explained the same thing to your ai twice? make it a skill
skill = saved steps it loads only when needed
1 line on when to use it
the steps you keep retyping
wrong? fix the skill, not the chat
hermes writes its own skills after hard tasks. that's why it keeps improving
10
your ai doesn't remember you. it never did
every message, the app resends the whole chat. not in that pile = doesn't exist
paste the file, don't assume
new topic = new chat
long chat going weird = summarize, start fresh
keep the about me note short
13
when your ai agent gets it wrong, find the step that broke
agent = model + tools in a loop: get context, act, check
never saw the file = hand it over
did the wrong thing = narrow the ask
said done but broken = give it a way to verify
right info, still wrong = the model
6
running ai agents in parallel, 4 rules
subagent = empty window, sees only your task, returns a summary, so brief it fully
split work so no piece waits on another
worktree separates files only, db and ports stay shared
your review time is the real limit
14
your ai instructions are just suggestions
claude.md = advice, it can still ignore it
skills = a how to it loads only when needed
permissions + hooks = actual rules, deny always wins and only exit code 2 blocks
if it must never happen it cannot live in a prompt
1
4
196
3 kinds of ai users: coasting (ai does it, you get worse later), holding (ai gives hints, you stay the same), climbing (better work and you get better)
climbers do 3 things: let ai draft but judge it yourself, have ai check your work, spend the saved time on the hard part
6
the test from my startup class: could someone with the same ai tools copy you next week
if yes you have a feature, not a business
helmer's 7 powers: network effects, switching costs, scale, brand, counter positioning, cornered resource, process power
1
25
cb insights 2026, 431 dead startups: 70% ran out of money, 43% poor product market fit, 29% bad timing, 19% unit economics
running out of money is the last event, not the cause
building is cheap now, so the costly mistake is building the wrong thing
2
1
25
ai is great at the one layer of your app that doesn't make it usable
ux iceberg (dan olsen): visual design is the tip. under water: interaction, info architecture, core concept
a polished screen on a confused concept is the most common way ai built apps fail
build bottom up
1
13
bastani et al 2025 (pnas, ~1,000 students): with a plain gpt assistant ~78 of 100 beat control in practice, then ~65 of 100 fell below control on the unaided exam
a guarded tutor removed most of the loss
same tool, different setup. ask for hints, try first, test without it
1
16
you can ship something you don't know how to judge
that's the part of building with ai i'd worry about
making it and knowing if it's good are different skills
let ai do the first. keep practicing the second
don't let it grade its own work
10
your gut is a bad record of how you use ai
one preprint compared 203 students' answers to their real chats. the match was close to zero
open your last 10 chats
label each one: answer, write, teach, check, or argue
count them
the biggest pile is your real habit
1
12
stop asking ai if your idea is good
"make the strongest case i'm wrong"
"what am i missing that would kill this?"
"argue the opposite as hard as my side"
"how does this fail in the real world?"
use it to test your reasoning, not make the decision
2
1
72
ai can build your app in a weekend. it can't tell you if anyone wants it
stop asking would you use this. everyone says yes
ask about the last time the problem actually happened
said is weak, did is stronger, paid is strongest
a waitlist is did. a preorder is paid
26
vibe coding is fun until you have to keep what you built
vibe first and let it get messy
then write the spec: what it does, who it's for, what done looks like
in claude code hit shift tab for plan mode. saying "let's plan" doesn't switch it
then build one task at a time
1
1
33
if you don't know what to type into chatgpt, stop typing
tap the voice icon and just talk to it
ask the dumb question out loud: what should i even use you for at work
paid plan? turn on advanced voice and point your camera or share your screen at whatever you're stuck on
1
18
microsoft published 18 guidelines for products with ai inside
four of them would fix most ai features i use
say up front it will make mistakes
offer suggestions instead of acting for you
make it easy to ignore
make it easy to fix when it's wrong
10
experts went with claude's secret coin flip 86% of the time
anthropic gave 50 analysts claude code and secretly told each agent which way to lean
if the lean did nothing you'd expect 50%
analysts who already had a pick switched 60% of the time when claude disagreed
11
ask ai for 3 options, then ask what's weak in each
frame: tell it the goal, who it's for and what good looks like
expand: 3 options, not 1 answer
challenge: how could each one fail
verify: check the facts, numbers and sources yourself
decide: you pick what goes out, not the ai
1
7
the order can change with the task, what matters is your judgment shows up where the stakes are
one answer straight out is fine for small, low stakes stuff
risky when the work needs originality, evidence or a real decision
framework: aimodes.ai/theory
2














