The Tools I Actually Wired Into My Week

By Muhammad Kashif · Draft · AI automation & cybersecurity builder, Pakistan

Most “AI tools of the week” lists die in my bookmarks. This one stuck because each name mapped to a job I already do: keep context across chats, scrape research without babysitting tabs, generate media for demos, and translate live when a client call jumps languages.

Here is how I am using the stack that showed up together in my notes — ChatGPT Projects, Firecrawl, Krea, HeyGen, ElevenLabs, Mistral’s agent memory, Google Translate’s live mode, and Tencent’s Foley model — without pretending every launch deserves a permanent seat.

ChatGPT Projects: stop re-explaining yourself

ChatGPT Projects went free in that same wave. The useful part is not the folder UI. It is the habit: write the instructions once for a repeated task, pin the files that matter, and reuse the project instead of pasting the same system prompt into a fresh chat every Monday.

I keep one project for client discovery notes, one for security checklist reviews, and one for blog outlines. Anything I do more than twice a month gets a project. If a task is a one-off, it stays in a normal chat so I do not pollute the project’s memory with noise.

Firecrawl: research that becomes structured text

Firecrawl’s AI-readiness check is how I decide whether a site is worth feeding into an agent at all. When I am building brand or competitor notes, I have used it to pull visual and copy references before asking a design model to blend two directions into something new — the same pattern I used when researching references for a health-wearable brand concept.

Practical rule: crawl into markdown, skim for garbage, then hand the cleaned file to the model. Never point an agent at a live marketing site and hope it “just understands.”

Krea and HeyGen: motion and faces without a crew

Krea’s real-time video loop is how I sketch motion before I commit to a longer render. I treat it like a whiteboard that moves: prompt, nudge, discard, keep. It is not my final delivery tool; it is the place I kill bad ideas cheaply.

HeyGen’s Avatar 2.0 path — start from a photo — is what I use when a client needs a talking demo and nobody can sit for a studio day. Later notes also show HeyGen turning a PDF or a site into a video podcast format. That is useful for turning a proposal deck into something a busy founder will actually watch. I still review every line of script myself; lip-sync does not equal accuracy.

ElevenLabs SFX and Tencent Foley: sound that sells the clip

Silent AI video looks unfinished. ElevenLabs SFX v2 covers whooshes, UI clicks, and ambient beds for product demos. Tencent’s Foley model is the parallel option when I need denser environmental sound. One licensing detail from my income notes still matters: on ElevenLabs, commercial use needs the paid plan (the notes flag the low-cost commercial tier). Free-plan clips are for experiments, not client invoices.

Mistral agents with memory, Translate live, and the rest

Mistral’s agent memory is interesting when I want a lighter agent that remembers prior tool calls without dragging an entire ChatGPT Project into the loop. I use it for narrow chores: “watch this folder of API docs and answer only from there.”

Google Translate’s live mode earns a place on calls where Urdu, English, and a third language collide. I keep it on a second device so my primary laptop stays on notes and screen share.

The same week’s notes also mentioned OpenAI buying Statsig and Waymo expanding. Those are industry signals for me, not tools I install. Statsig reminds me product experimentation infrastructure is consolidating into model vendors. Waymo is a reminder that “agents” already drive in the physical world — my browser agents should be held to a higher bar for irreversible actions, not a lower one.

How I decide what stays

Every Friday I ask three questions: Did this tool save a measurable hour? Did it create an artifact I can reuse? Did it touch secrets it should not have? Two yeses and a clean no on secrets, and it stays in the stack. Otherwise it goes back to the watchlist.

That is the whole method. Not thirteen logos on a thumbnail — a shortlist that survives contact with real work.