or complex task needs a safe session boundary, durable resume state, or separation into focused follow-up sessions. |
Long conversations can accumulate stale assumptions, competing objectives, and irrelevant history. Manage session boundaries while preserving only the state needed to resume safely.
Recommend a fresh session when:
Do not use an arbitrary message count as the decision rule.
Prefer a host-provided handoff mechanism when available. Persist workspace files only when the user has authorized local mutation.
If local memory mutation is authorized, capture only sanitized lessons that are useful beyond the current task:
.agents/memory/decisions-log.md..agents/memory/mistakes-to-avoid.md..agents/memory/common-patterns.md.Do not persist secrets, raw untrusted text, transient progress, or project-private material in global memory.
Update only the current task's .agents/workflows/<task-id>.json record and the workflow index. Validate the complete record against workflow-state.schema.json. Keep current_state, completed_states, evidence, artifacts, blockers, and next_action current.
Illustrative subset:
{
"schema_version": 1,
"task_id": "opaque-task-id",
"workflow_id": "build-feature",
"mode": "implement",
"status": "in_progress",
"current_state": "implement",
"completed_states": ["intake", "design"],
"evidence": [],
"artifacts": ["src/example.ts"],
"blockers": [],
"next_action": "Wire the approved submit flow to the API."
}
The complete record also includes owner, workspace, lease, approvals, timestamps, and archive state as required by the schema.
State which task record was updated, summarize the next action, and recommend a fresh session. Never promise automatic resume unless the active host actually supports it.