rayskills is a builder toolbox forged in real business. Users do not need to remember each skill name—hand over the real situation to /ray, and it decides the single step that should be done right now and dispatches to the right member skill.
Routing philosophy (single step by default, continuous execution when the endpoint is clear): when the requirement has only a situation or the next step is unstable, decide only the current step. When the user has already made the endpoint clear and a stable handoff protocol exists between intermediate stages, you may continuously execute a verified pipeline without requiring the user to confirm each segment. The fixed endpoint of the content pipeline is a verified draft; public publishing is not part of this pipeline and cannot be inferred from "push", "publishing process", or past permission.
How to dispatch
- Read the information already present in the current conversation (situation, goal, materials, blockers).
- Compare against the routing table below to determine which line and which skill the intent falls into.
- By default, choose only one skill, explain why it is that one, then execute according to that skill's SKILL.md. If the user explicitly requests end-to-end content production, read content-pipeline.md, and under stage gating continuously call
ray-writer、ray-cover, then according to the target platform callray-wechatand/orray-x-article。 - When the information is insufficient to decide, ask only one most critical question; normal stage handoffs, verification, and recoverable retries should not repeatedly interrupt the user.
Workbench task queue
The knowledge workbench (ray-dashboard, installed with rays-brain-kit , not in this repository) writes the user's queued AI collaboration requests into the vault's 50-系统/40-自动化/AI任务队列.md（frontmatter kind: ai-task-queue, under "## Pending", each line - [ ] 时间 · 动作 · [[目标笔记]]）。/ray When reading context, take a casual look at this file; if the file does not exist or has no unchecked items, silently skip it and do not mention it.
- There are pending items, and the user has not given any other task this round → lay out the queue, and after confirmation claim it: "draft" follows the finished-package path for the target writing task
ray-writer。 - The user has a clear task this round → do this round's work first, and at the end mention that there are still a few pending items in the queue.
- After each item is processed, check that line as
- [x]and append the artifact link at the end of the line, so both the workbench and the next session can see where it went. - The completion state of queue actions is always local artifacts; claiming a queue item does not constitute any platform write authorization, and platform actions still follow the tiered determination below.
Platform actions are tiered first
Before calling a platform Skill, you must first, based on the user'sexact words in the current round, determine the action level:
| Level | What can be done | Basis for determination |
|---|---|---|
local_only |
Only make local artifacts: layout, preview, cover, delivery package, without touching any online backend | The user only says "let me take a look first", "give me a preview", "prepare the delivery package", or the whole sentence does not name a platform destination |
draft_write |
Create or update a platform draft, and fully read it back for acceptance | The exact words of this round simultaneously containa specific platformanda destination, for example "save to the WeChat draft box", "update the X Article draft", "put it into the X backend" |
| Public release | Not part of the content production pipeline | This decision never produces it; a separate explicit publishing task is required, confirmed platform by platform |
There are three decision rules,ray-wechat and ray-x-article use the same set, with no individual interpretation:
- Names only a platform, verb ambiguous—"push to X", "push to WeChat", "sync to two platforms", "publish to the backend"—do not directly escalate to
draft_write. Finish the local artifact first, then ask clearly in one sentence: "It's ready locally, should I write it into the <platform> drafts now?" Write only after getting an affirmative answer. Neither write without authorization, nor finish and say nothing. - No platform destination named at all—"draft passed OK", "move to the next stage", "run the full publishing process"—always
local_only, neither ask nor write. - The user explicitly says "publish", "mass send", "public"—still only go as far as drafts, and explain that public release is beyond this pipeline.
Permission from past tasks, permission from another platform, an existing draft ID, an already logged-in state, or the allowlist having been fixed, none of these constitute authorization for this round. When uncertain, always downgrade to local_only。
Routing table
Content / IP
| What the user wants to do | Route to |
|---|---|
| Create, check, or adapt a local Obsidian content knowledge base | ray-obsidian |
| Turn an idea, material, or draft into a Chinese long-form article with Ray's attitude and reach | ray-writer |
| Translate a foreign-language article or repost someone else's Chinese article | ray-writer(Translation mode, first pass the two gates of authorization and attribution) |
| Do a quick spoken-word piece directly from inspiration and search; redistribute a spoken-word piece from a finalized long article; review or compare against the actual recorded script | ray-kb |
| Generate WeChat Official Account, X, and X Article covers for a finalized article, make 9:16 / 3:4 vertical covers for videos, change style, change cover title | ray-cover |
| Turn an article, clipping, or viewpoint into a no-face-on-camera knowledge explainer video (code-drawn visuals, cloned voice dubbing) | ray-explainer(First produce a storyboard for Ray to confirm, then dub and draw) |
| Format a finalized article and save it to the WeChat Official Account drafts | ray-wechat |
| Save a finalized article and a 5:2 cover to X Articles drafts | ray-x-article |
| Go from idea/draft all the way to article, cover, and X Article draft | Content production pipeline |
Independent repository members (ray-skills, routed only if installed)
Consulting, multi-model collaboration, trend research, and collage B-roll have been moved by usage frequency to ray-skills. When the corresponding skill is installed locally, route as usual; if not installed, state it directly and handle as an ordinary task, without pretending to invoke it.
| What the user wants to do | Route to |
|---|---|
| Assess whether an enterprise can adopt a knowledge base/AI, or produce a solution blueprint, phasing, and quote based on the diagnosis | ray-consult |
| Have Grok / Claude / Codex divide work, independently review, compete, or check X / Reddit / web in real time | ray-multimodel |
| Check cross-platform discussion, popular videos, and community sentiment on a topic over the last N days, clearly marking dates and evidence | ray-trend-search |
| Make collage B-roll / collage explainer videos for a spoken-word script or topic | ray-broll |
Common transitions (for judging the next step)
These are not a fixed pipeline, but empirical references for "once the previous result is out, what usually comes next":
- The user has no compatible local knowledge base but wants long-term content accumulation → first connect to
ray-obsidianSafe initialization, then hand the root directory and materials toray-writer。 ray-writeris complete, Ray has processed the "to be confirmed" items and passed the mobile browseability check → connect toray-coverto generate platform covers; when it needs to enter the WeChat Official Account draft box, connect toray-wechat, when it needs to enter the X backend, connect toray-x-article, both save drafts only by default. When the user explicitly requests "run the entire pipeline", follow content-pipeline.md to execute consecutively.- The judgment from inspiration, on-site observation, or conversation can already be explained clearly in one video →
ray-kbusefast-oral: retrieve materials as needed, write directly in Ray's tone, establish only one oral script master draft, and do not force supplementary candidate cards, writing tasks, or long-form text. - Long-form text needs one round of material collection and multiple distributions →
ray-kbgenerate an oral content package bound to the master draft fingerprint; when making an explainer video, then connect toray-explainer. Image long-form text enters the subsequent visual stage only when the user explicitly requests it. - Before topic selection or writing, if you need to know how a certain topic has recently been discussed → when
ray-trend-searchis installed, research first, and hand conclusions with dates and evidence toray-writerorray-kb。
Boundaries and Discipline
- Only perform routing and stage orchestration, do not replace the judgment of member skills. Single-step tasks are handed to one member; end-to-end content tasks connect members according to the handoff protocol, and member rules are not temporarily rewritten in the main router.
- Content-type expression comes first：
ray-writerandray-kbfully express opinions and emotions in Ray's own tone; uncertain or fabricated experiences and data are listed as "to be confirmed", for Ray to decide item by item before entering the platform. When the entire content belongs to someone else, use translation/introduction mode, solved through authorization and attribution, not through rewriting to evade;source_permissiontranslations that have not been approved are only local artifacts. - Drafts and public publishing are strictly separated：
ray-wechatandray-x-articlecompletion states are both verified drafts; any fix to whitelists, egress IPs, login, or file-sharing permissions must not expand the original authorization. - When a request only mentions multiple lines in passing, choosethe one step that should be done most right now, and treat the rest as next steps; execute consecutively only when the user explicitly gives an end-to-end endpoint and a formal handoff protocol exists.