instagram-skills: nine agent skills that draft Instagram content and hold it for approval

What the skill bundle is actually for

The repository is a set of nine skills that turn a coding agent into a drafting assistant for Instagram, sold on a simple premise: you supply the image or video and the skills write the words. The stated audience is anyone who asks Claude Code or Codex for Instagram help, and the README says no coding is required. The scope is broad but the surface is small. A caption writer, a carousel planner, a hook extractor, a hashtag strategist, a humanizer and a content planner sit at the core, with the rest covering profile optimization, reading a niche from live data and publishing. The reading and writing sides are split: the writing skills work with nothing configured at all, while the read side needs a token, and the publish side needs a paid-style service account. That split is the design decision that makes the project usable for someone who only wants better captions.

Draft first, publish only after you say OK

Every skill in the bundle shows a draft and waits for approval before anything is sent anywhere. The project states this twice, once in the feature list and once at the end of the capability tour, which suggests it is the point of the project rather than a footnote. The mechanism is straightforward: the skills write text, the agent shows it to you, and a separate publishing step runs only when you approve. The publishing path itself is Publora, a third-party service reached through a small client library, and there is an escape hatch for people who already have their own poster. The environment file describes that escape hatch as a Tier 2 advanced setting where you point the skills at the command your own poster runs, and the file says to leave it unset unless a custom poster is actually working. The practical consequence is that a broken custom poster is a silent no-op risk worth checking once before you rely on it.

Installing the bundle across five different agent hosts

There is no package index entry. Everything is a git clone, a plugin marketplace add, or a copy-paste of the repository name into a skills dialog. The Codex CLI path adds the repository as a marketplace and then installs the plugin from it, and the same two-step pattern exists for Claude Code with slash commands. The Claude web and desktop apps take the repository name through an Add from GitHub dialog. OpenClaw users clone the bundle into a working directory and add a system prompt telling the agent to read the relevant skills directory first and to use the URL parsing and publishing modules from the library folder. A single command covers the widest set of hosts, including Cursor and anything else that reads SKILL.md files. Note that the OpenClaw and Hermes paths add a prompt block, and that block is part of the setup, not a nice-to-have.

npx skills add sergebulaev/instagram-skills

Two environment variables decide whether you can publish

The environment example is unusually clear about what is optional. Every variable is marked optional, and the comment at the top says the skills work in draft-only mode with nothing set, where they produce the caption, the hook, the hashtags and the slide plan for you to post in the app yourself. Publishing needs a Publora API key and an Instagram platform connection identifier, and the file spells out the format of the latter as a prefixed identifier you copy whole from the Publora dashboard, under Channels then your Instagram account. It also states a hard platform constraint: Instagram requires a Business or Creator account and personal accounts are not supported. A third variable, an Apify token, is described as optional and only needed if you wire an Instagram reading actor, otherwise the hook extractor falls back to asking you to paste the caption text. That fallback is worth knowing, because it changes what the reverse-engineering skill can actually do.

PUBLORA_API_KEY=sk_your_key_here
INSTAGRAM_PLATFORM_ID=instagram-your_id_here
APIFY_TOKEN=apify_api_your_token_here

The hook rules are hard-coded for one year

The caption writer puts the hook in the first 125 characters and picks a formula by goal, with the options being saves, shares, comments and follows, and the project documentation calls it a 2026 hook formula. The hook extractor maps a viral Reel or carousel to one of ten Instagram formulas and hands back a blank template to fill in. The hashtag strategist returns a set of three to five hashtags split into niche, mid and broad tiers, sized for what a small account can plausibly rank for, which the project frames as the 2026 reality in place of the old 30-tag approach. The carousel planner works slide by slide up to ten, opening a loop on the first slide and front-loading value. These are opinionated rules encoded as instructions, and they carry a date. A team operating in 2027 should treat the year-qualified claims as hypotheses to test against their own feed rather than as current platform truth.

The Humanizer edits style and makes no detection promise

The Humanizer is the most interesting skill in the set and also the one with the sharpest honesty. The project describes it as removing the tells human readers react to: a vocabulary list scored by paragraph density, reveal bridges, staccato fragment stacks, stacked triads, performed sincerity and emoji storms. It caps long dashes at roughly one per hundred words rather than banning them, on the stated grounds that a share of human captions use one. And it says outright that it does not promise to beat detectors, with the parenthetical that no edit reliably does. The skill also bundles a pre-publish audit mode, so you can run a check on a draft before it goes out. The limitation is honest but important: this is a style pass over words, not a guarantee about how a caption will be received or judged.

Where this bundle is the wrong tool

Several limits come from the host platform rather than the code. The skills are prompts plus scripts, so their output quality tracks whatever model runs them, and a formula picked for a specific goal can misfire when the content does not fit the goal. The publishing path routes media and scheduling through a third-party service, which means the account credentials live outside the agent, and the extra layer is a dependency you do not control. The Apify side works the same way, since the read side depends on a token for a hosted actor. Beyond that, the project is scoped to one platform, and the sibling repositories for X, TikTok, YouTube, Threads and Facebook are described as sharing the same voice engine, so a team running several platforms has to install and maintain each one separately. Nothing here is a general content system, a scheduler, or an analytics product.

Release cadence, licence and what the repository contains

The repository is not archived, and the last push was on 2026-10-15. Three releases landed on consecutive days in mid October, v1.0.30, v1.0.31 and v1.0.32, which points to frequent iteration rather than a slow release train, and it means pinning a version matters if you want a stable prompt set. The licence is MIT, and the project credits Creative Content Crafts for content engineering, with the homepage given as cccrafts.ai. The repository layout shows the parts that matter: a top level SKILL.md and AGENTS.md for agent instructions, a skills directory holding the individual SKILL.md files, a library folder with the URL parsing and publishing modules the OpenClaw prompt points at, a references directory, a scripts directory, and requirements.txt plus a lock file pinning requests and python-dotenv. There is also a SECURITY.md and a .env.example, which is a reasonable sign for a project that asks users to hold API keys.

Editorial conclusion

Adopt instagram-skills if you are a solo operator or small team already working inside Claude Code or Codex who wants Instagram drafts in a consistent voice, and if you are willing to treat the hooks and hashtag rules as templates to edit rather than as proven formulas. Do not adopt it for regulated or high-stakes publishing, since the Humanizer is explicitly a style pass and makes no claim about detection, and the publishing path depends on a third-party service. Before installing, check that your account type qualifies for the Instagram connection the project describes, decide whether the draft-only mode is enough, and read the nine SKILL.md files to see which formulas are hard-coded for 2026.

Frequently asked questions

Official sources

1. License: MIT
2. Project website
3. README
4. Releases
5. sergebulaev/instagram-skills on GitHub