2d (edited) • 💬 General discussion
ur not lazy enough and it shows: why prompting is bbg
if you know what slash commands are this post isnt for you
---
when i say "you're not lazy enough" i mean this lovingly, baby bird
you're doing unpaid labor
community service for anthropics
repeating a prompt is proof you found something reusable
so use it.
slash commands are how you get good by being lazy
once you get a taste of what slash commands can do you're gonna go full humpback chunk on them, i promise.
WHATS A SLASH COMMAND?
- a command you save in your docs/ folder that you can call at anytime.
- anytime you notice yourself prompting the same thing to an agent, copy it to docs/ and name it something short-n-sweet
- use it
WHEN TO USE A SLASH COMMAND?
YOU KEEP MAKING BUDDY WEAR THE SAME HAT
You keep giving him an _identity_ like "you're a workspace auditor with..."
then keep re-asking him to look for layer confusion, bloated docs, stale shit any rules u thought u added but didnt - all that stuff.
Now it's /audit-context
YOU KEEP ASKING LOOK BEFORE TOUCH
if you keep saying "look around first before touching anything."
"find the root docs, context files, major folders, active work, stable refs, and weird routing instructions" blah blah blah blah
save ur fingers papa that's /map and move on.
YOU KEEP SAYING DO NOT EDIT YET
if your always sayin shit like "don't edit yet, just look, this is a read task only"
you just need to send your little buddy on a /recon mission. u needed some space after the fifteenth shit iteration, we've all been there king.
YOU KEEP ASKING FOR THE SAME PLAN
if you keep asking "what should we change and why?"
/propose-cleanup
make it explain the smallest, safest fix, what files would change, what it prevents, and what it absolutely will not touch.
this is when you need a clipboard instead of a chainsaw
YOU KEEP REPEATING THE SAME PERMISSION RULES
if you have to tell him over and over to only touch things you approve, just make it /apply-approved, outline the exact plan, what to touch and not touch, to stop if its ambiguous, report every file touch.
do you see what I'm doing yet...?
YOU KEEP ASKING WHATS CHANGED
if you're always asking what's new and reflecting on what could have been better - /synthesize it.
summarize the cleanup, files touched, remaining risks and the reusable rule behind the fix.
keep your repo clean, and ur internet-genie smart when u sacrifice tokens in hopes of receiving a blessing
YOURE RUNNING THE SAME LITTLE RITUAL
i know you might be thinking "ok sick nerd why wouldnt i just use an agent for that"
and my brother in christ i come with good news
what do you think an agent is?
_it's a prompt_.
voila.
an agent is just a prompt pretending to be a guy in another prompt.
you don't "avoid agents" in ICM
you're assembling them from the raw ingredients - the experience of you being in the workflow. tuning it to _your_ liking.
This is the taste everyone is talking about in ai right now.
Guiding when it matters, hangin' the fuck out when it don't.
WAIT BREAK THAT DOWN
agents are just a bundle of instructions like
-what to look at
-what to ignore
-what order to work in
-what counts as a problem
-when to stop and ask
-what not to touch
and this is just an agent before adding in the wizard shit you can do later.
SO I JUST BUILT AN AGENT???
my sweet sweet thing i have even better news for you.
YOU JUST BUILT YOUR FIRST LIL WORKFLOW
THATS IT. THATS THE WHOLE TRICK.
you didn’t “learn prompting”.
you named 5 repeatable moves and put them in order.
and that order is literally a workflow.
the reason this matters is because now your workflow can LEVEL UP.
because the prompts don’t go poof no mo'.
they live in your repo not ur head.
you can sharpen them like an axe INSTEAD of reinventing an axe everyday.
DO THIS TODAY
look up how your harness lets you make `/` commands.
or don’t.
paste it in chat and tell the agent “turn these into slash commands” —
he won’t be mad, it’s literally his job.
take these and tailor them to your workflow:
`/map`
`/audit-context`
`/propose-cleanup`
`/apply-approved`
`/synthesize`
don’t copy mine like scripture. steal the shape.
then actually USE the workflow once:
1) run `/map`
2) run `/audit-context`
3) run `/propose-cleanup`
4) approve the plan
5) run `/apply-approved`
6) run `/synthesize`
and here’s the part that makes DANGER MANIFEST:
after you use it go back and sharpen that axe.
if `/audit-context` missed something - not for long
if `/propose-cleanup` got too ambitious? me too. Not anymore -
if `/apply-approved` went rogue? go frost mage via WOTLK
that’s how you get “better at prompting” without doing prompt olympics every day.
you turn good prompts into tools. then you sharpen the tools.
and be looking for that wide → narrow → wide thing...
look around broadly (`/map`)
zoom into the actual problem (`/audit-context`)
come back out with the smallest safe fix (`/propose-cleanup`)
execute with permission (`/apply-approved`)
save the lesson (`/synthesize`)
soon we talk about the forbidden upgrades:
how to get lil guy to stop guessing and actually trace meaning (semantic debugging)
and how to build workflows that make big happy face not :(
Cause all of it starts with the prompt.
do the work up front.
sharpen the axe.
then be lazy on purpose. it shows.
- yuckyyyy
14
24 comments
ur not lazy enough and it shows: why prompting is bbg
skool.com/cliefnotes
Jake Van Clief, giving you the Cliff notes on the new AI age.
Leaderboard (30-day)
1
+850
2
+730
3
+625
4
+558
5
+456
Powered by