Skip to content

Latest commit

 

History

History
456 lines (323 loc) · 9.06 KB

File metadata and controls

456 lines (323 loc) · 9.06 KB

1. Product Vision

Build an AI-powered content generation platform for software engineers.

The application should automatically generate 10–15 high-quality social media posts per day focused on:

- Artificial Intelligence
- Machine Learning
- LLMs
- Agentic AI
- Full-stack development
- JavaScript
- TypeScript
- Python
- Cloud
- AWS
- Kubernetes
- Open Source
- Trending GitHub repositories
- Trending AI papers
- Tech news

The primary platforms are:
- X (Twitter)
- LinkedIn

Each generated post should feel human-written, educational, and engaging rather than generic AI-generated content.

2. Goals

Goals:

- Generate selectively every day, with one or two draft windows and no obligation to fill them.
- Research current trends before writing.
- Avoid hallucinated facts.
- Cite sources internally.
- Never duplicate previous posts.
- Maintain a consistent writing style.
- Produce viral-quality content.

3. Features

Core Features

- User authentication
- Dashboard
- AI post generation
- Trend discovery
- GitHub trending integration
- Hacker News integration
- Reddit integration
- AI paper integration
- Save drafts
- Schedule adaptive draft windows and recommend X/LinkedIn publishing independently from measured results
- Publish to X, but not without my approval, my approval is must
- Publish to LinkedIn, but not without my approval, my approval is must
- Analytics
- Post history
- Search previous posts
- Topic management
- Writing style management
- AI model settings

4. AI Workflow

Workflow:

1. Collect latest news.
2. Collect GitHub trending repositories.
3. Collect AI research papers.
4. Collect Reddit discussions.
5. Cluster topics.
6. Rank by importance.
7. Generate content ideas.
8. Generate hooks.
9. Generate posts.
10. Score posts.
11. Rewrite low-quality posts.
12. Save approved posts.
13. Schedule publishing.

5. Agents

Create specialized AI agents.

Research Agent
- Finds trends.

Summarizer Agent
- Summarizes articles.

Content Planner
- Selects topics.

Writer Agent
- Writes posts.

Editor Agent
- Improves grammar.

Fact Checker
- Verifies claims.

SEO Agent
- Improves discoverability.

Publisher Agent
- Publishes posts.

Analytics Agent
- Learns from engagement.

6. Writing Style

Writing style:

- Sounds like a senior software engineer.
- Avoid marketing buzzwords.
- Educational.
- Opinionated only when supported by evidence.
- Short paragraphs.
- Uses code snippets where helpful.
- Uses emojis sparingly.
- No clickbait.
- Avoid AI clichés.

7. X Rules

Generate:

- Single tweets
- Tweet threads
- Polls
- Quote tweets
- Tips
- Code snippets
- Comparisons
- Hot takes
- Tutorials

Character limit:
280

Optimize for engagement.

8. LinkedIn Rules

Generate:

- Long-form posts
- Storytelling
- Technical breakdowns
- Lessons learned
- Architecture posts
- Career advice
- AI insights

Length:
500–2000 characters.

9. Tech Stack

Note: below is for reference, we must use those technologies which provide free service if number of requests is low, since this app is not for public to use, only I'm going to use it, so the requests wouldn't be that high and there're many platforms that give free access if the requests are low, for example I can use vercel for deployments, Tavily provide free api access if requests are low, supabase provide a free db access which is limited to 500 mb once the storage reaches like 450mb we should have a cron that will wipeout the DB since by that time the trends would have changed so no need to have them stored, instead of S3 we've Cloudflare R2; for AI model I will get Deepseek subscription which is the cheapest of all AI providers, for others we must try to find if there is any other service providing the same for free if the requests are low and are within some limits.

Frontend
- Next.js
- React
- Tailwind
- shadcn/ui

Backend
- FastAPI
or
- Node.js + NestJS

AI
- LangGraph
- LangChain

Database
- PostgreSQL

ORM
- Prisma

Queue
- Redis
- BullMQ

Scheduling
- Cron

Storage
- S3

Authentication
- Better Auth

Deployment
- Docker

10. Code Standards

Requirements:

- TypeScript everywhere.
- Modular architecture.
- Clean Architecture.
- SOLID principles.
- Repository pattern.
- Dependency Injection.
- Unit tests.
- Integration tests.
- E2E tests.
- Documentation.
- API versioning.

11. Database Design

Entities

Users

Posts

Topics

Sources

Tags

Research

Drafts

Schedules

PublishingJobs

Analytics

Templates

WritingStyles

Models

PromptVersions

12. AI Prompt System

Prompt templates should be versioned.

Support:

System prompt

Developer prompt

User prompt

Few-shot examples

Output validation

JSON schema

Automatic retries

13. Content Quality

Example:

Score each post based on:

Novelty

Accuracy

Hook quality

Readability

Virality

Technical correctness

Engagement potential

Overall score

Reject anything below a threshold (for example, 8.5/10) and regenerate it.


14. APIs

Include integrations such as:

  • X API
  • LinkedIn API
  • GitHub API
  • Hacker News API
  • Reddit API
  • arXiv API
  • Google News RSS
  • RSS feeds from major engineering blogs

15. Folder Structure

Ask Claude to generate something like:

apps/
packages/
agents/
prompts/
workers/
database/
scripts/
docs/
tests/

16. Roadmap

Ask Claude to implement in phases instead of trying to build everything at once.

Phase 1 — completed
- Authentication, dashboard, slot generation, screenshots, and manual posting workflow

Phase 2 — completed
- Product-first research, owned-project fact cards, GitHub/RSS and selective external evidence

Phase 3 — completed
- Platform-native X/LinkedIn generation, grounding audits, scoring, and rewrite loops

Phase 4 — completed
- Manual performance snapshots, analytics, and engagement opportunities

Phase 5 — completed
- Generation provenance, controlled growth experiments, bulk metrics, and approval-gated learning

Phase 6 — completed
- Grounded branded technical cards, LinkedIn carousels, visual settings, and media experiments

Phase 7 — completed
- Owned-repository sync, meaningful-change filtering, fact review, and approved evidence sources

Phase 8 — completed
- Manual platform distribution cycles, ranked conversations, relationship tracking, grounded replies, and audience content signals

Phase 9 — completed
- Goal-driven product campaigns, evidence-gated narrative stages, manual orchestration, and campaign analytics

Phase 10 — completed
- Privacy-safe tracked redirects, explicit conversions, full-funnel attribution, and controlled CTA experiments

Phase 11 — completed
- Production health, stage-level operational telemetry, cron monitoring, deployment validation, and checkpoint-safe recovery

Phase 12 — completed
- Deterministic seven-day growth reviews with prior-period comparison
- Exactly three approval-gated continue, reduce, and test decisions
- Safe content-mix changes with stale-review protection and draft-only experiment creation
- Historical review evidence, next-week briefs, and PDF/CSV exports

Phase 13 — completed
- 1h, 24h, 72h, and 7d capture queue for X and LinkedIn
- Comparable-age coverage and deterministic snapshot quality alerts
- Idempotent DevPulse, X, and LinkedIn CSV imports with audit history
- Explicit account follower and profile-view checkpoints
- Weekly-review decisions gated on valid 24-hour cohorts and coverage confidence

Phase 14 — completed
- Deterministic seven-day execution plans generated from persisted weekly reviews
- One product-first anchor per day while remaining slots keep the approved content strategy
- Individual rejection, whole-plan approval, cancellation, skipping, and completion controls
- Approved anchors guide matching slot generation without bypassing draft review or manual publishing
- Passive iCalendar export plus explicit publish and valid-24h measurement confirmation

Phase 15 — completed
- Adaptive generation cadence with two X-oriented daily draft windows by default
- Independent LinkedIn weekly publishing days and per-platform recommendation queues
- Evidence, overall-quality, novelty, project-cooldown, and content-type-cooldown gates
- Intentional slot skips when no candidate clears the configured publishing bar
- Measured posting-hour recommendations with conservative small-sample fallbacks
- Daily manual engagement sequence before publishing, after publishing, and at the 24-hour checkpoint

One additional recommendation

One feature that can make this stand out from typical AI post generators is to make it research-first instead of prompt-first. Rather than asking an LLM to invent posts, have the system:

  1. Continuously ingest fresh sources (GitHub, arXiv, Hacker News, Reddit, company engineering blogs, AI news).
  2. Deduplicate and rank topics by relevance and novelty.
  3. Generate multiple content angles for each topic (tutorial, opinion, comparison, quick tip, thread, architecture breakdown).
  4. Learn from engagement data over time to improve future recommendations.