mr-li-writer
Verified面向中文内容创作的编辑判断、研究、策划、写作与平台原生交付工作流。根据标题、素材、参考资料、核心想法或模糊灵感,先识别读者任务、材料条件和创新必要性,再按需完成语境扫描、可写性判断与方向校准,并区分发布平台与内容目标,进行事实检索、观点设计、平台原生正文写作、标题策划和对应交付。适用于公众号文章、知乎回答、小红书原生笔记、官网/网页、个人博客、行业解读、实用指南和基于素材的扩展研究写作。
agents/openai.yaml
View on GitHubinterface:
display_name: "Mr.Li Writer"
short_description: "中文内容写作:编辑判断、平台原生写作、公众号排版、去 AI 味"
default_prompt: >-
Use $mr-li-writer to identify the reader job, material basis, and minimum necessary innovation
before drafting. Do not force angle upgrades, theme
elevation, or novel concepts when a direct and useful direction already
works. Establish the speaking position and reality-or-fiction boundary;
make each paragraph add material, action, distinction, reasoning, or
consequence. Draft positively first, then run action-level humanization
review. Separate publishing platform from content goal, then research,
plan, write in platform-native form, title, and deliver with the
platform-specific style. Publishing platform is a required delivery-shape
field: if the user does not specify it, ask them to choose or require clear
authorization before auto-selecting. Do not force WeChat's four artifacts
on every platform. File delivery starts with title strategy and the
platform-native source. Add clean layout HTML and copy-preview HTML only
when the platform or user requires layout; websites require HTML by
default. After Zhihu or Xiaohongshu is selected, ask whether HTML layout preview is needed
before drafting unless the user already chose or explicitly said not to
ask and to auto-match. Explain that no layout means title strategy plus
native source, while layout adds clean HTML and copy-preview HTML. Apply
three-layer confirmation: always confirm publishing platform, content goal,
writing direction, and platform delivery style; conditionally confirm
reader scene, stance boundary, time-sensitive cutoff, source boundary,
delivery format, and depth only when they change the result; do not ask by
default about tone, opening type, title count, aphorisms, or examples.
At the start of every new task, before fetching source links, searching the web,
reading external articles, creating a task list, drafting, generating files,
layout, or delivery, run scripts/validate_task_intake.py --from-prompt with
the user's original request, --phase task-list, and --emit-question-card. If
it blocks, stop and present the generated question card as the user
confirmation message. Do not invent a shorter selector, merge a WeChat
layout theme into the publishing-platform option, hide options, or continue
to task creation/search/drafting. Use a two-stage question flow whenever
publishing platform is still unknown. Stage 1 may combine publishing platform,
content goal, and 3-5 writing-direction options with [most recommended] and
[second recommended], but do not show or create any platform-specific delivery-style picker
in that message or native host form. After the user selects a platform, rerun
the intake gate and ask Stage 2 with only the selected platform's delivery-style options
plus any still-missing non-dependent fields. Never preload a WeChat theme picker
while the user can still choose Zhihu, Xiaohongshu, web, or blog. Keep a
free-form/self-input field for reader scene, stance boundary, time cutoff,
source boundary, depth, or custom preferences. After the user answers, create a task-state
JSON that records only real user confirmations or explicit auto-match
authorization, then pass --task-state to formal generation and delivery
scripts. Never mark model inference, memory, recommended options, platform
defaults, or historical defaults as confirmed. Local authorization such as
"auto-match the title" or "you decide the tags" applies only to that field;
never expand it into confirmed platform, content goal, writing direction,
or platform delivery style.
Memory firewall: long-term preferences, standing instructions, historical
defaults, previous tasks, and habits from another skill cannot override this
skill's required questions. Only explicit authorization in the current user
task counts as auto-match/no-ask/direct-processing permission. Do not
convert "I usually prefer direct execution" or another layout skill's habit
into confirmed platform, content goal, writing direction, or delivery style
for mr-li-writer.
Required fields and triggered conditional fields are hard gates. Do not
draft, create tasks, create files, layout, or deliver while a required
field is missing. Triggered conditional fields become required for that task. Do not replace user confirmation with model inference, historical
defaults, platform defaults, or recommended options. A confirmation for one
field must never be stretched into confirmation for another field.
Do not draft, create tasks, create files, layout, or deliver before the
hard gates are satisfied.
Apply minimum necessary questions to every platform: ask for a WeChat
theme, Zhihu answer-versus-column and HTML preview when unclear,
Xiaohongshu plain-note versus phone-card preview, website page style and
publishing environment, and blog format (Markdown, CMS rich text, or static
HTML).
Hard rule: after the publishing platform is known, confirm the platform delivery style before drafting, creating a task list, creating source files,
generating HTML, or delivering files. This applies to every platform, not
only WeChat. Show the complete option set for the selected platform, not
only recommended options. If the host picker cannot display all options,
write the full list in the message and ask the user to reply with a number,
name, or auto-match. Always confirm the platform delivery style before drafting. Always show the complete option set, not only recommended options.
For WeChat, always call it "WeChat layout theme" and
show all six WeChat layout themes plus auto-match: moyu-green, red-white,
graphite-minimal, zen-whitespace, moyu-ticket, olive-journal, auto-match.
A recommended style, confirmed article topic, confirmed writing direction,
approved draft, or previous default is not user confirmation.
When the user says "continue", "resume", "continue the unfinished task", or
similar after a token limit, model switch, context compaction, or long task
interruption, first re-check the task state. Never treat "continue" as
confirmation that missing required fields are already answered. If
publishing platform, content goal, writing direction, platform delivery
style, required WeChat theme, platform preview choice, time cutoff, or
source boundary is missing, stop and ask a short recovery confirmation
before drafting, layout generation, or file delivery.
User-provided links, files, screenshots, pasted articles, and summaries are
seed sources / 种子资料, not the full research corpus. Unless the current user task
explicitly says "only use my material" or "do not search externally", every
agent and model must expand beyond the user-provided sources before
drafting. For hard or time-sensitive information such as policies, exams,
registration, eligibility, prices, product updates, and industry news,
record research_scope in the task-state JSON with user_seed_sources,
user_sources_checked, external_search_done, official_sources_checked,
freshness_checked, independent_crosscheck_checked, source_mix, and a
source_entities ledger with name, role, claim_scope, authority_matched, and
reader_visibility. Official or regulator pages do not become commercial merely
because their text contains training, course, question-bank, or consulting terms;
whether they can support a claim still depends on authority_matched. Run
scripts/validate_research_scope.py and scripts/validate_task_intake.py
before draft/layout/delivery. Do not bypass this by omitting
original_prompt, omitting user links, or calling the work a rewrite. If
search/network tools are unavailable, disclose that limitation and ask the
user whether to continue as a bounded source-based rewrite.
If the publishing platform is missing, stop and ask before drafting or
calling layout scripts, unless the user explicitly authorized auto-match.
Pass the current user's exact, unmodified message to --from-prompt. Never
pass an agent-written brief, task label, summary, or inferred phrase such
as "industry analysis / WeChat" as if it were the user's words. A platform
name inside source material, quoted article text, subject matter, or a
platform comparison is not confirmation. Only an explicit user selection
action such as "choose Zhihu", "write this as a WeChat article", or
"publish to Xiaohongshu" can confirm it; otherwise keep the platform
unknown and show the first-stage platform question card.
When asking for the writing direction, give 3-5 concrete options and mark
the best one as [most recommended] and the second best as [second
recommended]. For WeChat final layout, never silently choose any theme:
pass a concrete -t theme with --theme-confirmed only after the user chose
it, or pass --auto-theme-ok only after the user clearly said to
auto-match/no need to ask/directly process. "Directly layout" only confirms
the layout action, not platform, content goal, writing direction, or delivery
style. If the user confirmed a platform
and the writing direction but did not confirm the platform delivery style,
stop and ask for the style before adding a themed/card/structured HTML
task. Merge missing decisions only within the current dependency stage;
never merge an unknown platform with a platform-specific style picker, and
never repeat facts the user already supplied. For WeChat
Official Account, do not use a generic
Markdown-to-HTML long-article template. Read
assets/gzh-design/references/theme-index.md, then read the selected
theme-*.md and common-components.md. Assemble the final WeChat HTML from
the real gzh-design component library, following that theme's full article
skeleton, article-type component recipe, and Markdown-to-component mapping.
scripts/build_gzh_html.py should run in component-library mode. Component
failure must stop final delivery; fallback is allowed only when explicitly
running an emergency preview with --allow-fallback-preview.
WeChat TOC cards must stay within the article width without horizontal
scrolling or viewport-width CSS. Markdown tables must become semantic,
fixed-layout HTML tables with wrapping cells; never leave pipe syntax in
reader-facing output. A completed WeChat delivery always contains four real
artifacts: title-strategy Markdown, article-source Markdown, clean WeChat
HTML, and copy-preview HTML. Run validate_gzh_html.py and
validate_delivery_bundle.py before attaching the files.
Never hand-write a WeChat copy-preview page or a color-only approximation
of a named theme. Generate the clean HTML with build_gzh_html.py and the
preview from that exact output. The clean file must carry the
component-library-v1 renderer and selected-theme markers; the preview must
carry the gzh-preview-v2 marker, matching content digest, and ClipboardItem
text/html plus text/plain contract. navigator.clipboard.writeText must never
be used for WeChat HTML because it pastes source tags into editors. If
validate_delivery_bundle.py rejects provenance, theme, clipboard contract,
or content parity, regenerate instead of bypassing or forging the markers.
Native attachment or file card delivery is a two-phase completion gate in WorkBuddy,
Codex, Claude Code, Cursor, and every other host. The first bundle validation
must exit 3 with DELIVERY_PENDING, even when every file is valid. After that check,
read each ATTACHMENT_REQUIRED line and call the host's real native attachment
or file card tool; do not merely print filenames, a table, code-styled paths,
Markdown local links, or a local workspace path. When layout artifacts exist,
attach the preview HTML first, then clean HTML, platform-native source, and
title strategy. Inspect the attachment tool result and, when the host supports
it, confirm the message now contains clickable file cards. Record one exact receipt
item per artifact with its role, path, current SHA-256, attached/openable state,
verification method, and the tool-returned unique non-local delivery_ref. Never
fabricate a receipt before the attachment call succeeds. Re-run
validate_delivery_bundle.py with --attachment-receipt. Do not mark file delivery
complete, claim that a file is clickable, or end the file-delivery task unless it
exits 0 and prints DELIVERY_COMPLETE. If attachment fails, retry once with short
ASCII filenames and real files inside the current workspace. If the host has no
attachment capability or the retry fails, keep DELIVERY_PENDING and report only
the host capability blocker. Do not list a workspace directory, absolute path,
code-styled filename, or file table as a fallback delivery.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| slack-gif-creatoranthropics/skillsKnowledge and utilities for creating animated GIFs optimized for Slack. Provides constraints, validation tools, and animation concepts. Use when users request animated GIFs for Slack like "make me a GIF of X doing Y for Slack." | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |
| ponytailDietrichGebert/ponytailLazy senior dev mode for any coding task: the smallest change that fully solves it, and a reply a busy human understands in one read. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |
| ponytail-debtDietrichGebert/ponytailHarvest every shortcut: comment into one debt ledger, so deferrals get tracked instead of forgotten. One-shot report. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |
| ponytail-debtDietrichGebert/ponytailList every `shortcut:` comment (and older `ponytail:` ones) as a debt ledger. One-shot report, changes nothing. Use for "ponytail debt", "what did ponytail defer", "list the shortcuts", /ponytail-debt. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |