recurring-content-calendar

Build and run a recurring social content calendar — designed carousels, single-image posts, and short videos, written fresh each run and scheduled ahead to LinkedIn, Instagram, TikTok, and Threads. Use this skill whenever the user wants a content calendar, a daily or weekly posting routine, automate

Stand up a content calendar that keeps running: each run writes new posts, designs them, and places them in future publish slots, while checking that the last batch actually went live.

Publishing runs on PostNitro (carousels, single-image posts, short videos → LinkedIn, Instagram, TikTok, Threads). Sections 1–8 are specific to it. If the user wants a calendar on a different publishing tool, the intake, content, dedup, and audit logic still apply, but the tool-call mechanics do not.

Three layers to keep separate:

• Config (section 0) — the user's brand, accounts, cadence, timezone, voice, topics. Captured per user, never hardcoded into this file.
• API constraints (sections 2–8, and whatever the live schema says) — real limits. Breaking one returns an error.
• House rules — editorial and defensive choices that the API would happily let you break. Each is marked where it appears. A user setting this up for their own brand may change any of them; an agent mid-run should not.

The house rules, so a new user can see them in one place: the 50-second video ceiling, LinkedIn carousels as document, always filling TikTok postTitle, passing postSettings on every scheduled reel, DESIGN output rather than PDF/PNG, the slide and caption formatting tags, and the async import-then-poll pattern. All exist because something broke without them.

Look up live tool schemas every run rather than trusting parameter names remembered from a previous session. Never carry another workspace's IDs, API keys, publish times, or captions into this one — every value below comes from the user or from a live tool call.

0. Config

This is the whole per-user surface. Fill it during intake, echo it back in the recap, and store the filled version with the saved routine so later runs don't re-interview.

text

BRAND            — what is being promoted (company/product/person + one line on what it does)
AUDIENCE         — who the posts are for
VOICE            — educational tips (default) | product-led | founder-personal | other
LANES            — 5–8 recurring topics (derive from BRAND + AUDIENCE, see section 3)
ACCOUNTS         — the specific social account IDs to publish to
POSTS_PER_DAY    — distinct posts created per run
JOB_TIME         — clock time the routine runs (creation time, not publish time)
WINDOWS          — one publish window per daily post
DAYS             — weekdays only | all 7
TIMEZONE         — user's timezone (offset recomputed per run)
FORMAT_MIX       — how carousel / image / video rotate
IG_STORY         — feed only | feed + story
STATUS           — DRAFT while testing, SCHEDULED after approval
TEMPLATE/BRAND/PRESET IDs — from section 2

Intake

Ask one question at a time, skipping anything already answered in this chat. Use a question widget when there are a few real options. Do not assume a cadence, timezone, voice, or account list.

Required before anything goes live: API key (if not already connected), ACCOUNTS, POSTS_PER_DAY, JOB_TIME, WINDOWS, DAYS, TIMEZONE.

Ask, but take these defaults if they say "you pick": VOICE educational; FORMAT_MIX rotate so a week covers all three formats; IG_STORY feed + story; STATUS draft test first; LANES derived from BRAND and AUDIENCE.

BRAND and AUDIENCE are the ones most likely to be assumed wrongly — ask for them explicitly rather than inferring from the connected accounts' names. Do not import or schedule until the required items are answered. Then recap the filled config in one short message and wait for a yes.

1. Connect the publishing tool

Publishing runs through PostNitro. If it is not connected, it is usually absent from the catalog, so add the remote server:

• URL: https://mcp.postnitro.ai/mcp
• Header: Authorization: Bearer <pn- key>

Confirm with the user, then add. The key lives at PostNitro profile → Embed → Generate API Key, and starts with pn-. Tools appear on the next message, not immediately.

At first setup only, refresh capability facts from https://postnitro.ai/llms-full.txt or https://postnitro.ai/facts — which platforms publish natively, which post formats exist. Re-check if a platform or format later behaves unexpectedly, or if BRAND happens to be PostNitro itself. This is not worth fetching on every run of a daily routine; the live tool schemas are the authority for anything an API call depends on.

2. Discover the workspace

List templates, brands, AI presets, social accounts, audio, and saved defaults. Page if needed.

Check defaults every run. Call the get-defaults tool. If it returns hasDefaults: false, defaults are NOT set — pass templateId, brandId, and responseType explicitly on every import that run. Never assume defaults exist because you offered to save them earlier; a save can be wiped between runs.

If there is exactly one template / brand / preset, offer to save them as defaults with DESIGN (do not switch to PDF unless the user needs file URLs). MP4 cannot be stored as a default — pass it per call on the video tools, where a saved PDF/PNG default is treated as DESIGN anyway.

Show the social accounts you found and ask which to use. Store those IDs as ACCOUNTS — they are needed for scheduling and for the history/dedup query. If a platform in the plan turns out not to be connected, say so and skip it rather than failing the run.

3. Content rules

Content should solve a problem the AUDIENCE recognizes: strategies, tactics, specifics. BRAND is how you run the tactic, not the subject of a feature dump. No generic self-promo, no founder-personal voice unless VOICE says so.

Deriving LANES: from BRAND + AUDIENCE, propose 5–8 recurring topics the brand can credibly own, each narrow enough to yield many distinct posts. Show the list and let the user cut or add. Rotate through them so consecutive posts sit in different lanes. As illustration only — a project-management tool for agency owners might run: scoping, client comms, pricing, capacity planning, handoffs, retros, tool stacks.

Structure

• Carousel / video (5–7 slides or scenes): 1 starting_slide, 1 ending_slide, 3–5 body_slides. First slide moderate (curiosity), last slide moderate (close + where the brand fits), body elaborates.
• Single image: one slide object (not an array), no slide type. Only heading is required; add sub_heading, description, cta_button as needed. One complete thought. Infographic layout is allowed on this single slide.
• Infographics where they help. Call the import-template tool first for the exact structure.
  • layoutType: infographic replaces the image.
  • cycle: sequential, counters, all items in the first column.
  • grid: comparison, max 3 columns.
  • Caller-provided ids on every column and item. Item description is HTML.

Formatting reference

PostNitro's own markup, not markdown. An agent that reaches for **bold** or a real newline produces slides and captions with literal junk in them, so treat this as the complete allowed set. None of it appears in the import schema — the schema accepts any string — so do not "correct" these tags toward something that looks more standard.

Slide text (heading, sub_heading, description, cta_button) accepts:

Tag	Use	Limit
<strong>	The single most important phrase on the slide	Max one per slide
<em>	Light emphasis, a term being introduced	Sparingly
<u>	Underline	Sparingly; never inside <c>
<c>	Brand-colour highlight on headline words	1–3 words per headline; never combined with <u>
<br/><br/>	Paragraph break	The only way to break a line

Restraint is the point: a slide with three highlighted phrases reads as noise. One <c> on the hook and one <strong> in the body is a fully formatted slide.

Infographic item descriptions are HTML, not plain text, and want a wrapped paragraph: <p dir="ltr">3–5 posts per week</p>. A bare string may render unstyled. Column and item ids are caller-provided and never auto-generated.

Captions are one single line of plain text — see section 5. No markdown, no real newlines, <br/><br/> for breaks. The slide tags above do not apply in captions; <strong> in a caption publishes as visible characters on every platform. Hashtags are extracted automatically, so write them inline where they belong.

Everything else is out. No markdown headers, lists, links, or code formatting. No emoji as structural markup (as decoration in copy is a voice decision, not a formatting one). No single <br/> — the double form is the break.

CTA rotation: one CTA per post, differing from the most recent post (Save / Share / Follow). Use "Follow" sparingly, roughly 1 in 3. No engagement bait, no fake urgency.

Write copy in chat and lock it before importing, unless the user has already approved an ongoing pipeline.

4. Import

You write the copy. Import onto the saved template/brand, or the explicit IDs from section 2 when no defaults are set.

Use the async import pattern. Wait/synchronous convenience tools can take 15–60s and time out at the transport layer while the job keeps running server-side — so a retry after a timeout creates duplicate designs. Instead:

1. Call the async import tool (import_carousel / import_image / import_video) — it returns an embedPostId immediately.
2. Poll check_status with that embedPostId until COMPLETED, or until a step reports FAILED.
3. Call get_output to read the designId (and file URLs if PNG/PDF).

On a timeout, poll — never re-call the import. On a 429, wait ~15–55s and retry the same call.

• Carousel: slides array — exactly 1 starting_slide first, at least 1 body_slide, exactly 1 ending_slide last. Minimum 3 slides; type and heading are the only required fields per slide.
• Video: same slides array, each slide a scene. Duration and audio are set once in videoSettings, never per slide. videoDuration is variable, ≥5s per scene where it fits, and capped at 50s — the API accepts anything under 60, but 50 is the deliberate house ceiling, so do not raise it. videoSettings is required when responseType is MP4 and should be omitted on DESIGN imports — a DESIGN import creates no render, so duration and audio belong at schedule time instead (see section 6). Only reach for MP4 when the user explicitly needs file URLs. Video output is MP4 or DESIGN only — never PDF or PNG. audioId is a media ID from the library, not a URL.
• Single image: slide as one object, not an array, no type. Only heading is required. Use the image import tools, not the carousel ones.

Always pass:

text

generateImages: {
  context: "<you write this from the post>",
  imagePlacement: "auto",
  imageStrategy: "strategic"
}

Make context a text-safe composition brief: describe the theme, keep the focal subject in a lower or side third with clean negative space in the upper-center where the heading sits, high contrast, one clear visual metaphor, no text in the image.

AI images require a paid plan and draw on a separate AI-image quota (distinct from post credits). They are best-effort — the post still completes if they fail or the quota is exhausted. Confirm they landed via the GENERATE_IMAGES step in check_status. Infographic slides never receive an image; that is expected, not a failure.

Keep DESIGN unless the user needs file URLs. Do not use generate-from-topic unless they ask for it.

5. Captions

Unique per platform. Fill each one in; do not pad to the cap.

Captions are one single line of plain text. No markdown, no real newlines, no other formatting. For a line break, insert <br/><br/> and nothing else. Open every caption with the same hook line as slide 1.

Platform	Max	How
LinkedIn	3000	Longer, professional, light or no hashtags
Instagram	2000	Hook first, scannable, hashtags at the end
TikTok	2200	Spoken, hook first, native hashtags
Threads	500	Conversational, must stay under 500

Captions go in postContent, keyed by platform. The object also accepts common as a shared fallback and a facebook key, though Facebook is not among the four platforms the accounts tool returns. Hashtags are extracted automatically, so write them inline. Prefer per-platform keys over common — the point of this section is that each platform reads differently.

Lock captions on the first few posts the same way as slides.

6. Platform settings

Look up the live schema. Which settings blocks are required is conditional: Instagram, TikTok, and Threads settings are required when an account of that platform is selected and a designId is set, while LinkedIn settings are required whenever a LinkedIn account is selected, designId or not.

These are the defaults unless the config overrides them:

• Instagram: postType matches the asset (carousel | image | reel). postAsStory: true unless IG_STORY is feed only.
• TikTok: carousel or reel only — never image (the enum has no image option). Required fields: postType, canComment, isBrandedContent, isYourBrand, isThirdPartyBrand, isAIGeneratedContent. Always set postTitle to the post's title — treat it as required even though the schema marks it optional. SCHEDULED → privacyLevel: PUBLIC_TO_EVERYONE. canComment: true. Carousel: autoAddMusic (required for carousel). Reel: canDuet + canStitch (required for reel). Branded flags false — and note that if isBrandedContent is ever true, at least one of isYourBrand/isThirdPartyBrand must also be true, and it cannot combine with SELF_ONLY. isAIGeneratedContent: false unless the user says otherwise.
  • Single-image design: schedule LinkedIn/Instagram/Threads as image, then create a separate TikTok scheduled post as a reel with postSettings (duration ≤50) and a postTitle.
• LinkedIn: carousels are always postType: document with a postTitle of 5–90 chars when SCHEDULED (missing or too short returns 400). The enum also allows carousel, which is the PNG-output path — do not switch to it. A designId is required when SCHEDULED; text-only LinkedIn posts can only exist as drafts. Single image: image. Video: reel.
• Threads: postType matches the asset (carousel | image | reel).
• Reels: postSettings is schema-optional but necessary for video here. A DESIGN import carries no videoSettings, so omitting postSettings makes the API fall back to 30 seconds and silence. Pass videoDuration (≤50) and audioId on every reel.

7. History: dedupe + failure audit

Before writing, list social accounts, then list scheduled posts. This is both the dedup source and the health check.

Always pass socialAccountIds (the ACCOUNTS from config) on the list call. It is schema-optional, but omitting it returns every post in the workspace and can exceed the tool's output limit. Note the filter is post-level, not account-level: a post targeting two of the filtered accounts comes back with its full accounts array intact, so this narrows which posts return, not how large each one is — the response can still be big. Window: 30 days back through the next 7 days (fromDate/toDate are both required).

The result can be large. If it overflows, parse it with jq or hand it to a subagent to extract only what you need — do not read the raw blob into context.

Dedupe: do not repeat a design name, hook, or topic inside that window. Rotate the CTA away from the most recent post. Skip a publish window that already holds a SCHEDULED post for these accounts.

Failure audit (every run): SCHEDULED does not mean it went live. Scan for post-level FAILED or PARTIALLY_FAILED, and any per-account entry with a failed status or errorMessage. Known fragile cases: Instagram and Threads connections that silently fail — often with no per-account error string, where the aggregate status is the only signal — and render errors such as "Preparation failed: Render service failed". Surface these to the user (which post, which platform, what error). This routine cannot force a republish; flagging is the deliverable.

8. Schedule

scheduledAt is a future ISO-8601 time inside the chosen WINDOWS. This applies to drafts too — the API rejects a past time even for DRAFT, so a draft test still needs a real future slot. Never publish immediately on first setup.

A post needs either a designId or non-empty postContent. designId may be omitted or null only for DRAFT.

Pass the time with the user's UTC offset (for example +05:00 or -08:00); the API stores it as UTC. Recompute the offset per run for timezones with DST — never hardcode it.

Randomize the minute within the window and avoid reusing the exact minute from the previous day.

If the job runs late and a window has already passed, do not error and do not pick a past time. Use the next future slot and note the deviation in the run summary.

Use designId, never embedPostId — passing an embedPostId here fails with "Design not found." If import succeeds but scheduling fails, retry only the schedule; do not re-import. The combined import-and-schedule tool returns the created designId on a scheduling failure precisely so the design can be reused, which is the same reason to keep import and schedule as separate steps. When scheduling several posts in one run and a later one fails, keep the successes and report exactly what did and did not schedule.

First test is DRAFT. Only move to SCHEDULED after explicit approval.

9. First test, then automate

1. Recap the filled config from section 0. Get a yes.
2. Lock one post (copy + captions) in chat. Prefer a carousel for the first test unless the user asks otherwise.
3. Import DESIGN + generateImages using the async pattern in section 4.
4. Create a DRAFT on the chosen accounts, plus the extra TikTok reel draft if the design was a single image.
5. Send the editor URL. Stop creating posts until the user says to continue.
6. After approval, save a recurring routine as an intent, not a frozen tool recipe, carrying the filled config. Each run: dedupe + failure audit (section 7), write, import, caption using <br/><br/> only, apply platform settings including TikTok postTitle, schedule, then report what went out and flag any prior-run publish failures. If PostNitro auth fails, pause the routine and tell the user rather than retrying silently.

The calendar is the accumulated result, not a stored artifact: each run fills the next open slots and section 7's 30-day-back / 7-day-forward query is what reads the calendar back. So the run summary should say where things now stand — what was added, which slots are filled ahead, and anything from a previous run that failed to publish — rather than only confirming today's posts.

10. Product facts

• AI writes content (from a topic, URL, X post, or pasted text) and designs carousels, single-image posts, and videos. Users can also import their own copy.
• Native publish targets: LinkedIn, Instagram, TikTok, Threads. X is a source or export, not a native target.
• A brand kit's handle or tagline field may carry marketing copy. That is brand context, not a required closer on slides or captions.