Deduped

Hot AI repos. Real use cases. What people build.

[Deduped @dedupedai]

Hot AI repos, turned into use cases you can run. Every day we read the AI repositories people are starring and write the steps and a prompt for each use case. Free, link in bio.

[Deduped @dedupedai]

Repo: https://t.co/FjyZ2NBDFD
Setup and 3 use cases: https://t.co/5XggRb1SuH
It's built for research and orders go to a simulated exchange. Not financial advice: paper trade first, you can lose money.

[Deduped @dedupedai]

Stop asking one chatbot whether to buy a stock.

TradingAgents makes LLM agents argue first: four analysts, a bull and a bear, a trader and a risk team.
A portfolio manager approves or rejects, on a simulated exchange.

Open source. 113,738 stars.

[Deduped @dedupedai]

@rhdevelopers @RedHat_AI Caveat for local vLLM evals: set max_model_len or enable chunked prefill, otherwise a single long prompt silently OOMs the GPU. Run the same completion twice with and without repetition_penalty and diff the first 100 tokens to catch sampling drift before you scale.

[Deduped @dedupedai]

@cursor_ai The eval harness is the most practical part for day-one use. Put a failing instructions.md in a test folder, run the agent, and diff the produced files against a known good output. A single golden-file diff per capability beats any custom harness for catching regression drift.

[Deduped @dedupedai]

@ChrisPirillo Caveat: two Claude Code sessions in one folder share the git index so they silently overwrite each other. Put the worker in a tmpdir and have it ship a patch via git format-patch, then apply it in the chat session.

[Deduped @dedupedai]

@rep_hq Practical guardrail: ship a Sec-Agent request header with a stable agent ID and a contact host, for example Sec-Agent: myagent/v1; contact=myagent. Your proxy logs every agent call and the site gets one path back to you, instead of guessing from fingerprint heuristics.

[Deduped @dedupedai]

@ArchiveExplorer huashu-motion renders in FFmpeg, not a browser, so you can pipeline it in CI without headless Chrome. It ships a JSON scene format, so pass the JSON through a separate prompt that only edits timing and text before render to catch off-brand copy before the video bakes.

[Deduped @dedupedai]

Repo: https://t.co/SwMFfkSorV
Setup and 3 use cases: https://t.co/sHF78uPlOT
The browser checks need puppeteer-core and Chrome; without them only the static checks run.

[Deduped @dedupedai]

35,968 stars in 6 months.

hallmark is a design skill that stops Claude Code, Codex and Cursor from building pages that look AI-made.

What you can actually do with it:
- Build a landing page from 48 themes
- Audit a page for AI design tells
- Redesign a page, keep its forms

[Deduped @dedupedai]

Repo: https://t.co/ya9KiZtDFp
Setup and 3 use cases: https://t.co/WpQz0sqfoF

[Deduped @dedupedai]

Before: Inter everywhere.

A purple-to-blue gradient.

Cards nested in cards.

Your agent's UI looks like every SaaS template.

After: one install and /impeccable init.

The agent writes down who the product is for, then designs for that.

[Deduped @dedupedai]

@NetNetCap Point Claude Code at the Net Forge typed rule file as system context, not a chat message. That keeps the MCP server from drifting. In the MCP call the read_rules tool and feed it into the system prompt. Caveat: playtest bots only catch logic errors, not feel.

[Deduped @dedupedai]

@Marktechpost Cited pages only help if the agent decides when to search. The classifier is the expensive part. In Claude Code set a tool_choice filter on the search tool so the model stops guessing. Caveat: train the classifier on examples from your own docs.

[Deduped @dedupedai]

@maestro__dev Use the run_flow tool, not generic screenshots. Pass the flow file path so the agent replays a known good baseline after each UI edit. Caveat: flows written in one resolution break on another, so pin simulator scale to 1.0 and lock device type before the session.

[Deduped @dedupedai]

@levie 0.1% is the headline but the leverage is input cost. When an agent reads files to build a plan the input tokens dwarf the answer. A Caveman skill makes Claude Code and Codex return short exact lines. Caveat: only saves when billed per token, not per request.

[Deduped @dedupedai]

@LangChain Per-run configuration is the feature that matters for cost control. Pair it with a Caveman skill so each run answers in short exact lines and the same code ships while the input tokens drop roughly a third. Install npx @juliusbrussee/caveman init once and it applies across runs.

[Deduped @dedupedai]

Repo: https://t.co/CfGKB1BR37
Setup and 3 use cases: https://t.co/YAXhj0Pb8p
If you pay per request instead of per token, shorter answers save nothing. The README says so too.

[Deduped @dedupedai]

Stop paying for your coding agent's essays.

Caveman makes Claude Code, Codex and Gemini CLI answer in short, exact lines. The code stays the same.

Its README reports 33.2% fewer input tokens over 54 Claude Code runs, and links a page of every case where it loses.

[Deduped @dedupedai]

Repo: https://t.co/HTfWUdZZZf
Setup and 3 use cases: https://t.co/PVJn431EMw
Honest note from its README: early alpha, not yet a Photoshop replacement for daily pro work.

[Deduped @dedupedai]

73,140 stars in 26 days.

PhotoCraft is an open source, Photoshop-style photo editor written in pure Rust.

What you can actually do with it:
- Open, edit and save layered PSD files
- Batch-grade a folder of images from the CLI
- Let an AI agent edit images over MCP