/copy-research

name: copy-research
description: "Dispatch copy-lead helpers: Tavily, Codex luna/terra, grok/X, OpenCode DeepSeek, Cursor Grok 4.6 medium-fast. Use when: глубокий ресёрч, firecrawl, tavily, сподручные, конкуренты, twitter, язык ЦА. SKIP: SEO SERP (seo-specialist); one URL (WebFetch)."

Copy research

Write notes to `.agents/copy/research/inbox/YYYY-MM-DD-<slug>.md`. Then **you** fill audience/page files and `mv` the note to `used/`. Research ≠ buyer quote. Never one `web.md`.

First load `references/helpers.md` (who runs the CLI). Then `references/search-playbook.md` for query rules. Do **not** `npx skills add` unless the human has the key **and** asked. Twitter/OAuth skills: never.

Who runs what

| Need | Runner | |---|---| | Одна страница | You: `WebFetch` | | Короткий поиск с цитатами / URL | skill **`tavily`** → `/search` | | Отчёт: рынок / ЦА / ландшафт | skill **`tavily`** `/research` **or** Firecrawl (key) | | Открытый веб, несколько URL | **Codex `gpt-5.6-luna` + fast** (`terra` if thin) | | X / Twitter, живой сленг | **grok** CLI | | Пачка вариантов / дешёвый черновик | **OpenCode** `deepseek-v4-flash` | | 2–3 сильных угла | **OpenCode** `deepseek-v4-pro` | | Второй проход с репо | **cursor-agent** `cursor-grok-4.6-medium-fast` | | Уже в репо | `Agent(Explore)` | | Неясно, с чего начать | `Agent(general-purpose)` first, then CLI |

Recipes: `references/helpers.md`. Helper writes inbox only. You lift.

Tavily

Load skill `tavily`. Key: `$TAVILY_API_KEY` (see the skill for the Env Catalog and terminal fallback). Search = `/search`. Cited report = `/research` (ask duration).

Firecrawl deep-research (report-scale)

Official skill: https://www.skills.sh/firecrawl/firecrawl-workflows/firecrawl-deep-research Source: `firecrawl/firecrawl-workflows` · skill `firecrawl-deep-research`. Needs `FIRECRAWL_API_KEY`.

Use **only** when the human wants a cited report (landscape, how the category talks, multi-angle + sources). Not for one URL, top-N, or «найди 3 заголовка».

1. If `FIRECRAWL_API_KEY` is unset → stop. Tell the human to export it. Do not invent sources. 2. If the skill is not installed:

npx skills add https://github.com/firecrawl/firecrawl-workflows --skill firecrawl-deep-research

3. Ask one question if duration is unknown: «Сколько крутить ресёрч?» (их онбординг). 4. Load skill `firecrawl-deep-research`. Output also copy to `research/inbox/YYYY-MM-DD-<slug>.md`. 5. Lift language into `audience.md`. Do not treat report sentences as buyer quotes. Then `mv` the note to `used/`.

CLI helpers

Load `references/helpers.md` and run that recipe. Do not paste a second copy here.

After the worker

1. Read the inbox note. 2. Lift **language** into `audience.md` / persona `words_they_use`. 3. Do not paste a tweet as `happy_customer_pattern`. 4. `mv` the note to `research/used/` (noise → `dead/`). Refresh `INDEX.md`. 5. Then headlines / UX as usual.

NEVER

• Twitter/OAuth/post skills from skill.sh
• Firecrawl deep-research for a single URL or a top-N list
• Firecrawl without `FIRECRAWL_API_KEY`
• `run-init` / writer lanes
• Treat scraped text as a confirmed customer quote
• Append into one `web.md`
• Let a helper edit canon / `locked` files
• Pick a Cursor model other than `cursor-grok-4.6-medium-fast`