LinkedIn Content Engine

A local app that researches, writes, checks and publishes LinkedIn posts, and helps you comment and reply, with you approving every word.

Built by Arun Kirupa, founder of Pro Marketer. It runs on FastAPI, the Claude API and the LinkedIn REST API. The writing playbooks in knowledge/ come from linkedin-skills by Serge Bulaev (MIT).

📖 New here? Read the full User Guide: setup, daily routine, every page explained, and troubleshooting.

---

Screenshots

Home: what needs you today, setup checklist, what's coming up, Engage tasks, AI spend	Studio: pick a goal and a proven hook formula, add your facts, get a checked draft with a live LinkedIn preview
Queue: every draft gets a quality score before you approve it	Queue editor: the quality check lists AI tells and reach problems, with Auto-fix and Fix with AI
Plan: a week of posts across your content pillars; draft any item into the Queue	Engage: comments and replies you approved post one at a time, with Copy + Open as a fallback
Brand Voice: the guided editor for how you sound, plus Learn my voice from your own posts	Schedule: posting times and scheduled posts in your timezone
History: published posts; type in LinkedIn's numbers so the app learns what works	Settings: LinkedIn connection test, limits, posting times, AI usage by feature
Profile Optimizer: a 9-part scorecard with headline, About and experience rewrites

Screenshots use demo data: posts already published on LinkedIn and a sample queue.

---

How it works

flowchart LR
    R[Research<br/>Google Trends · Reddit · RSS<br/>competitors] --> D
    B[Brand Voice<br/>soul/soul.md] --> D
    K[Playbooks<br/>hook formulas · voice rules] --> D
    D[Drafting<br/>Studio · Plan · Idea Lab<br/>scheduled batches] --> Q{Quality gate}
    Q -->|auto-fix, 1 AI repair,<br/>fact-check, reach score| U[Queue<br/>you review]
    U -->|Approve / Schedule| P[Publish to LinkedIn<br/>daily cap + spacing]
    P --> F[First comment<br/>links go here]
    P --> S[Stats<br/>API or typed in]
    S --> L[Learnings<br/>feed future drafts]
    L --> D
    E[Engage<br/>comment + reply drafts] --> A{You approve}
    A -->|paced API posting| LI[LinkedIn]
    A -->|fallback| C[Copy + Open post]

Loading

1. Your voice comes first. soul/soul.md, edited on the Brand Voice page, holds who you are, your business, your audience, your pillars, your voice fingerprint and words you never use. Every Claude call gets this as its system prompt, along with the relevant playbooks from knowledge/.
2. Drafts come from four places:
   • Studio: goal → formula → your facts;
   • Plan: a guard-railed week of posts;
   • Idea Lab: variations of a rough idea;
   • scheduled batches from research.
3. Every draft passes one quality gate:
   • formatting for LinkedIn (the 25-line limit, clean characters);
   • safe automatic fixes;
   • an AI-tell audit (question openers, reveal bridges, clichés, too many em dashes, links in the body, placeholders);
   • at most one "Fix with AI" pass;
   • a combined fact-check and reach score.
4. You review in the Queue. Nothing is published without your approval. Approved posts go out at your posting times, never more than POSTS_PER_DAY. If a post has a link, it goes into the first comment, which the app posts right after.
5. The loop closes. The app learns from real results, either read from LinkedIn (if your app has r_member_social) or typed into History. Learnings feed back into future drafts. Autoresearch experiments only start after 10 real posts.
6. Engage grows your reach between posts. It drafts comments on posts in your niche and replies to comments on yours. Approved items post through the LinkedIn API one at a time, 90–180 seconds apart and under a daily cap. If LinkedIn refuses, everything switches to Copy + Open post.

What's inside

Area	Page	What it does
	Home	Setup checklist, drafts to review, what's coming up, Engage tasks, AI spend
Create	Studio	Write: goal → proven hook formula (F1–F20) → your facts → draft. Repurpose: turn an article, newsletter or transcript into up to 3 posts. Hook Lab: see why a hook works and save it as a template
	Plan	A 3–5 post week across your pillars. No pillar above 60% and no formula twice in 7 days, both enforced in code
	Idea Lab	Turn a rough idea into several variations
Review	Queue	Quality score on every draft. The editor shows the issues, a LinkedIn preview with the "…see more" fold, and Auto-fix / Fix with AI
	Schedule / History	Scheduled and published posts, failed posts with retry, and stats you type in from LinkedIn
Engage	Engage	Comment on others' posts (T1–T7), reply to comments on yours (R1–R5), daily targets, follow-ups, activity
Insights	Research, Analytics, Learnings, Competitors	Trending topics, what works for you, competitor patterns. Experiments appear once autoresearch is on
Setup	Brand Voice	Guided editor for soul/soul.md, plus Learn my voice
	Profile Optimizer	9-part profile scorecard with rewrites
	Settings	Connection test, limits and toggles, posting times, AI usage by feature

Images: the Gen Image button creates a post image with kie.ai, or with Google Gemini when only that key is set. Images avoid invented numbers, logos and real people.

With KIE_IMAGE_MODEL=auto (the default) the post decides the model:

The post is about	Model	Cost and speed
Illustration, cartoon, sketch	bytedance/seedream	~3.5 credits, ~7s
Products, photoreal scenes, detail, text in the image	nano-banana-2 (2K)	~12 credits, ~45s
Everything else	google/nano-banana	~4 credits, ~15s

If a model is down at kie.ai, the app retries once with google/nano-banana, so a post still gets an image. Set KIE_IMAGE_MODEL to fast, detail, poster or any kie.ai model id to always use one.

Safety and security

• Approval gate. Posts, comments and replies are only published after you approve them. Post now, the scheduler and Engage all enforce it.
• Limits. POSTS_PER_DAY (default 1) applies on every publish path, plus a minimum gap between automatic posts. ENGAGE_DAILY_CAP limits comments, which are spaced out. If LinkedIn doesn't confirm an item, it's marked "check on LinkedIn" and never retried, so nothing posts twice.
• No invented facts. Drafts only use your Brand Voice and the notes you give. Missing facts become [placeholders], which block approval until you fill them.
• Local and locked down.
  • The app listens on 127.0.0.1 and refuses a network address unless DASHBOARD_PASSWORD is set.
  • It only answers requests addressed to localhost / 127.0.0.1, which blocks DNS rebinding.
  • Cross-site requests are blocked, and pages can't be framed by other sites.
  • Server-side fetches only go to public hosts.
• Untrusted text stays inert. Scraped and pasted text is marked as untrusted in every prompt and escaped on every page. A test fails the build if a page brings back an unsafe pattern.
• Secrets stay local. Keys live in .env, which is gitignored. The database, which holds your LinkedIn login, is owner-only. A pre-commit check blocks commits containing keys.

Quick start

git clone https://github.com/Only-Pro-Marketer/Linkedin.git
cd Linkedin
python3.12 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
cp .env.example .env        # add your keys
python app.py

Open http://localhost:8000 and follow the checklist on Home:

1. Add ANTHROPIC_API_KEY to .env.
2. Fill in Brand Voice. Paste 3–6 of your posts into Learn my voice for a head start.
3. Connect LinkedIn (Settings → Connect).
4. Check your posting times in Settings.
5. Write in Studio, review in the Queue, and approve.

Key	Required	Where to get it
ANTHROPIC_API_KEY	Yes	console.anthropic.com
LINKEDIN_CLIENT_ID / LINKEDIN_CLIENT_SECRET	To publish	linkedin.com/developers: enable Sign In with LinkedIn using OpenID Connect and Share on LinkedIn; redirect URL http://localhost:8000/auth/callback
KIE_API_KEY	No	kie.ai for post images. Used instead of Gemini when set
APIFY_TOKEN	No	Optional scraping (competitors, your own posts, fetching comments). Everything also works by pasting
REDDIT_CLIENT_ID / GEMINI_API_KEY	No	Extra research source / Gemini images

LinkedIn permissions

• w_member_social (default): publish posts, comments and reactions.
• r_member_social (needs LinkedIn approval): read post stats automatically. Without it, type the numbers into History.

Run with Docker instead

Same app, no Python setup. You still need .env filled in first.

docker compose up -d --build

Open http://localhost:8000. docker compose logs -f shows what it's doing, and docker compose down stops it.

• The port is published to your machine only, so the dashboard is never reachable from your network.
• Your keys (.env), the database, your brand voice and generated images live on your computer and are mounted into the container, so nothing is lost when the container is rebuilt or removed.
• Keys and the database are never copied into the image.

Configuration

Secrets live in .env. Limits and toggles can also be changed in Settings; those values override .env.

Setting	Default	What it does
POSTS_PER_DAY	1	Hard cap on publishes per day (1–100)
MIN_HOURS_BETWEEN_POSTS	3	Gap between automatic publishes
ENGAGE_DAILY_CAP	30	Comments and replies per day
SCHEDULER_ENABLED	true	Background jobs on or off
AUTO_REPAIR / FACT_CHECK_ENABLED	true / true	AI repair pass and fact-check for new drafts
AUTORESEARCH_ENABLED	false	Claude-scored experiments (idle until 10 real posts)
CLAUDE_MODEL	claude-opus-5	Model for every Claude call
KIE_IMAGE_MODEL / KIE_IMAGE_ASPECT	google/nano-banana / 4:5	Post images
POSTING_TIMEZONE	America/Toronto	All schedule times use this zone
HOST / DASHBOARD_PASSWORD / ALLOWED_HOSTS	127.0.0.1 / empty / empty	Network exposure; set a password before opening the app to a network

What it costs to run

• Claude: pay per use. Settings → AI usage shows calls and estimated cost by feature for the last 30 days.
• kie.ai images: about 4 credits per image with Nano Banana.
• Apify (optional): about $0.002 per scraped post, $0.01 per profile and $0.005 per comment. Daily imports cost cents per competitor.
• LinkedIn API: free.

Development

pip install -r requirements-dev.txt
python -m pytest -q

requirements.txt holds only what the app needs to run. The optional screen-record and deep-research features need a browser, which comes with the dev requirements (playwright install chromium once). Keeping it out of the runtime list keeps about 134 MB out of the Docker image.

Tests use a temporary database, a fake Claude and mocked LinkedIn, kie.ai and Apify, so they need no keys or network. See CLAUDE.md for the architecture and the rules the code relies on.

app.py                     FastAPI app: middleware, routers, startup recovery
config.py / app_settings.py  .env settings / values edited in Settings
llm.py                     The only place that calls Claude (caching, refusals, usage log)
knowledge/                 Vendored playbooks from linkedin-skills (MIT) + loader
soul/soul.md               Your brand voice
content/                   brand, prompt_builder, pipeline (quality gate), humanizer + quality_rules,
                           ai_fix, review (fact-check + reach score), generator, image_generator, templates/formulas
studio/                    writer, repurposer, hook_lab, planner, voice, profile
engagement/                comment_drafter, reply_handler, publisher (paced API posting), followups
linkedin/                  api_client (posts, comments, reactions), poster (approval gate + limits), url_parser
post_queue/                queue state, scheduler jobs, calendar slots
research/ analytics/ autoresearch/   research sources (incl. optional Apify), learning loop, experiments
dashboard/                 routers/, routes.py, security.py, templates/, static/js (app.js, post-editor.js)
database/                  models, engine (SQLite WAL), migrations (versioned, backed up)
tests/                     pytest suite

Post status flow: QUEUED → APPROVED → (SCHEDULED) → POSTING → POSTED, or REJECTED / FAILED (with retry).

---

About the author

Arun Kirupa is the Founder and Head of Strategy at Pro Marketer, a Toronto e-commerce growth agency he started in 2018.

• Helps e-commerce owners grow with the proprietary Growth Trust Model®: a 90-day plan built around CRO, AOV, retention and paid media.
• Shopify specialist for supplement, apparel and skincare brands. Pro Marketer is a Klaviyo Partner and a Zoho Authorized Partner.
• Host of the Pro Marketer CheckOut Podcast, where he interviews e-commerce founders.
• Founder of the "DTC Owner" community on Skool.
• Previously a contract CMO leading a 40+ person marketing team.

He built this engine to run his own LinkedIn the way he runs client growth: a clear process, real numbers, and a human approving every word.

• LinkedIn: linkedin.com/in/arunkirupa
• Website: promarketer.ca

License

MIT, see LICENSE. knowledge/ is adapted from linkedin-skills by Serge Bulaev (MIT, see knowledge/LICENSE).