Mildo 🥽

[Mildo 🥽 @Mildo_ai]

@zeuuss_01 WEB

[Mildo 🥽 @Mildo_ai]

@zeuuss_01 Web

[Mildo 🥽 @Mildo_ai]

@VaibhavSisinty Fable

[darkzodchi @zodchiii]

https://t.co/s7ybsbBEED

[Mildo 🥽 @Mildo_ai]

@liu8in Yes please

[Mildo 🥽 @Mildo_ai]

@johnvirality 100

[John @johnvirality]

claude fable 5 just made it possible to post 100 AI UGC videos per day across 4 platforms in JUST 30 minutes of production

everything else runs 100% autonomously

clippers might be cooked 😭

this mythos model watches raw video footage and finds viral moments transcripts would miss

it scrolls tiktok while you sleep and builds trend reports, generating 100 production packages and renders them through higgsfield MCP without you touching another tab.

so i documented the ENTIRE machine with every prompt, every setup instruction, and every workflow step

here's what's inside:

→ the overnight market research system that runs while you sleep

(cowork scrolls 5 platforms, analyzes 60-80 pieces of content each, returns a trend intelligence report with 10 specific content ideas by the time you wake up)

→ raw-pixel clip identification that finds moments human clippers miss

(facial expression shifts, product reveals, body language peaks, visual incongruities. all timestamped and ranked by predicted virality.)

→ batch script and asset generation: 20 complete production packages per prompt, run 5x for 100 total

(6-shot scripts, character prompts, product frame prompts, voice direction, platform captions. 15-25 minutes total.)

→ the higgsfield MCP pipeline that renders all 100 videos automatically

(fable 5 calls seedance 2.0 directly. character reference locked. lip sync aligned. zero human involvement.)

→ the virality predictor that filters your top 20 candidates before posting

(hook score, hold rate, brain region activation. bottom 5 get diagnosed and revised automatically.)

→ the CPM math: 400 platform-posts/day × 3,000 avg views = 36M views/mo = $180k/mo at $5 CPM

→ all 6 copy-paste prompts that run the entire machine

all from my personal experience in looking behind the scenes on how Rizz App + Looksmax AI + Memix scaled past 7-figures with this method

like + comment "100" and i'll send you the ENTIRE system

(must be following + RT for priority access)

will stop sending these out in 24h...

[RasputinKaiser @RasputinKaiser]

My agent usually operates with "Tranche method (often written as T0, T1, T2, etc.), which is a form of time-phased/phase-based planning.
You break work into numbered tranches (or bands), execute one tranche at a time, and keep scope strict per tranche so you can deliver incrementally and re-prioritize between slices."

[Mildo 🥽 @Mildo_ai]

@YVOVY_AI What tools did u use for motion and interactivity. ?

[Mildo 🥽 @Mildo_ai]

@drianlarde_main @LexnLin will it work?

[Mildo 🥽 @Mildo_ai]

@everestchris6 REEL

[Chris @everestchris6]

claude opus 4.8 + OpenClaw now finds restaurants with weak food photos, rebuilds their best dish into a cinematic reel, and mails the owner a postcard with the QR...on autopilot.

here's how agencies can land recurring contracts with this system:

- scans every restaurant in a city in real time
- pulls their real reviews, ratings, and reviewer-uploaded food photos
flags the weakest shot of their signature dish
- samples the brand color straight from the restaurant's own dish photo
rebuilds that exact plate into a cinematic 9:16 reel
- writes a printed postcard about their best dish
- mails it to the registered office, addressed to the owner, with a QR to the live reel

every step from the scrape to the reel to the mailbox is automated

reply "REEL" + RT and i'll send you a free guide so you can build this too (must be following so i can DM you)

[Mildo 🥽 @Mildo_ai]

@everestchris6 reel

[Mildo 🥽 @Mildo_ai]

@aripratama293 Mumbai marine drive

[CopyRebeldia @CopyRebeldia]

Un laboratorio chino acaba de humillar a media industria del video.

Subes una foto y un audio, sale un avatar hablando en sincro. Open source.

Lo que antes era agencia, cámara y edición ahora parece un repo.

Se llama LongCat-Avatar.

[Mildo 🥽 @Mildo_ai]

@Zephyr_hg Sales

[neilson @neilsonks]

open-sourcing a 3D gen toolkit for Claude Code

input image → environment, meshes, physics, lighting, & audio

[Dilum Sanjaya @DilumSanjaya]

Fun interactive science app ideas | Part 3

Played around with generating 3D biological structures and made an app to explore them interactively

UI Design
GPT Images 2

Code
Gemini 3.1 Pro

More demos ↓

[Akshay 🚀 @akshay_pachaar]

How to write Skills that never fail.

Writing good skills for agents is one of the highest-leverage skills you can develop today.

But many Claude skills fail silently. They never trigger, they trigger on the wrong prompts, or they skip the steps they were written to enforce.

The fix starts with understanding how skills actually load.

Skills run on 𝗽𝗿𝗼𝗴𝗿𝗲𝘀𝘀𝗶𝘃𝗲 𝗱𝗶𝘀𝗰𝗹𝗼𝘀𝘂𝗿𝗲, a 3-tier loading system that keeps your context lean.

𝗧𝗶𝗲𝗿 𝟭: 𝗠𝗲𝘁𝗮𝗱𝗮𝘁𝗮 (𝗮𝗹𝘄𝗮𝘆𝘀 𝗹𝗼𝗮𝗱𝗲𝗱)

The YAML frontmatter (name and description) sits in the system prompt at session start. Roughly 100 tokens per skill.

𝗧𝗶𝗲𝗿 𝟮: 𝗦𝗞𝗜𝗟𝗟.𝗺𝗱 𝗯𝗼𝗱𝘆 (𝗹𝗼𝗮𝗱𝗲𝗱 𝗼𝗻 𝘁𝗿𝗶𝗴𝗴𝗲𝗿)

The full instructions enter context only when Claude decides the skill is relevant, based on the description alone.

𝗧𝗶𝗲𝗿 𝟯: 𝗕𝘂𝗻𝗱𝗹𝗲𝗱 𝗿𝗲𝘀𝗼𝘂𝗿𝗰𝗲𝘀 (𝗹𝗼𝗮𝗱𝗲𝗱 𝗼𝗻 𝗱𝗲𝗺𝗮𝗻𝗱)

Scripts run via bash with only stdout entering context. Reference files load when SKILL .md branches to them. Assets never enter context at all.

This is why the description field is doing 90% of the work. If selection fails, the body might as well not exist.

Six practices that compound over time:

𝟭. 𝗪𝗿𝗶𝘁𝗲 𝗱𝗲𝘀𝗰𝗿𝗶𝗽𝘁𝗶𝗼𝗻𝘀 𝘁𝗵𝗮𝘁 𝗮𝗿𝗲 𝗽𝘂𝘀𝗵𝘆, 𝗻𝗼𝘁 𝗽𝗮𝘀𝘀𝗶𝘃𝗲

A 650-trial activation experiment found directive descriptions hit 100% activation. Passive "Use when..." phrasing collapsed to 37% under load.

The template that wins:

"<Domain> expert. ALWAYS invoke this skill when the user asks about <triggers>. Do not <alternative action> directly, use this skill first."

Positive routing plus a negative constraint. Either alone is insufficient because Claude finds the bypass.

𝟮. 𝗔𝗱𝗱 𝗲𝘅𝗽𝗹𝗶𝗰𝗶𝘁 𝗲𝘅𝗰𝗹𝘂𝘀𝗶𝗼𝗻 𝗰𝗹𝗮𝘂𝘀𝗲𝘀

Anthropic's own docx skill ends with: "Do NOT use for PDFs, spreadsheets, Google Docs, or general coding tasks unrelated to document generation."

This is how you stop two skills from fighting over the same vocabulary.

𝟯. 𝗞𝗲𝗲𝗽 𝗦𝗞𝗜𝗟𝗟.𝗺𝗱 𝘂𝗻𝗱𝗲𝗿 𝟮𝟬𝟬 𝗹𝗶𝗻𝗲𝘀

The 500-line cap is a maximum, not a target. Past 200 lines, instructions at the bottom get ignored. Move detail to a references/ folder and link to it.

For pure judgment skills like creative writing or design, aim for 30 to 60 lines. More constraints destroy the output you're trying to produce.

𝟰. 𝗣𝗮𝗶𝗿 𝗲𝘃𝗲𝗿𝘆 𝗴𝗲𝗻𝗲𝗿𝗮𝘁𝗶𝗼𝗻 𝘀𝗸𝗶𝗹𝗹 𝘄𝗶𝘁𝗵 𝗮 𝘃𝗲𝗿𝗶𝗳𝗶𝗲𝗿

Activation is one reliability problem. Execution adherence is the harder, silent one.

The model rationally skips verification steps because they delay output and add no visible content. Ship a paired verifier with binary, mechanical assertions: regex, parsers, file existence checks. Use LLM-as-judge only for genuinely subjective dimensions.

𝟱. 𝗜𝘁𝗲𝗿𝗮𝘁𝗲 𝗼𝗻 𝘁𝗵𝗲 𝗿𝗶𝗴𝗵𝘁 𝗹𝗮𝘆𝗲𝗿

- Skill never fires → Change the description.
- Fires on wrong prompts → Add exclusions to the description.
- Fires but produces wrong format → Change the body.
- Same helper code every run → Promote to scripts/.

Teams often keep tweaking the body when selection is the actual problem.

𝟲. 𝗕𝘂𝗶𝗹𝗱 𝗮 𝗿𝗲𝗴𝗿𝗲𝘀𝘀𝗶𝗼𝗻 𝗳𝗶𝘅𝘁𝘂𝗿𝗲 𝗽𝗲𝗿 𝘀𝗸𝗶𝗹𝗹

Three saved prompts, expected behaviors, re-run on every model update.

Model behavior shifts silently, and without a fixture you discover regressions through user complaints.

The deepest insight: maturation in this craft involves removing something, not adding it. Compress descriptions, sharpen triggers, trust the model with enough latitude to apply judgment correctly.

Skills are Markdown with a tiny bit of YAML and some optional scripts. The leverage is in what you leave out.

If you want to go deeper, I wrote a detailed article covering the anatomy of the .claude/ folder, a complete guide to CLAUDE(.)md, hooks, skills, agents, and permissions, and how to set them all up properly. Link in the next tweet.

[Meta Alchemist @meta_alchemist]

Codex's app has been super slow for me lately.

at first, I thought the problem was Codex itself.

It wasn’t.

After cleaning things up properly, Codex felt roughly 10X faster. 0 slowness. Before this, I had 8GB of logs built up, and it slowed things down like crazy.

Here’s the 15-point cleanup system, which worked perfectly for me. It won't delete anything.

Copy paste these 15 bullet points when your Codex starts to slow down:

> it will inspect things first
> back up & archive important files
> and make your Codex blazing fast again.

15 ITEMS TO KEEP CODEX FAST

1. Check what is actually taking space.

Inspect sessions, archived sessions, worktrees, archived worktrees, logs, config, and the local state database.

2. Back up the important files first.

Back up config, global state, session index, state database, memories, skills, plugins, and automations before changing anything.

3. Check if Codex is open.

If Codex is running, only inspect. Apply cleanup after closing it so the local database is not being touched from two places.

4. Find the giant active chats.

Look for the biggest active session files. These are often old conversations that are still treated as active history.

5. Archive old non-pinned chats.

Move chats older than 7-10 days into archived sessions, unless they are pinned or clearly still current.

6. Keep only recent work active.

Your sidebar/history should not be carrying weeks or months old execution threads.

7. Use handoff docs instead of massive chats.

If an old thread matters, turn it into a handoff doc, archive the thread, and resume in a fresh chat from the doc.

8. Normalize weird paths.

On Windows, clean up path mismatches like normal C:\... paths vs extended \\?\C:\... paths.

9. Prune dead config projects.

Remove project paths from config that no longer exist or point to temporary folders.

10. Move stale worktrees.

Don’t keep old Codex worktrees in the hot worktrees folder. Archive them instead of deleting them.

11. Rotate large logs.

Move oversized old logs into an archive folder so Codex can recreate fresh ones.

12. Check heavy background processes.

Look at Node/dev-server processes. Don’t auto-kill them, but close the ones you don’t need.

13. Verify the cleanup.

Afterward, confirm config still parses, the database opens, active session size dropped, archived sessions increased, and no bad paths remain.

14. Turn this into a weekly script.

The cleanup should not be a dramatic one-time rescue mission. Make it repeatable.

15. Make it boring.

Weekly maintenance should back up first, archive old sessions, normalize paths, prune config, move stale worktrees, rotate logs, and give you a report.

The biggest lesson for me: giant chats should not become permanent memory.

Chats are for execution. Handoff docs are for memory. Archives are for history. Fresh threads are for speed.

P.S. Before doing all this, make comprehensive handoff documents for each active chat, too, with prompts prepared for each to reactivate them after.

This will start new chats from the exact places you left off, but at blazing-fast speed.

Like this, things simply work perfectly.

I even told my Codex to automate these weekly, and it has set it up for every Sunday.

Save this for when you will need it, as Codex app does get heavy as you use it more, especially if you are using many terminals and long sessions a lot.