Tt Caption Writer Agent Skill

TikTok Caption Writer

Write the caption that sits under the video: a tight caption inside the 2,200

char API limit, a mixed-reach hashtag set, and the platformSettings.tiktok

flags that control reach and interaction. On TikTok the caption supports the

video, it does not carry it. The hook is in the first 1-3 seconds of the clip

(see tt-hook-scripter); the caption's job is to add context and earn a comment.

When to use

• User has a video (or a script) and needs the caption, hashtags, and settings
• User is about to upload and wants the posting flags right
• User wants Claude Code or Codex to schedule a rendered .mp4 via Publora

What this skill produces

• Caption (<= 2,200 chars on the API; aim much shorter), first line front-

loaded with a reason to read or comment

• Hashtag set (3 to 5, mixed reach), placed at the end
• platformSettings.tiktok flags with a plain-English summary
• On approval, the publish call (draft -> upload -> schedule) when a video file

is supplied

The TikTok settings (platformSettings.tiktok)

Build these with lib.tiktok_settings(...). Defaults match Publora's.

| Setting | Values / default | What it does |

|---|---|---|

| viewerSetting | PUBLIC_TO_EVERYONE (default), MUTUAL_FOLLOW_FRIENDS, FOLLOWER_OF_CREATOR, SELF_ONLY | who can view. Effectively required; an empty value is rejected |

| allowComments | true (default) | viewers can comment |

| allowDuet | false (default) | viewers can Duet |

| allowStitch | false (default) | viewers can Stitch |

| commercialContent | false (default) | the video is commercial |

| brandOrganic | false (default) | promoting your own brand |

| brandedContent | false (default) | paid partnership / sponsored |

Two gotchas you must surface to the user

• Boolean inversion bug. Publora currently maps allowComments,

allowDuet, allowStitch to TikTok's disable_* flags, so the booleans can

land inverted (sending allowDuet: true may disable duets). The status cannot

be confirmed from the API alone. **Test with a SELF_ONLY draft before

trusting these values.** Tell the user this whenever they change a flag.

• Commercial disclosure. If commercialContent is true, at least one of

brandOrganic (your own brand) or brandedContent (paid partnership) must

also be true, or Publora rejects it. lib.tiktok_settings enforces this.

Posting reality

• Unaudited apps post PRIVATE only. Until the publishing app passes TikTok's

review, posts are forced to SELF_ONLY regardless of viewerSetting. Warn the

user if a "public" post lands private.

• For reach, use PUBLIC_TO_EVERYONE. For an end-to-end test, use SELF_ONLY.

Steps

Voice profile first (all drafts). If ../../references/voice-profile.md has filled: yes, load it and match the user's voice fingerprint, hard rules, and CTA/link style throughout. If it is not filled, mention once that tt-humanizer --mode profile can learn their voice from a few posts, then proceed with the generic voice rules.

Never invent the specifics. The rules below ask for a concrete number, a date and a named entity, because that is what separates a real post from a generated one. Take them from what the user actually said in this conversation. Do not invent a figure, a date, a client name or a result, and do not soften a vague claim into a plausible-looking number. If the user has nothing concrete for a beat, ask them once, and if they still have nothing, drop the claim rather than decorate it. A published invented number is a retraction; a missing one is only a weaker post.

• Gather inputs, goal first. Ask (or infer) what the caption should earn,

then shape for it: comments -> end on one specific question tied to the video;

saves -> promise reference value ("the 3 settings are in the caption");

shares -> a claim people will want to show someone; profile taps -> an open

loop the pinned video pays off. Then gather the video topic/script, the

niche, whether it is commercial, and whether a rendered

.mp4 path exists.

• Draft the caption. Front-load the first visible line with a reason to read

or a specific question to comment on. Keep it tight. Em dashes under the cap

(about one per 100 words; a short caption rarely needs one), no AI vocab

cluster. Optionally restate the hook's open loop as a comment prompt.

• Pick hashtags. 3 to 5, mixed reach: one broad, one to two niche-defining,

one to two specific to the video. Put them at the end. Never stuff.

• Set the flags. Default to PUBLIC_TO_EVERYONE, comments on, duet/stitch

off, no commercial flags. Adjust per the user, and surface the two gotchas.

Build with lib.tiktok_settings(...).

• Char check. Confirm caption + hashtags <= 2,200. If over, tighten.
• Humanizer pass. Scrub 2026 AI vocab by density, cap em dashes (about

one per 100 words, never swap one for a period), break stacked triads,

generic openers, reveal bridges and sincerity openers ("not gonna lie").

Leave the rhythm alone; never chop a line for punch. Canonical rules:

tt-humanizer V3.

• Approval card. Show: caption, char count, hashtag set, settings summary,

and the resolved viewerSetting (flag if it may post private).

• On approval. Call `lib.publish("video", caption, target_url=<upload or

profile URL>, video_path=<path or None>, platform_settings=<settings>,

scheduled_time=<iso or None>, platforms=[<TIKTOK_PLATFORM_ID>])`. With a video

path and Publora configured, it runs draft -> upload -> schedule. Without one,

it returns the caption and settings to upload in-app.

Hard rules

Global voice rules: see root SKILL.md Voice rules. Additional skill-specific

rules:

• Caption + hashtags must fit 2,200 chars (API), hashtags included. Aim far under.
• 3 to 5 hashtags, mixed reach, at the end. Never 10+, never mid-sentence.
• The caption never does the hook's job. The hook is in the video.
• Always set viewerSetting. Never ship an empty value.
• Surface the boolean inversion warning whenever an interaction flag changes.

Anti-patterns (skill will refuse)

• Hashtag walls (10+), or hashtags jammed into the sentence.
• Em dashes above the cap (more than about one per 100 words), or an em dash swapped for a period.
• "The result?" reveals and "No X. No Y. Just Z." staccato stacks added for punch.
• "Follow for part 2 / like and subscribe / don't forget to share" stacks.
• A caption that repeats the spoken hook word for word.
• commercialContent: true with neither brand flag set.
• Claiming a post will be public when the app is unaudited (it will be private).

Resources

• ../../references/voice-rules.md - caption and hashtag rules
• ../../references/algorithm-heuristics.md - caption, hashtag, and settings heuristics
• references/settings-matrix.md - every platformSettings.tiktok value, defaults, and the known bugs
• lib/publora_client.py - tiktok_settings(...), publish_video(...), create_draft(...)

Optional illustration

Offer a generated image when a visual would lift reach. Draft a prompt and call

lib.illustrate(prompt, kind="story"), pulling brand handle/color from Voice &

Brand Profile section 6 for a pixel-exact overlay. Show the returned url + cost,

for a TikTok photo post, attach the image via media_urls=[url] (TikTok photo carousel); for a video, TikTok's API has no cover-image field, so set the cover in the TikTok app. Full workflow (incl. quote-cards):

../tt-humanizer/sub-skills/illustration.md. No Pixfaro key -> it drafts the prompt for you to generate manually.

Related skills

• tt-hook-scripter - the in-video hook (the caption does not replace it)
• tt-humanizer - scrub the caption before publishing
• tt-content-planner - schedule captions across a posting week