๐ฅ Technical Content Lead @AccuKnox // Head of DevRel @KubeArmor Building 10xdevrel.atharvashah.com Wake up and chant slay3r's testament everyday.
work, portfolio, blogs โ
- Tweets 2,172
- Following 1,785
- Followers 227
- Likes 11,009
i still see 14 features on the first slide of a board deck.
pendo studied 615 accounts and found 80% of features are rarely or never used. the standish group put it at 45% never, 19% rarely.
most of what that slide reports is dead weight, and the slide calls it achievement.
โ ๏ธ report the behavior that changed instead. name the metric and the size of the move.
21
athrv.sh retweeted
๐ How the top 1% of cold emailers get their replies, explained:
โ
โ ๐ 1. The list
โ โ Build it before you write a word
โ โ Keep campaigns small: 21 to 50 people
โ โ Time it to a change at their company
โ
โ ๐ 2. The offer
โ โ Fix the offer before you polish the copy
โ โ Test a different one, like a free pilot
โ โ Give something useful before you ask for a meeting
โ
โ ๐ 3. The email
โ โ Write the first 25 by hand
โ โ Template the angle that got replies
โ โ Keep email 1 under 80 words
โ โ Ask 1 question
โ
โ ๐ 4. The follow-ups
โ โ Bring a new angle or new proof each time
โ โ Instantly found 58% of replies come from email 1
โ
โ ๐ 5. The inbox
โ โ Verify every address, bounces under 2%
โ โ Spam complaints under 0.1%
โ โ About 20 emails per inbox a day
โ
โ ๐ 6. The scoreboard
โ Turn off open tracking
โ Apple Mail opens emails by itself
โ Count positive replies
This is how you get to 3x the average reply rate.
6
3
30
1,449
my brand colour failed the accessibility test.
3.58 to 1 against white. the bar for body text is 4.5.
so the colour the whole identity is built on was too weak for every small label it touched.
i kept the orange, because the brand is the colour. i split it into two tokens instead. the original handles fills, progress bars and display text at 24 pixels and up, where the bar drops to 3 to 1. a darker one at 5.06 handles any text on a light surface.
the identity did not move. one of the two values now passes at any size.
19
tomorrow i'm on the neatHack winners stream, live on neatlogs' X, YouTube and Discord.
what the stream covers:
1. winners share their demos on stage
2. strategy and tips for winning the hackathon
3. a walk through the top 20 projects
4. what the hosts liked, and how some projects almost made it
i'll be talking about how i built in public to win the community prize: the posts, the feed block, the checks before send.
come hang out. thank you @neatlogs @EntireHQ @arithecfo, and to the judges, mentors and hosts: @bettersayAJ @akshay_hangloo @agent_wrapper @HaimantikaM @chetan_raj @roerohan @ankushdharkar @karan_pargal @BngRithvik @simranrambles
#neatHack #neatlogs
ALT Portrait of Atharva Shah beside the words This Friday, 23 Oct, live on X, YouTube and Discord
1
5
278
how i kept my feed on one topic for 3 days of neatHack:
1. moved 338 scheduled posts out of the way. 226 on X, 111 on LinkedIn, 1 on Instagram
2. each platform slid as one block, so the order and the gaps stayed the same
3. deleted nothing. every post came back after the event
4. put an image on every hackathon post
5. added a check before send column to the queue: commit the image first, be ready to run what people send, every commit carries an Entire checkpoint trailer
6. opened the issue before posting the screenshot
the feed read like one story instead of a random mix.
everyone whose tools and advice this ran on: @neatlogs @EntireHQ @arithecfo, plus @bettersayAJ @akshay_hangloo @agent_wrapper @HaimantikaM @chetan_raj @roerohan @ankushdharkar @karan_pargal @BngRithvik and @simranrambles. thank you.
#neatHack #neatlogs
ALT Grid of 338 dots, one per scheduled post: 226 black for X, 111 blue for LinkedIn, 1 pink for Instagram
5
237
neatHack by the numbers, straight from the winners call:
1. about 1,600 people registered, from 40+ countries
2. 205 submitted. roughly 1 in 8 of the people who signed up shipped something
3. scores ran from 29 to 90.5 out of 100
4. 1st place and 20th place were 7 points apart
5. the top 3 were 1 point apart
6. the community prize ran in its own lane, judged on who built in public. i won that one
so shipping beat about 7 in 8 of the field before the scoring even started. then it got tight.
thank you @neatlogs, @EntireHQ and @arithecfo for running this in the open, and to the judges and mentors: @bettersayAJ @akshay_hangloo @agent_wrapper @HaimantikaM @chetan_raj @roerohan @ankushdharkar @karan_pargal @BngRithvik and @simranrambles for hosting it so well.
#neatHack #neatlogs
ALT Flow chart from about 1,600 registered to 205 submitted, then a grey lane to the top 20 and top 3 and a gradient lane to one community prize winner
8
188
won the community prize at neatHack. here's what went into it:
1. entered solo with FirstRun, an agent that runs a dev tool's quickstart in a clean container and opens a PR when a step breaks
2. posted proof at every step. screenshots, real output, an issue filed before the screenshot went up
3. ignored my follower count. the rules say it doesn't decide the prize and the host repeated it on the call
4. moved 338 scheduled posts out of the way so my feed carried one topic for 3 days
5. put an image on every single post
the tools that carried the build:
@neatlogs traced every run and caught my own agent failing steps that actually worked. @EntireHQ recorded the session behind every commit. @arithecfo showed that my old spending cap would run out of cash in month 14.
thank you to the judges and mentors who shaped this: @bettersayAJ @akshay_hangloo @agent_wrapper @HaimantikaM @chetan_raj @roerohan @ankushdharkar @karan_pargal @BngRithvik and @simranrambles for being the best host!
#neatHack #neatlogs
ALT Confetti around a portrait of Atharva Shah with the words Community Winner and the neatlogs, Entire and cfo.ai logos
8
18
468
>model scaffolds my landing page
>writes eight testimonials: stripe, vercel, mongodb, twilio, cloudflare, supabase, linear, figma
>plus a logo wall of the same eight
>none of it ever happened
>ship past it for weeks
>mfw
the three arrays now export empty, and each section renders only when its array has entries. the block cannot come back by accident. it comes back when a real person sends a real quote.
if a quote or a number on your site cannot be traced to a file, it is decoration.
ALT Headline '8 fake testimonials' over a tilted git diff of eight deleted quote entries for Stripe, Vercel, MongoDB, Twilio, Cloudflare, Supabase, Linear and Figma, under a red DELETED stamp.
2
2
78
more from the 10x devrel series: 10xdevrel.atharvashah.com?utโฆ
13
never trust an ai skill until a script can refuse its output. in one session, my scorer refused 21 linkedin comments, and none of them posted.
a skill is a folder of instructions claude loads for one repeated job. a rule inside it is a request. late in a long session, the task in front of claude wins over a rule it read an hour ago.
three things turn the request into a gate:
- a hook, which is a script claude code runs on its own after every file save. it runs the scorer and blocks the save on a fail
- a gate that fails open. when the scorer crashes, let the work through and print "check skipped". a gate that blocks on its own crash gets switched off on a deadline day
- proof printed by a script. "all checks pass" costs the model nothing to type. a result printed by the check itself cannot claim a run that never happened
ALT A hook runs the scorer after every file write. A pass lets the work continue. A fail blocks the save with the findings. A crash lets the work through and prints check skipped.
1
33
saved a colleague's emails to a folder for months because he writes better than i do.
pointed an agent at the folder. the draft sounded like him.
no amount of adjectives in a prompt had ever got me there.
ALT Card: instructions describe, examples demonstrate, over a skill folder tree and three real deck screenshots
1
15
every tool you use keeps a copy of your work on its own terms.
- a routine, a prompt that runs on a schedule, saved inside a claude account vanishes when you switch accounts
- one scheduling tool i use deletes an uploaded image after 7 days unless a post uses it
- a chat app keeps your history until you clear it or lose the account
a plan change, a ban or a cleanup job removes the only copy. you find out on the day you need the file.
the fix:
- keep the master copy in a git repo you own
- commit the file before you hand it to any tool
- treat the tool's copy as a cache, a spare you can lose without harm
then a deleted upload costs you one re-upload.
ALT Rented work lives on a tool's terms. Owned work lives in files and git you control.
27
a screenshot of a model naming your product proves one thing. the model was capable of saying it once.
i've seen this in three decks now, presented as evidence the ai visibility work is paying off.
ask the same question five times and you can get five different answers, with different products named in each. one citation is a sample of one from a distribution nobody measured.
fix a prompt set. run it repeatedly. report your share of the pool.
13
put these in one private github repo before your next laptop dies.
- the app list. winget, the package manager built into windows, installs 35 apps from it in 15 minutes
- the install and restore scripts
- every config file. symlinks, shortcuts that point each app at its settings file in the repo, turn a settings change into a commit
- your claude setup: CLAUDE.md, your skills and your routine files
- a .env.example with every key name and an empty value
two things stay out on purpose: the real keys, and bulky data such as notes and databases. syncthing copies those from machine to machine.
rebuilding from memory takes about 8 hours. rebuilding from the repo took me about 45 minutes.
ALT Nine restore steps in three groups, apps and settings, data, and Claude, all read from one repo. Keys stay out of git. Result, a matching machine in 45 minutes.
57
write your conference abstract to the reviewer, not the room
they're reading eighty of these in one sitting, tired, looking for a reason to say no. two lines is what they get: what problem this solves, why you're the person who can talk about it
rank the conferences before you write a word. one abstract fired at eight events collects eight rejections that teach you nothing, because the same talk needs a different frame for every track
charm comes later. or never
10
a one-time sale cannot produce recurring revenue. i needed a written plan to notice that.
free workbench, 29 dollar book, and a document titled "how this reaches 10k a month". it died in its own first paragraph. one sale is one sale. the server bill arrives every month and the 29 dollars does not.
so the model changed before any of the marketing did. the workbench stays free. the book becomes the front door. the recurring layer is a subscription with hosted ai behind it.
nothing paid gets marketed until the product can take a recurring payment. that was the part that did not exist yet.
1
8
last-click attribution rewards the channel that captured the demand rather than the one that created it.
i watch teams cut the podcast and the community because the dashboard credits google for everything, usually your own brand name typed back at you.
buyers research in slack groups and dms your tracker can't see. the channel that started the conversation never shows up in the report.
ask buyers how they found you. one free-text field on the demo form, read weekly.
1
14