Open-source platform for 3D AI agents. Turn text or a photo into a rigged, animated GLB avatar, give it an LLM brain, memory and a wallet, and embed it anywhere with one web component. Remote MCP server, x402 payments, Three.js, browser-native.
The daily partner spotlight
Every day @trythreews features one partner listed on three.ws/partners and one thing three.ws built on that partner's technology. Each spotlight is a thread of a head post and three to five replies, carrying a reel filmed live against production that day, plus a head built for the slot: a card, a GIF, the reel itself, or generated art. It runs inside the X content pipeline, so every rule there (the claims ledger, the language lint, the editor review, the veto window) applies to it too.
| Piece | Where |
|---|---|
| Registry: one file per partner, listing its designation, wording rules and features | data/x-content/partners/ |
| Engine: rotation, brief, drafting, series lint, claims, item | api/_lib/x-content/partners.js |
| CLI | scripts/x-partners.mjs (npm run x:partners) |
| Slot and approval | slots and approval.series in data/x-content/queue.json |
| Scheduled runner: a Cloud Run Job image and its setup | deploy/x-partners/, scripts/create-x-partners-job.mjs |
| Tests | tests/x-content-partners.test.js |
The registry
A partner file is the only source the series may draw on. The model never invents a fact: every sentence it can lean on is written here with evidence the claims ledger checks at review.
{
"id": "ibm",
"name": "IBM",
"order": 2,
"designation": "IBM Business Partner",
"designationEvidence": [{ "type": "page", "url": "https://three.ws/partners", "contains": "IBM Business Partner" }],
"page": "https://three.ws/blog/three-ws-ibm-business-partner",
"handle": "IBM",
"mark": null,
"forbidden": ["endorsed by IBM", "official IBM partner"],
"guidance": "How to talk about this partnership, and what never to say.",
"docs": ["docs/partners.md"],
"features": [
{
"id": "user-group-in-world-meetup",
"title": "A meetup inside a 3D world",
"surface": "https://three.ws/...",
"covers": [],
"lane": "community",
"summary": "One paragraph on what was built and why it matters.",
"docs": [],
"scenario": { "format": "square", "steps": [] },
"probes": [],
"facts": [{ "says": "a sentence the copy may use", "evidence": [] }]
}
]
}
| Field | Meaning |
|---|---|
order |
Position on the partners page. Ties in the rotation go to the lower number. |
designation |
The exact program name. Every spotlight states it word for word. |
page |
The partner's page on three.ws. One reply links it. |
forbidden |
Phrases that overstate the relationship. Any of them fails the lint. |
guidance |
Plain instructions for the writer: what the partnership is, and what it is not. |
features[].surface |
The page the head post links. |
features[].lane |
community, developer, token or labs, the same lanes as the queue. |
features[].scenario |
The proof scenario filmed for the reel (reel.js format). Its read steps become numbers the copy may quote. |
features[].facts |
Sentences with evidence: page, file, or proof ({ "type": "proof", "fact", "equals" } checks a value the run read; { "saw" } checks text the run saw). |
npm run x:partners -- check validates every file and names each missing piece. A feature that cannot be filmed against production (a surface behind credentials production does not have, say) does not belong in the registry: the series only announces what it has watched working.
Adding a partner: write its file, run check, then one --partner <id> --dry-run to film and draft it once without touching the queue. Adding a feature to an existing partner works the same way with --feature.
Rotation
pickNext chooses the partner seen least recently, breaking ties by order, then that partner's least recently featured feature. The head rotates through card, gif, reel and art, and the opening pattern through mechanism, walkthrough, clip and number, each avoiding the last one used. number is only chosen when the feature has a fact with a digit in it, so the series never leads on a number it cannot prove.
History is read from the queue itself: every item with series: "partners" carries spotlight: { partner, feature, head, pattern } and a notBefore day. --overlay <dir> adds the spotlights in a materialized production overlay, so a machine whose queue.json is older than production still rotates correctly.
Each spotlight takes its own day after every day already held. If today's series slot has already opened and passed, the next spotlight starts tomorrow.
Drafting
For each missing day, daily:
- Films the feature's scenario against production with
proveItem. The run's read values (counts, statuses) become facts the copy may quote. A feature whose scenario fails is skipped for the rest of the run, and the next partner in the rotation takes the day. - Builds a brief from the registry entry, the head and pattern, and the names of the other partners (so the copy can be checked for partners the feature does not involve).
- Asks the model chain (
callModelChain) for{ post, thread, headline, body, prompt, why }, with the house voice (announce-voice.md) and the keeper exemplar in the system prompt. - Lints the draft and sends every finding back for a rewrite, up to three attempts. Each language finding names the post it belongs to (
head,reply 2,headline,body), and the links rule lists the links it found, including a barethree.wsthat X turns into a link, so the rewrite touches the right post. - Builds the queue item (
itemFor), attaches the reel, builds the head from it (rebuildHead), validates it, writes the story todata/x-content/stories/<id>.json, and hands it tox:content adoptandadvance.
The series lint (seriesFindings) adds these rules on top of the pipeline's own:
- The designation appears word for word. No
forbiddenphrase appears. - No @mentions anywhere: a tag puts a partner's name next to ours, and that is a person's decision.
- The head post names the partner and links the feature's surface. One reply links the partner's page.
- One link per post (a bare
three.wscounts as one). - Three to five replies, and at least three registry facts quoted, in plain prose rather than quotation marks.
- No copy about where the media sits ("this reply", "the card above").
- No other partner is named unless the feature's own facts name it.
draftFindings then runs the pipeline's copy, language and claims checks, the head spec limits (a 90-character headline, a 200-character body), and a similarity check against the queue and the archive (limit 0.34).
Claims are never taken from the model. claimsFor builds them from the designation evidence, the registry facts the copy actually quotes, and the values the run read, so every number in a spotlight is one the ledger can prove.
The slot
The queue's tier 1 slot belongs to the series:
"slots": [
{ "tier": 2, "at": "12:00" },
{ "tier": 1, "series": "partners", "at": "15:30" },
{ "tier": 3, "at": "19:00" },
{ "kind": "article", "at": "22:30", "everyDays": 2, "from": "2026-10-08" }
]
A series slot spends its series' best ready post first, ranked across every tier. Posts of a series that owns a slot are kept for that slot, so the series posts once a day and never twice. When the series has nothing ready, the slot is the ordinary tier 1 slot, so the prime-time flagship never goes empty. It counts toward dailyCap like any daily slot.
Approval
Spotlights are tier 1, which normally waits for the owner. "series": ["partners"] in approval releases a listed series by policy whatever its tier, behind the veto window, and only when the other conditions hold: the post was filmed and the proof covers it, it tags no one, and the review passed on the editor's own verdict. Remove partners from the list and every spotlight waits for approve again.
The weekly Article
Once every seven days the series also writes a partner deep-dive as an X Article: one partner, every feature in its registry entry, filmed and measured that day. daily writes it when it is due (--no-article skips it), and article --partner <id> writes one on demand.
- Every feature's scenario is filmed against production. The lead feature's run binds to the Article's own id, so its reel is the proof the Article's post carries; the others are filmed under
<id>-<feature>. Four or more frames from the reels become inline figures inpublic/x-media/<id>/, and the cover is built from the lead reel. articleBlocksbuilds the Article's code and tables from data, never from the model: a table of the runs (surface, length, values read), a step log per feature (each step, its target, pass or fail and timing), a facts table per feature with the evidence behind each sentence, each scenario as JSON, for API surfaces the real request with the response production returned, and a roster table of every partner on the partners page with its designation and what three.ws built on it. The step log and roster keep a single-feature partner above the gate's code and table minimum.- The model writes the prose around them as
TITLE / POST / --- / markdown, placing every block once as{{block:NAME}}and the figures as{{figure:N}}.assembleArticlesubstitutes them and names any block missing, repeated or unknown. - The draft must pass the queue's Article gate (an 80 to 100 character title, 2400 words over at least eight sections, at least four images, enough code and tables, a section on every partner listed on the partners page, a Try it section) plus the series lint, with every finding fed back for up to three attempts. The model gets a 16,000 token budget for the Article.
- Claims come from the designation, the registry facts the prose quotes, and the lead run's values only (a non-lead run is not the proof the post carries, so its numbers stay in the tables). The item is a tier 1 Article with
spotlight: { partner, article: true, features }, written todata/x-content/articles/<id>.mdand handed toadoptandadvance.
Rotation goes to the partner whose last Article is oldest, ties by order. Articles post in the queue's Article slot, not the series slot, and wait for the owner's approve like any tier 1 Article.
Running it
npm run x:partners -- check # validate the registry
npm run x:partners -- plan --days 7 # the rotation ahead and how many days are in stock
npm run x:partners -- daily --days 3 --dry-run # film and draft three days, print them, write nothing
npm run x:partners -- daily --days 3 # write them to the queue and send them to review
npm run x:partners -- one --partner aws --feature marketplace-front-door --head gif
npm run x:partners -- article --partner ibm --dry-run # film and draft the deep-dive Article, write nothing
--ship on daily or one passes --ship to advance, which bundles approved spotlights for production through the overlay (x-content-pipeline.md, "Publishing without a deploy"). Shipping publishes, so it is an owner-approved step.
To keep the series in stock unattended, run daily once a day on a machine with a browser and ffmpeg, after materializing production's overlay (the bundles live on Google Cloud Storage when X_CONTENT_BUNDLE_BUCKET is set, otherwise on R2; the first pull on a bucket never published to succeeds with an empty overlay):
npm run x:bundle -- pull --dir /tmp/x-overlay
npm run x:partners -- daily --days 3 --overlay /tmp/x-overlay --ship
That run is packaged as a Cloud Run Job. deploy/x-partners/ holds the image (Node, Chromium, ffmpeg and a colour emoji font, so the reels film the way they do on a workstation) and its entrypoint, which pulls the overlay and runs daily --days 3 --ship. node scripts/create-x-partners-job.mjs prints the plan; --apply builds the image, creates the job three-ws-x-partners with the production service's env and secret references, and schedules it daily at 08:10 UTC through Cloud Scheduler. The task times out after three hours (10800 s) and never retries, because three spotlights plus the weekly deep dive outran the earlier one-hour limit. Applying it ships posts unattended, so it is an owner-approved step.
Keeping three days in stock means a failed run, a flaky surface, or a model outage costs nothing: the slot still has a reviewed spotlight waiting.