Fifteen Instagram skills for Claude Code and Codex, merged from two MIT bundles, with the one thing neither of them had: a publish gate that is enforced rather than documented.
read your niche → write → show the draft → you approve → publish
↑
nothing passes this point
without a live, payload-bound approval
Two good bundles already exist, and this one takes from both.
sergebulaev/instagram-skills
is maintained, versioned and checked in CI. It closes the loop: it reads what is
working in your niche through Apify, writes, and publishes through Publora. Its
_env.py draws a trust boundary most projects never think to draw — a .env
may carry API keys but may never name a command to execute.
Jakeschincariol/instagram-agent-skill never connects to anything, and was pushed twice in forty-one seconds and left alone. But it carries three reference tables that are the real work: 26 hook formulas each with the trap that makes it fail, 199 anti-slop entries that come with their replacements, and a 12-criterion profile rubric.
The gap both share: the approval step before publishing was a convention. The upstream file said so in its own docstring — "a thin conventions layer, not runtime enforcement." A skill that forgot the card could still reach Instagram, and a draft that changed between the card and the call went out unapproved.
This bundle closes that gap, keeps everything else, and says exactly what came from where in NOTICE.
Three locks, in lib/approval.py. lib.publish() calls enforce() before any
branch that reaches a third party.
| Lock | What it stops |
|---|---|
| ARM | The bundle is read-only until INSTAGRAM_STACK_ALLOW_PUBLISH=1 is set in the environment you launch from. A .env cannot set it — same boundary _env.py already draws around executables. |
| BINDING | An approval is bound to a fingerprint of the exact payload shown. Approve caption A, publish caption B, and it fails. This is the lock that catches the realistic accident. |
| FRESH | Single use, expired after 15 minutes. An approval cannot be replayed onto a later post. |
Every attempt, allowed or refused, lands in .instagram-stack/publish-audit.log:
2026-10-30T16:24:01-0400 refused post 79aff2c0… never confirmed
2026-10-30T16:24:01-0400 refused post 9150402e… no approval for this exact draft
2026-10-30T16:24:01-0400 published post 79aff2c0… token 3adabf308e80
2026-10-30T16:24:02-0400 refused post 79aff2c0… already used
Six cases are proven in tests/test_approval_gate.py, and
scripts/check_approval_lock.py fails the build if a later edit lets
publish() step around the gate.
An agent running arbitrary code in this process can call confirm() itself.
These locks stop mistakes — a skill that skips the card, a test that
publishes for real, a draft edited after approval. They are not a sandbox, and
nothing inside a Python process can be.
If an account matters enough that a hostile agent is in your threat model, keep
PUBLORA_API_KEY out of the environment entirely. The bundle then drafts and
you post by hand, which is also how it ships by default.
git clone https://github.com/mehdijabry/instagram-stack.git
cd instagram-stackThat is the whole install for the drafting half. No keys, no accounts, no network. Point your agent at the folder and ask for a caption.
To let it read your niche or publish, copy .env.example to .env and fill in
only what you use. Then, deliberately, in the shell you launch from:
export INSTAGRAM_STACK_ALLOW_PUBLISH=1Leave that line out until you have watched the approval card appear on a dozen drafts. The bundle is useful without it.
Writing — produce text and files, touch nothing else
| Skill | What it produces |
|---|---|
ig-caption-writer |
Caption with the hook in the first 125 characters, a skimmable body, one CTA |
ig-carousel-planner |
Up to 10 slides: hook slide that opens a loop, value slides, payoff slide |
ig-reel |
Reel script: hook options, spoken line, on-screen line, beat sheet |
ig-story |
Story sequence frame by frame, sticker placement, order across the day |
ig-repurposer |
A LinkedIn post, blog, YouTube script or X thread rebuilt as native Instagram |
ig-hashtag-strategist |
3–5 sized tags (niche / mid / broad) a small account can actually rank in |
Reading — the only skills that leave your machine
| Skill | Reach |
|---|---|
ig-audience-insights |
Scans a hashtag and pulls profile stats through Apify. No login |
ig-hook-extractor |
Reverse-engineers the hook of a reel that worked, from a URL or a pasted caption |
ig-viral |
Finds what is travelling in your niche through the session's own browser, after you approve 6–12 accounts. Ranks by multiple of the account's baseline, not raw views |
Judging — read what you give them
| Skill | Reach |
|---|---|
ig-humanizer |
Strips the machine fingerprint using all 199 entries, then scores what is left |
ig-profile-optimizer |
Scores the profile out of 100 on 12 criteria, rewrites what loses points |
ig-audit |
Post-mortem on published posts, from statistics you paste |
ig-content-planner |
The week: what, which format, when, and who to engage. Schedules nothing |
Interacting — draft, you send
| Skill | Reach |
|---|---|
ig-comment |
Comments for other accounts' posts. Never auto-posts |
ig-reply |
Replies to your own comments, sorted by what is worth answering |
ig-dm from upstream B. The skill is well written and its guardrails are sound.
It is left out because a bundle aimed at an account a business depends on should
not make direct-message prospecting easy to run at volume — that is the shortest
path to a restricted account. Fetch it from upstream if you disagree; it is MIT.
What was locked inside single skills upstream is a library here.
from lib import hooks, scan_slop, desloppify, score_profile
hooks() # 26 formulas, each with its documented trap
desloppify(draft) # rewrite using the table's own replacements
scan_slop(draft) # report only — the writer decides
score_profile({"bio": 8}) # out of 100, with the three biggest losses nameddesloppify returns the rewritten text and every change it made. A humaniser
that edits silently is worse than one that does nothing. Structural entries —
"it's not just X, it's Y" — are reported, never auto-applied: deleting the
setup half is a judgement, not a substitution.
It removes a fingerprint. It does not make text undetectable, and no skill here may tell a user that it does.
lib.publish() routes to whichever backend is configured.
| Backend | When | Notes |
|---|---|---|
manual |
Default, no keys | Prints the copy-paste block. Needs no approval and no arming — it publishes nothing |
publora |
PUBLORA_API_KEY + INSTAGRAM_PLATFORM_ID |
Carried over from upstream A. A publishing API, not a bot driving the app |
diy |
INSTAGRAM_SKILLS_CUSTOM_POSTER |
Your own command, JSON on stdin. Must come from the launch environment |
Publora is the backend upstream A wired in, and this bundle kept the client
because it works. The sign-up link here carries no attribution parameters —
upstream tagged it so sign-ups were credited to the bundle author. Compare it
against Buffer, Later, Metricool or the Instagram Graph API directly before you
hand any vendor your business account; the diy backend exists so you can use
something else without touching this code.
python3 tests/test_approval_gate.py
python3 scripts/check_approval_lock.py
python3 scripts/check_no_secrets.py
python3 scripts/check_untrusted_rule.pyFill the voice profile properly. Both upstreams are calibrated on American
English. references/voice-profile.md is where a real business becomes
distinguishable from every other account using the same formulas: the words your
customers actually use, the objections they actually raise, and what you never
say. Skip it and the bundle produces generic content at speed, which is what
people already complain about.
Nothing here is a growth guarantee. These are hook formulas, a slop dictionary and a publishing client. They shorten the part of the work that is writing. They do not know your customers.
The work is upstream. See NOTICE for what came from where, file by file. Both projects are MIT; if you build on this, carry their names forward too.