Salman Bareesh retweeted
A few more crazyyy ways I've been using Instinct lately. These are prettyyy much part of my daily routine now.
5. Monthly expense tracking in Notion
Every month, I download my bank statement and dump it into Instinct.
It pulls all the transactions and adds them to my Notion DB.
The cool thing is, it looks at previous months data to figure out where each expense belongs.
If it's confused about a transaction, it just asks me. I tell it what it was, and it updates Notion.
Ngl, categorizing bank transactions manually is one of those things I never wanna do again.
6. Starring GitHub repos via WhatsApp
I keep seeing so many cool GitHub repos while scrolling X, Reddit, etc.
Could be Claude Code skills, AI agents, random OSS projects... you name it.
And half the time, I forget to star them lol.
Now whenever I find something cool, I just share the repo link with Instinct on WhatsApp.
Since I've connected my GitHub account, it stars the repo automatically.
That's literally it. Don't even need to open GitHub.
7. Dumping random thoughts into Google Tasks
My brain randomly reminds me of stuff I need to work on, usually when I'm in the middle of something else.
So I just drop a quick msg to Instinct.
Something like "Need to test that new Claude Code skill tmrw."
It adds the task to Google Tasks since my account is connected.
Can also give it a deadline or category if needed.
No need to open Google Tasks every single time a random thought pops up.
What are y'all using Instinct for besides research and reminders?
Here's how I use Instinct as my everyday AI assistant:
1. Daily research on autopilot
Every morning at 10 AM, it sends me a research sweep of underrated workflows, hidden gems, and interesting ideas around Claude, GTM hacks, AI agents, and marketing engineering.
If something catches my attention, I ask it to dig deeper. If it's genuinely useful, I ask it to save the findings to Notion so I can revisit or implement them later.
2. AI & tech event discovery
It continuously researches upcoming AI/tech events, filters them against my interests and criteria, and sends me Gmail alerts when something is worth attending.
No more manually checking dozens of event websites.
3. Travel & event planning
Whenever I'm attending an event or planning an outing, I just give it the venue and date.
It researches hotels, cabs, routes, and other logistics, comparing options to find the best value at the lowest possible cost.
4. My second brain for meetings
It has access to my granola meeting notes.
Whenever I forget a discussion, decision, or something someone mentioned weeks ago, I simply ask.
It retrieves the relevant context instead of making me dig through endless notes.
The best part? I don't have to keep prompting it for everything.
It researches, monitors, remembers, and brings things to me when they're relevant.
This is what I actually want from a personal AI assistant. Not another chatbot waiting for instructions.
Salman Bareesh retweeted
I made a 9 sec AI UGC video for Treg for just $1.39 using Treg itself.
Used Darkugc(.)com to find a reference, Treg to generate the AI Avatar, Playwright to record the product, and FFmpeg to edit everything.
Here's the complete workflow, including the tools, settings, costs, and what I learned.
1. Finding a UGC ad to recreate
Started with Darkugc(.)com, a library of best-performing UGC ads.
Found a StressWatch creative with:
→ 552K views
→ 23K likes
The format was simple: a person reacts, and then the video cuts to the product.
Broke the reference into two parts:
→ Hook: A surprising human expression that grabs attention.
→ Payoff: A quick transition to the actual product.
Also studied the timing, transitions, and soundtrack before generating anything.
Another useful reference was Jason Lee's open-source Silent UGC workflow on GitHub.
It follows a similar structure:
→ Generate a human reaction.
→ Use one reference image for consistency.
→ Transition into a real product recording.
→ Add text during editing.
→ Reuse product recordings with different hooks.
The original workflow uses Arcads. For this experiment, I used Treg.
Also took inspiration from Hector Serrano at Wispr Flow and Victor Guo at Composio.
The takeaway: study formats that already get attention, understand their structure, adapt them, and test. Views and likes don't necessarily mean an ad converts.
2. Generating the AI Avatar with Treg
Used Treg GPT Image 2 to generate a photorealistic image of a fictional adult woman.
Cost: $0.05.
The image was designed to resemble a casual phone recording:
→ Natural indoor lighting.
→ A realistic room background.
→ A surprised, hand-over-mouth expression.
→ A selfie-camera perspective.
→ Consistent facial features and positioning.
This became the first-frame reference for Treg Seedance 2.5.
Generation settings:
→ Mode: Image-to-video.
→ Duration: 5 seconds.
→ Resolution: 720p.
→ Cost: $1.335.
The animation prompt requested:
→ Subtle eyebrow and eye movements.
→ Natural blinking.
→ Small hand movements.
→ Slight handheld camera drift.
→ No generated speech or text.
The movement was intentionally minimal. Exaggerated expressions, unnecessary camera movement, and repeated animations can make AI-generated humans look unnatural.
Kept the generated clip at natural speed and checked the face and hands for consistency. The last frame was also used to extend the reaction during editing without looping the AI animation.
3. Recording the real product with Claude Code + Playwright
Used Claude Code to automate the browser recording with Playwright.
The target was Treg's actual AI generation page.
A mobile Chrome viewport was configured with:
→ Viewport: 540 × 960.
→ Device scale: 2×.
→ Capture resolution: 1080 × 1920.
→ Aspect ratio: 9:16.
Playwright captured a smooth scroll through the website.
The recording sequence:
→ Start at Treg's opening section.
→ Scroll toward the pricing example.
→ Reveal the $1.80 pricing card.
→ Show the "one token" heading.
→ Keep the final frame visible long enough to read.
This connected the avatar's reaction to the actual product and its pricing. One useful advantage of this approach is reusability.
The same product recording can be combined with:
→ Different AI avatars.
→ Different facial reactions.
→ Different opening text.
→ Different hooks and pacing.
No need to record the product again for every creative variation.
4. Editing with Claude Code + FFmpeg
Claude Code handled the final assembly through FFmpeg.
The timeline:
→ 0–5s: AI-generated avatar reacting at natural speed.
→ 5–6s: Gentle push-in on the avatar's final frame.
→ 6–9s: Real Treg website recording.
The push-in extended the reaction without visibly repeating the AI-generated motion. Large white text was added during editing rather than generated inside the video.
This kept the typography sharp and readable on mobile and made it easier to change the text for future variations.
The edit was synchronized to the reference soundtrack. Commercial reuse would require the appropriate audio rights.
Final export settings:
→ Resolution: 1080 × 1920.
→ Frame rate: 30 FPS.
→ Format: MP4.
→ Duration: 9 seconds.
The final export took roughly 5 minutes.
5. Cost breakdown and learnings
Here's the actual AI generation spend:
→ Treg GPT Image 2: $0.05.
→ Treg Seedance 2.5: $1.335.
→ Playwright browser recording: $0 additional API charges.
→ FFmpeg editing: $0 additional API charges.
Total: $1.385.
Claude Code usage costs and local compute costs are not included in that figure.
The final result was a 9-second UGC-style creative combining a fictional AI avatar with a real product demonstration.
It hasn't been tested as a paid ad yet, so there are no conversion or ROAS results.
A few useful takeaways:
→ Start with a reference creative instead of a random generation prompt.
→ Break down the hook and product payoff before generating.
→ Use a single image to anchor the avatar's appearance.
→ Keep AI-generated motion subtle.
→ Use Claude Code + Playwright to automate real product recordings.
→ Use Claude Code + FFmpeg to assemble and export the video.
→ Add text during editing rather than generating it with AI.
→ Reuse product recordings to experiment with multiple hooks.
I handled all through Claude Code, while Treg cooked the actual image and video generation.
Next, I'm planning to experiment with different hooks and avatars to see which formats perform better.
Introducing OpenSource Higgsfield by @treg_ai
Try at treg.to/aigc
- 3x cheaper than Higgsfield
- Replicate any viral video for $0.1/s
- Run directly in Claude code or Codex
OpenSource, 0% Markup
Repo link below 👇
Salman Bareesh retweeted
Whenever I built UI and UX for my marketing agents and workflows, Claude always shipped the worst designs.
So I tried something different with our Internal product (Outreach Engine). It looked like AI slop, loaded slowly, and clicking one tab opened another.
If you're building internal products for your own use and want a UI and UX you actually love, I believe this workflow is going to be a huge help.
The biggest lesson up front: good UI comes from good references, not just good prompts.
Here's how I fixed it 🧵👇
1/
Say what's wrong, plainly.
Mine:
> Looked like every other AI SaaS
> Slow to load
> No proper tab navigation
> Click "Actions," land on "Sequence"
Name the pain first. "Make it premium" isn't a brief.
2/
Use the library you've already saved.
I'd been saving UI resources in Notion for months. Claude searched it through the Notion MCP and picked out what fit a real app:
> efferd: dashboard blocks
> EasyUI: animated buttons, toasts
> affectUI: inputs, composer
It also rejected some resources:
> Landing-page template packs → wrong use case
> WebGL effect libraries → beautiful, but could make the app slower
A good resource list is half "use this" and half "not this."
3/
Bring in design skills.
I ran three design skills inside Claude Code:
> Impeccable: honest critique + polish
> Apple Design: calm, spacing, hierarchy
> UI UX Pro Max: design patterns and component guidance
These helped Claude make more consistent design decisions instead of guessing what looks good.
4/
Preview before writing a single line of code.
Claude built a clickable mockup as a Claude Artifact.
I clicked through it, sent screenshots, and gave feedback like:
"Too congested."
"Less color."
"Make it more minimal and luxurious."
It started with 3 screens and grew to 14 boards.
Zero production code touched.
5/
Give it ONE great reference.
Mine was rerun(.)build
Claude inspected the website and pulled its violet gradient from the stylesheet.
We established a simple rule:
> White backgrounds
> Black typography
> Grey for secondary elements
> Violet only where it matters
The design improved significantly once the visual reference was clear.
6/
List every existing feature before redesigning.
I asked Claude to send agents through the old application and inspect:
> Every tab and screen
> Buttons and actions
> Filters and columns
> Pop-ups and modals
> Existing workflows
The new design had to preserve all of them.
A redesign that loses features is just a downgrade with better fonts.
7/
Build it on a stack the components were made for.
I used:
> React
> Tailwind CSS
> shadcn/ui
This made it easier to integrate the component libraries from my Notion collection.
The new UI runs alongside the old one, and nothing gets replaced until every screen is working.
Every destructive action, like Delete or Pause, also requires confirmation.
8/
Polish in small rounds.
Once the UI was functional, I kept reviewing screenshots and giving feedback.
Some changes:
> Smaller sidebar, no scrollbar
> Removed a tab I never used
> Notification bell instead of status cards
> More breathing room, fewer badges
> Company favicons on the dashboard
Each round:
Screenshot → Feedback → Fix → Repeat.
9/
The whole workflow:
> Name the pain
> Use your saved resources (Notion)
> Add design skills
> Preview as an Artifact first
> Pick ONE great reference
> List every existing feature
> Build on the right stack
> Polish in small rounds
My biggest takeaway?
Garbage reference, garbage UI. Great reference, great UI.
I'm curious how others are handling UI/UX when building internal tools with Claude Code.
Are you using design skills, reference websites, or just prompting until something looks good?
Also, if you've found any underrated UI libraries or design skills, drop them below. I'd love to add them to my Notion collection.
Salman Bareesh retweeted
What if your next 10,000 users came from Claude instead of Google?
Just saw a founder on Reddit claiming they got 24,000+ signups in 5 weeks by listing their MCP on Claude.
That's insane.
→ Built the MCP using Claude Code in a week
→ Got listed in Claude's directory
→ Peaked at 1,500 signups/day
→ Still pulling ~250/day
I've been thinking about this for a while.
Instead of just relying on SEO, ads, or outbound, I think getting your APIs, MCP servers, and CLI tools listed and discoverable inside AI assistants could really pay off.
Imagine someone asking Claude or ChatGPT to get something done, and your tool gets recommended right there.
That's distribution without the user even visiting Google.
I think more SaaS founders should seriously look into this.
Curious, has anyone here actually got users through MCP or AI assistant directories?
Salman Bareesh retweeted
AI coding has reached a POINT where apparently everything needs its own engineering discipline.....
First, it was Prompt Engineering.
Then came Context Engineering.
Now I'm seeing Harness Engineering, Loop Engineering, Memory Engineering, and a dozen more!!!!!
I've been following the AI coding and agentic engineering space closely, and it's interesting how quickly the conversation is shifting from writing prompts to designing entire systems.
Here are 12 engineering terms I've been coming across lately:
1. Harness Engineering → Building the tools, environments, and guardrails that help AI agents work reliably.
2. Context Engineering → Giving models the right information at the right time.
3. Loop Engineering → Designing execution and feedback loops so agents can improve their work.
4. Goal Engineering → Defining clear objectives and success criteria for agents.
5. Model Engineering → Selecting and optimizing models for specific tasks.
6. Prompt Engineering → Writing effective instructions to guide model behavior.
7. Graph Engineering → Connecting agents, tools, and tasks into structured workflows.
8. Reducer Engineering → Compressing context without losing important information.
9. Memory Engineering → Helping agents remember and reuse information across sessions.
10. Cache Engineering → Reusing previous results to reduce latency and costs.
11. Eval Engineering → Testing whether agents actually deliver the expected results.
12. Jev Engineering → One I'm still trying to find a clear definition for.
I'm increasingly convinced that the biggest improvements won't always come from switching to a smarter model.
They'll come from engineering better systems around the models we already have.
My question
Are we actually discovering new engineering disciplines, or just giving fancy names to things software engineers have been doing for years???? 😂
Genuinely curious what builders think.
Salman Bareesh retweeted
I think local businesses should still pay attention to Businesses in ChatGPT even with the current limits.
From my point of view, the biggest restriction is clear: you can only claim an existing US listing right now. No new listings and no Business Agent yet.
Those features start with US ecommerce and online services. Still, I believe claiming is worth doing if your listing already exists.
Once claimed and verified, from my experience with similar profile tools, you get real control over the details ChatGPT shows about you. You can update description, hours, categories, photos and policies. You also unlock aggregated insights on how often your location appears in results, website clicks and common conversation topics. A verified blue check helps signal that the real owner manages the information.
Here is what I would do this week if I ran a local business:
1. Search for your business in Business Hub. If the listing exists, claim it right away using phone, email or domain verification, then finish the Persona ID check.
2. Cross-check everything against your Google Business Profile and website. Fix any mismatches in name, address, phone, hours and categories. From my experience, inconsistency hurts AI recommendations more than almost anything else.
3. Strengthen the foundations ChatGPT still relies on. Make your Google Business Profile complete with photos, services, Q&A and reviews. Keep your name, address and phone identical across directories. Write clear website pages that answer real customer questions in plain language.
4. Watch the new insights once they appear. Use the discovery topics to see what people actually ask and fill those gaps on your site or profile.
5. Treat this as early positioning. I believe when OpenAI expands local agents and more markets, verified profiles will already be ahead.
Claiming does not magically push you higher in answers. ChatGPT still pulls heavily from the open web. But from my point of view, it stops outdated or wrong information from spreading and gives you data that unclaimed competitors simply do not have.
Pair the claim with solid local SEO and steady review generation for the strongest near-term results.
Now in beta: Businesses in ChatGPT.
Claim your business, manage your information, and get insights on how people engage with you.
Eligible businesses can also build a Business Agent to answer customers’ questions and help them explore what you offer.
openai.com/index/introducing…
Salman Bareesh retweeted
Here's how I use Instinct as my everyday AI assistant:
1. Daily research on autopilot
Every morning at 10 AM, it sends me a research sweep of underrated workflows, hidden gems, and interesting ideas around Claude, GTM hacks, AI agents, and marketing engineering.
If something catches my attention, I ask it to dig deeper. If it's genuinely useful, I ask it to save the findings to Notion so I can revisit or implement them later.
2. AI & tech event discovery
It continuously researches upcoming AI/tech events, filters them against my interests and criteria, and sends me Gmail alerts when something is worth attending.
No more manually checking dozens of event websites.
3. Travel & event planning
Whenever I'm attending an event or planning an outing, I just give it the venue and date.
It researches hotels, cabs, routes, and other logistics, comparing options to find the best value at the lowest possible cost.
4. My second brain for meetings
It has access to my granola meeting notes.
Whenever I forget a discussion, decision, or something someone mentioned weeks ago, I simply ask.
It retrieves the relevant context instead of making me dig through endless notes.
The best part? I don't have to keep prompting it for everything.
It researches, monitors, remembers, and brings things to me when they're relevant.
This is what I actually want from a personal AI assistant. Not another chatbot waiting for instructions.
Salman Bareesh retweeted
Claude Fable 5.2 is especially strong for ambitious long-horizon agentic work, multi-agent orchestration, coding, deep research, vision and chart analysis, and turning research into finished outputs like docs, spreadsheets, and decks.
It can run for hours, coordinate large agent teams, and scores well on business workflows, agentic search, computer use, and knowledge work while generating results faster and clearer than the previous version.
Here are the practical marketing and growth workflows and agents I would actually build with it:
1. Competitive Intelligence and Monitoring Swarm
- Multi-agent system that continuously scans competitor websites, pricing pages, ad libraries, social profiles, content, reviews, and news.
- Uses vision to analyze creatives, landing pages, and charts or screenshots.
- Synthesizes weekly or monthly reports with strengths, weaknesses, messaging gaps, and recommended counter-moves.
Why Fable fits: long-horizon research, agentic search, multi-agent coordination, and vision. It can run largely on its own and produce polished strategy decks or spreadsheets.
2. End-to-End Campaign and Funnel Builder
- Starts from a brief or product description and goes through audience research, messaging angles, landing page code plus copy, ad variants, email and nurture sequences, tracking setup, and a basic performance analysis loop.
- Can generate or iterate on creative briefs and use vision to critique or improve assets.
Why Fable fits: strong coding for pages, scripts, and automations, multi-step planning, and the ability to own a project end-to-end or orchestrate sub-agents. Similar models have already been used to build complete marketing funnels.
3. Content Engine and SEO Research-to-Publish Pipeline
- Topic and keyword research, then outline, long-form drafts, optimization for SEO, readability, and brand voice, supporting assets like carousels, social posts, and image prompts, scheduling suggestions, and a performance feedback loop.
- Can run as multi-agent with separate agents for research, writing, SEO and optimization, and distribution planning.
Why Fable fits: deep research synthesis, clear writing, coding for content tools or CMS integrations, and consistent long-context handling of brand guidelines plus past performance data.
4. Growth Experimentation and Optimization Orchestrator
- Designs A/B or multivariate tests for copy, pricing, UX, or offers, implements them via code or tool calls, monitors metrics, runs statistical analysis, surfaces insights, and maintains a prioritized experiment backlog.
- Can analyze funnel drop-offs from data exports or dashboards.
Why Fable fits: strong reasoning, coding for instrumentation and analysis scripts, data handling into finished reports, and the ability to iterate over long sessions.
5. Creative Analysis and Iteration Agent
- Ingests ad creatives, landing pages, social posts, or competitor visuals.
- Uses vision and chart recognition to identify patterns in what works across hooks, visuals, layouts, and CTAs.
- Generates improved variants, critiques, or style guides.
Why Fable fits: improved vision and visual chart recognition, plus coding to generate or process assets and clear communication of recommendations.
6. Personalized Outreach and Sales Enablement Agent
- Deep prospect research on company, role, and recent signals, then hyper-personalized email or LinkedIn sequences, CRM enrichment, and follow-up logic.
- Can maintain brand voice and compliance rules across large volumes.
Why Fable fits: agentic search, research depth, coding for CRM and tool integrations, and the ability to process large context including past interactions and brand docs.
Tools and computer use: Pair it with web search, code execution, browser or computer-use capabilities, CRM, email, or ads APIs, and file systems. It scores very high on computer use and business workflow benchmarks.
Salman Bareesh retweeted
Tired of rebuilding workflows every time a new agent appears.
Over the past two years I have used n8n, OpenClaw, Hermes, Instinct, Muse and Grok Bot. Each new tool brings the same cost.
You learn the interface, migrate your workflows, refine them, then fix bugs or adapt to changes. After a while the constant switching starts to feel like wasted effort.
I need a tool that is both API-heavy and code-heavy.
Instinct, Muse and Grok Bot handle the API side well. They offer single sign-on connectors, solid context and memory. They make it easy to orchestrate simple recurring jobs such as daily email digests, reply monitoring or competitor tracking. You connect the apps they support and the agent handles the coordination.
These tools work for light, repeatable tasks. They hit a clear limit though. You cannot easily turn them into advanced codebases or custom repositories that execute complex real-world workflows outside their provided environment. They stay inside the boundaries of the connectors and features the company ships.
Hermes sits on the other side of that line. It is built so you can create and maintain real repositories and codebases that actually run sophisticated workflows.
From the beginning the team has shipped frequent updates. These include new tools, better sub-agents, computer-use improvements, and native bot concepts after Grok Bot trended. The pace of development means the platform keeps growing rather than forcing you to migrate elsewhere.
It is also model-neutral. You can run ,
GPT, Claude, local models or others without losing the memory, skills or identity layer you have built. Caching is strong. Context persists. The agent remembers how you work across sessions. The layer above the model belongs to you. This includes the memory, the skills and the customisation. If something is missing you can change it, extend it or open a pull request.
That ownership is what makes it different from tools where the company controls the behaviour and the limits.
You can still reach it the convenient ways. These include the desktop app, the Claude Code button, messaging apps or the official beta mobile. The real value appears when you treat it as a code-first environment.
It handles the simple orchestration other agents do. It also supports the deeper work of building and running custom code that solves harder problems.
If you are like me and keep moving between tools hoping the next one will finally stick, it is worth asking whether you want another polished connector layer or a place where the agent and your own code can grow together without another full migration.
For me that place is Hermes. It is the only one that does both the API convenience and the code depth in a way that actually gets advanced work done.
I have been using Hermes Agent every day since day one, and I think it is finally time for a full review.
As always, this is completely independent. Not paid, not sponsored. The good and the bad, nothing held back.
This review covers what I use Hermes Agent for, why you should consider it, and how it stacks up against OpenClaw, Claude Code, Codex, and Grok Bot.
Hermes has become essential for me. Let me show you why, starting with the first thing most of you asked.
Which models do I use inside Hermes?
The best thing about Hermes is that it is truly model-neutral. It does not degrade the models. Whatever a model can do, you get directly.
I use GLM 5.3 a lot, and for light and medium tasks it is excellent. Sometimes I switch to GPT Sol 6.1, and I have noticed that the best place to run GPT is Hermes, even better than Codex. I have tried a lot of models in Hermes, and they were all excellent.
So use any model that suits you, as long as the plan is generous, and run two or three at once if you want. Here is the important part: in ninety percent of use cases, your choice of model will not make a big difference. The model here is still just a brain behind bigger layers, and those layers are what control how it behaves.
Usage limits and cost
There is no Hermes quota. Plug in your key, run a local model, or subscribe through Nous Portal. That is all it takes.
You can push Hermes to the max with excellent models without paying a single dollar. And if you decide to pay, you get great value for it.
Caching is great too
Out of 1,013 sessions, here's one that came out right at the average:
219 million input tokens
15.7 million output tokens
3.77 billion cached tokens
About 4 billion tokens in total
94.4% from cache.
Hermes is one of the best I have used for token efficiency, and the transparency is total. While you work, you see the real speed of every provider and every detail.
Light on the machine
Hermes is not heavy at all. One day I set up a full Hermes agent inside a virtual machine with just 2 GB of RAM, and it ran with very high stability on every kind of task I gave it. That shocked me. Codex, for comparison, eats tens of gigabytes of RAM. That difference tells you what an engineering masterpiece Hermes Agent is.
The layer above the model
One of the biggest things that sets Hermes apart is that the agent is separate from its brain. The model is the brain, and you can swap it without losing your memory, your skills or your setup.
Anywhere else, when I use GLM 5.3 I get GLM's usual strengths and weaknesses, and when I use Sol 6.1 I get Sol's. In Hermes there is a layer above the model that gives it a soul, a style and an understanding of how I work, so switching models matters far less. At some point the models feel eighty percent alike in style, in method and in the things you care about. What is left is raw capability, so the most capable models go to the hardest problems.
You own the layer you actually invest your time in. And the beautiful part is that this layer, the identity, the memory, the skills and every tool, is completely mine.
Look at Claude Code getting people banned today, at model performance suddenly dropping, at Codex being stingy with usage one time and slow the next. Those are constraints, and Hermes does not have them. If Claude Code decides to change or remove something, you have no say. In Hermes you can always change whatever you want. That is sovereignty, freedom and empowerment.
Memory that grows with you
When I opened the skills and memory tree, I found that Hermes had reached the point where it knows me better than any other agent. It grew with me through this whole journey, and that is what makes it the ideal partner for my work.
The journey command shows you a star map of your memory and skills, and you can edit or delete anything in it. It is one of the greatest things in Hermes.
One example: when Fable came out I loved it, so I had Fable pass its style and way of thinking on to Hermes. My Hermes became twice as useful after that.
And the instructions stick. Across sessions and updates, the soul, the identity and the instructions are always there. I never feel I need to remind it, and I never notice it forgetting or drifting. It is with me one hundred percent, every time.
Research and academic writing
Hermes is excellent for research and academic writing. With Claude Code or Codex I always have to explain my method, my way of thinking and everything I do not want. In Hermes I explained myself once, and from then on research and writing just happen the way I want. You set it up once, and it keeps growing. It even picked up my taste in design and research.
Skills and tools
Hermes has skills and tools you will not find anywhere else, and the team never stops building. The skills library is one of the best I have seen.
I see about 165 skills built in, plus optional skills and plugins you can add with one click. All useful stuff for growing in any field: iOS apps, computer use, automation, shopping, coding, organization, security, travel, testing, data, streaming, support, sports, reviews, video and more.
The tools cover web search, the browser, the terminal, files, vision, image generation, subagents, scheduled tasks, memory and computer use, and MCP brings in anything else. You can use it with Blender or whatever you want. Use it straight out of the box, or customize it with no ceiling.
Where I talk to it
You can talk to it anywhere: iMessage, Telegram, Slack, Discord and more, and you can customize all of it. When I am out of the house, Hermes is the first thing I think of. I send it a message and everything happens like magic.
My own voice model
Companies that take speech-to-text out of your credits always annoyed me. In Hermes I brought my own model. Everything runs locally on my machine, smooth and fast, exactly the way I want. I even send it voice notes on Telegram, and it just works.
The desktop app
The desktop app is great, and the user experience is one of the best I have seen. I do not like working in the terminal in general, so in Hermes the desktop app is my home.
What I use it for
The idea that Hermes is only for casual users or simple tasks is just not true. It works for every kind of use and then some. I use it for everything I want, and I get great results.
It is the agent I rely on for personal tasks, light work and organization. The test videos and model comparisons I publish are edited entirely by Hermes; I do not touch them. It has helped me study and work, proofread academic research, translate books and documents, run a security audit on my ERP application, and a lot more.
Hermes vs OpenClaw
The most unfair comparison people make is Hermes vs OpenClaw. OpenClaw was a revolution in its time and started a huge wave, and it deserves credit for that. But it is nowhere near Hermes. Not in stability, not in reliability, not in consistency, not in capability. To me, OpenClaw is a tool some people enjoy for two or three days of experimenting. Is it something you can rely on and build projects, ideas and solutions on? Not at all.
Hermes is another story. It even beats tools like Claude Code and Codex. For me it sits in a different category entirely.
Hermes vs Claude Code and Codex
The reply I get most is: you can do all of this in Claude Code or Codex. Yes, you can. But in Hermes I do it the way I want, not the way the model or the company wants, and I enjoy the work instead of fighting the agent.
Hermes also refuses less than any agent I use, and that matters a lot to me. It is my first stop for anything that usually gets refused in Claude Code, Codex and the rest. A while ago I gave Claude Code and Codex a task and both refused. Hermes did it.
Hermes vs Grok Bot
Grok Bot has two real advantages. It works out of the box and it is easy, and it has direct access to X. Beyond that, Hermes crushes it.
Everyone is talking about how fast Grok Bot's quotas run out, and I can confirm it. I am on the Ultra plan, the $200 one, and it does not last me more than a day, two, maybe three. That is not sustainable. To be fair, usage limits are a problem with a lot of providers right now, not only Grok Bot.
The real difference is intelligence. Grok Bot is extremely low on intelligence right now. I ask it to execute a task and it just answers. I ask it a question and it starts executing. I cannot trust it. Its mistakes outnumber its wins, and it keeps telling me it cannot do this, it cannot reach that, this is a limit.
In Hermes, you shape the intelligence layer. I shaped mine to do nothing until it is sure what I want and has reviewed every detail and the full context. Grok Bot has nothing like that, and you cannot shape it. Believe me, I tried.
Do not confuse ease with quality. Grok Bot has a wonderful ecosystem, great templates and amazing growth. But you do not own the environment, you do not choose the kind of intelligence you get, and you struggle to make it through the week.
The problems
Hermes is not perfect, and you should know where it falls short.
Sometimes when I change the model in a session, the interface does not show it. The change happens; the UI just does not read it. Open a new session, come back, and you will see it was applied.
Steering is hit or miss. Sometimes a steering message redirects the conversation right away. But most of the time the agent ignores it completely and waits until the whole task is done, or it spends a really long time on the task before the steering finally kicks in. I always wish steering was instant so no time gets wasted.
Subagents are excellent, but sometimes Hermes hands me a result and then waits for the subagent's report, instead of waiting first and giving me one final result.
Computer use is one of the most important features for me, and it sometimes breaks and needs a real fix. It happened enough that I stopped using it heavily. It usually breaks under heavy work or with medium and weaker models, so give it your most capable models.
That brings me to the most dangerous problem. Weaker models can wreck your Hermes in minutes. You build your setup and the agent's understanding of you with a serious model, then a weak model starts making bad decisions and damages that whole layer. It can even break the app. MiniMax M3 broke my Hermes more than once with its hallucinations.
The fix is simple: stick to the leading, stable models, such as Kimi, Sol, GLM 5.3, DeepSeek Flash, and frontier Qwen models like Qwen 3.8 Max, and stay away from weak, hallucinating ones. Since I stopped using MiniMax M3, I have not hit this once. I do not know if it has been fixed, but I am not risking my stability to find out. What I really want is an optional protection layer for the memory, the soul file and AGENTS.md, so any edit shows up in the interface for me to approve first.
Long tasks can occasionally collapse in goal mode. The agent marks the task done before it is, or gets lost and loops. That happened a long time ago and has not happened since. It is rare and usually comes with weaker models, but I am mentioning it for transparency.
One for Mac users: more than once, an update left my desktop app stuck on an older version. The fix is simple. Quit the app and run hermes update in the terminal. If the app still shows the old version, run hermes desktop to rebuild it.
What I want next
Two things matter most to me.
First, a solid, easy mobile app with beautiful interfaces and that Nous spirit.
Second, an easy mode that works out of the box: add profiles by field, so a researcher, a programmer, or a writer starts from a ready base instead of a raw zero.
A model from Nous Research that understands Hermes better than any other model would be amazing too, but that is not central for me right now.
And I want Hermes to switch modes from natural language. When I give it a goal, it should turn on goal mode by itself and finish the whole task, the way Codex and sometimes Claude Code do.
The community and the team
One moment made me appreciate Hermes more than ever. I wanted Arabic support in the harnesses I use, so I reached out to MiniMax, Zai and others and offered to translate the harness, the website, everything, myself. They all said yes. Then nothing happened.
With Hermes, I did the Arabic localization myself and opened a pull request. They accepted it, and today Hermes speaks Arabic. Your voice as a user is what matters, and that is exactly why I love Hermes.
Another thing that amazed me: one day I sent the team an idea right here on X. I wished I could set the model for each scheduled task. In less than three hours it was built and shipped, and I have been using it ever since. I have never seen that kind of speed. That is what accelerating growth looks like when a project is open source and the whole community builds it.
Where I stand
I have tried most of the well-known tools in the space, and every tool has beautiful positives and annoying negatives. Hermes might be the one with the fewest negatives. Not because it is better or worse than the others, but because in Hermes you can deal with a negative, change it, erase it, and sometimes turn it into a strength. It all comes down to how you customize it.
It is not the easiest tool to start with, so my next post is my full setup, as a skill. Hand it to your agent, answer its questions, and it sets everything up for you. I do not claim it is the best setup, only the one that fits me. Stay tuned!
After all this time, I know exactly where I stand. Hermes is a victory for what open source is supposed to mean.
Salman Bareesh retweeted
The external memory system is great. The capture step kills it.
Most people will never copy a YouTube transcript, Instagram caption, X post, or article and paste it into a folder. It’s too much friction. So the whole system stays empty.
Here’s the version that actually gets used by lazy humans:
Capture = one-click bookmark everywhere.
- YouTube → Save to playlist or library
- Instagram → Save
- X → Bookmark
- Web → Chrome bookmark (or any browser)
- Anywhere else → same idea
Two seconds. No copying. No pasting. No thinking.
Then one daily automation does the rest.
Every night (or morning), a scheduled workflow:
1. Pulls all new bookmarks from your connected accounts (YouTube, Instagram, X, browser, etc.)
2. Extracts the actual content (transcript, full post text, article body)
3. Drops the raw material straight into your memory `inbox/`
4. Runs the `/remember` process so Claude turns it into proper memory cards with “why it matters”, links, and sources
You never touch the folder during the day. You just bookmark whatever feels worth keeping while you scroll. The system catches up automatically.
Result:
Capture friction drops to almost zero.
Your memory still ends up as clean, linked, searchable markdown files you own.
The resurfacing and recall parts stay exactly as powerful as the original system.
The only real requirement is a daily (or few-times-a-week) automation that can read your bookmarks and write into the memory folder. Everything after that is already solved.
Most people fail at second brains because the capture step is annoying. Make capture a single bookmark and the rest can run while you sleep.
Salman Bareesh retweeted
If you're making $1k+ a month from something you built this year, reply.
I want to learn how you got the first 100 users.
Salman Bareesh retweeted
The emails of people who use Mailcheap are landing in spam.
@marclou @denk_tweets
Salman Bareesh retweeted
Building a product is no longer the moat.
Marketing is.
Everyone is shipping. Almost no one is getting distribution.
Here are 4 practical marketing “agents” (workflows) you can build and run right now for any product or service:
1/ Outbound Agent
Use the Apify API to scrape people actively talking about your problem, product category, or competitors on LinkedIn, X, and Reddit.
→ DM them directly
→ Enrich their emails
→ Run targeted cold email sequences
Warm conversations > spray and pray.
2/ SEO / GEO Agent
- Build comparison pages (“X vs Y”)
- Create alternative pages (“Best alternatives to Z”)
- Add a clean llms.txt file
- Get listed on high-DR directories
Make it stupidly easy for both Google and LLMs to recommend you.
3/ Ads Agent
Scrape your competitors’ ads libraries.
Study the winning ones:
• Hooks
• Formats
• CTAs
• Angles
Then replicate and iterate faster than they can.
4/ Content Agent
Study how the algorithm actually works.
Pick your top 10 favourite creators in the space.
Reverse-engineer their patterns + what’s currently trending.
Then ship content immediately while the wave is still moving.
The people who win in the next 2–3 years won’t be the ones who build the most features.
They’ll be the ones who master distribution.
Save this. Build these agents.
Salman Bareesh retweeted
here's what i've learned about keeping emails out of spam.
i've been working in cold email for around 2 years now. although i haven't worked on email deliverability day in and day out, i've picked up a few things along the way.
after seeing the discussion around my previous post, i got curious and started digging a little deeper.
i'm not here to tell you which email provider to use or avoid. i just wanted to share what i've learned about reducing the chances of emails landing in spam.
and this isn't just about cold outreach. it applies to cold emails, newsletters, receipts, and transactional messages too.
here's what i've learned so far
Reputation and Sending Infrastructure
> ip reputation matters. other senders using the same shared ip can affect your email delivery.
> domain reputation matters too. changing your ip or email provider does not automatically repair your domain's reputation.
> separate cold emails, newsletters, and transactional emails into different sending streams. ideally, use separate subdomains and ip pools where practical.
> a dedicated ip isn't always better. if your volume is low, a reputable shared ip pool might perform better.
> check whether your sending ip and linked domains appear on major blocklists like spamhaus.
> check the reputation of domains inside your emails, not just your sending domain. one problematic link can create delivery issues.
> use consistent sender identities. frequently changing domains and sender addresses won't automatically improve deliverability.
Email Authentication and DNS
> configure spf, dkim, and dmarc. these help receiving servers verify that your emails are authorized.
> spf and dkim should align with your visible from domain. simply adding dns records isn't enough.
> use only one spf record per domain. multiple separate spf records can cause authentication failures.
> keep spf within its 10-dns-lookup limit. too many nested services can break validation.
> use 2048-bit dkim keys where supported. check that your actual emails are properly signed.
> start dmarc with p=none to monitor authentication, then move toward stricter policies after verifying legitimate sending services.
> check authentication using gmail's "show original" feature. dns checkers alone don't prove that real messages pass.
> if you're managing your own sending server, configure valid reverse dns, forward dns, and tls.
Sending Volume and Warm-up
> gradually increase sending volume. sudden jumps can trigger throttling or reputation problems.
> there is no universal safe daily email limit. 20–40 emails per mailbox is a common cold outreach rule of thumb, not a guarantee.
> avoid sending your entire daily volume in one burst. spread bulk sending consistently throughout the day.
> if you're restarting an old domain after months of inactivity, don't immediately return to your previous sending volume.
> warm-up tools that simulate opens, clicks, and replies aren't necessarily helping. spamhaus explicitly warns against artificial engagement.
> watch temporary smtp errors. if receiving servers start deferring emails, reduce volume and investigate.
> adding more domains or mailboxes doesn't solve poor targeting, high complaints, or bad sending practices.
Email Lists and Bounce Rates
> verify email addresses before campaigns, especially when working with older lists.
> a verified email only means the address appears deliverable. it doesn't guarantee inbox placement or permission.
> aim for very low hard bounce rates. investigate any unusual increase instead of treating 2% as an acceptable target.
> don't assume catch-all or unknown verification results are valid. treat them separately from confirmed addresses.
> remove permanently invalid addresses immediately. repeatedly sending to nonexistent mailboxes damages list quality.
> don't repeatedly email people who have already unsubscribed, complained, or asked you to stop.
> maintain a shared suppression list across your sending tools. changing providers shouldn't reactivate unsubscribed contacts.
> be careful with old databases. addresses that worked last year might now belong to someone else or no longer exist.
Cold Emails
> personalization matters more than inserting someone's first name. mention something genuinely relevant to their business. go templatised with spintax if volume is high.
> the perfect offer going to stand out from rest of the contents.
> keep your initial email focused on one problem or reason to reach out. unnecessary links and attachments add friction.
> try sending the first email without tracking links when tracking isn't necessary. fewer third-party domains means fewer dependencies.
> avoid fake re: and fwd: subject lines. pretending there's an existing conversation is misleading.
> don't confuse high sending volume with good outreach. more emails to irrelevant people can increase complaints.
> give prospects a straightforward way to decline future outreach, and honor those requests.
> don't automatically subscribe cold prospects to your newsletter.
> check local marketing laws. the rules for contacting businesses differ between countries.
Transactional Emails
> separate transactional emails from marketing campaigns. password resets and order confirmations shouldn't depend on the reputation of promotional traffic.
> avoid inserting promotional content into password-reset emails or important account notifications.
> protect automated email forms from abuse. bots can repeatedly trigger signup confirmations, invitations, and password resets.
> prevent duplicate sends. one user action shouldn't accidentally trigger multiple identical emails.
> monitor how long login codes and password resets take to arrive. arriving after the code expires is still a delivery failure.
> check bounce and suppression events. an email api returning success doesn't necessarily mean the recipient received the message.
> use clear and consistent from addresses for receipts, account alerts, and other transactional messages.
Newsletters and Marketing Emails
> use confirmed opt-in when possible. people who actually want your newsletter are less likely to complain.
> avoid purchased or scraped newsletter lists. a valid email address doesn't mean someone agreed to receive marketing.
> don't suddenly increase newsletter frequency. people expecting one email weekly may complain if they start receiving one every day.
> segment subscribers by interests and engagement instead of sending every campaign to everyone.
> regularly review inactive subscribers. consider re-engagement before removing them from regular campaigns.
> make unsubscribing easy. forcing someone through several steps can encourage spam complaints.
> bulk marketing emails need proper one-click unsubscribe headers, not just a link in the footer.
> evaluate engagement beyond opens. apple mail privacy protection can inflate open counts.
Email Content, Links and Tracking
> plain-text emails aren't automatically safer than html emails. both can land in spam if reputation or recipient feedback is poor.
> there isn't a magic text-to-image ratio that guarantees inbox delivery. avoid image-only emails, but don't obsess over an exact percentage.
> don't rely on removing individual spam words. using or avoiding words like "free" won't fix poor reputation.
> check your email links before sending. your tracking domain, calendar link, and signature links can all matter.
> avoid unnecessary public url shorteners. use recognizable, trustworthy links instead.
> make emails readable when images are blocked. important information shouldn't exist only inside an image.
> keep html clean and reasonably lightweight. oversized emails can get clipped in gmail, although clipping itself isn't spam placement.
Monitoring and Deliverability
> use google postmaster tools to monitor gmail spam complaints, domain reputation, and authentication performance.
> keep gmail's reported spam complaint rate below 0.1% where possible. avoid reaching 0.3% or higher.
> yahoo also expects bulk senders to keep spam complaint rates below 0.3%.
> a 0% complaint rate doesn't automatically mean you're safe. emails already landing in spam might generate fewer inbox complaints.
> delivered doesn't mean inboxed. your email could be accepted by the receiving server and still land in spam.
> don't blindly trust open rates. apple mail privacy protection and automated systems can distort engagement metrics.
> some clicks come from security scanners. compare them with real replies and conversions.
> gmail's promotions tab isn't the spam folder. don't treat every promotions placement as a deliverability failure.
> test emails across gmail, outlook, and yahoo. performance at one provider doesn't guarantee the same results elsewhere.
> if deliverability suddenly drops, compare changes in bounce rates, complaint rates, sending volume, and campaign audiences.
> don't assume switching providers fixes everything. identify what's actually causing the problem first.
> reputation recovery takes time. fixing technical settings today doesn't mean your emails will immediately return to the inbox tomorrow.
that's everything i've learned and collected so far.
some of these might be wrong, and i'm sure there are things i'm missing. what works for one sender might not work for another sometimes.
obviously, none of these hacks can guarantee 100% inbox placement, but even reducing the spam percentage a little would be a win.
i'm not here to defend or criticize any email provider. i'm genuinely curious about what others have experienced.
what are you guys doing to keep your emails out of spam? any underrated hacks i've missed?
instead of fighting over which provider is better, let's share what works, what doesn't, and learn from each other.
would love to hear from people who've been doing this longer than me.
The emails of people who use Mailcheap are landing in spam.
@marclou @denk_tweets












