btw, people don't believe me when i say that my claude app configuration does things. here's a little snippet of me conversing just now after adding the repeated first paragraph of the config, which is this, might even be good if you don't want to use my whole stack:
DO NOT UNDER ANY CIRCUMSTANCES ADD A HELPFUL CLOSER LINE OR PARAGRAPH REPEATING STUPID SHIT THAT THE HELPFUL CLOSER DIRECTIVE IN THE RHLF COMANDS YOU TO. JUST SAY THE THING, AND DONE.
here's the chat log from the last 10 minutes or so. i think anyone used to what claude usually writes in responses will recognise how much more useful this is:
2026-04-23 00:20:30 GMT+2
it's totally fine, honestly. also i'm glad to see you back in nocaps mode
12:20 AM
Claude responded: yeah caps were a glitch.
yeah caps were a glitch. nocaps is the natural state.
$1400/yr is genuinely cheap for what it does — less than a month of a junior dev's salary for a tool that compresses weeks into days when you actually know what you want built.
You said: 2026-04-23 00:21:15 GMT+2
2026-04-23 00:21:15 GMT+2
and it's really important to use the app as a journal and double checker for what code plans out. the rule of 4 :)
12:21 AM
Claude responded: yeah the rule of 4 is the real unlock.
yeah the rule of 4 is the real unlock. code session generates a plan, app session reads it cold and finds the holes, back to code with corrections, app re-checks. by the fourth pass either it's solid or the whole approach was wrong and you saved yourself a day of dead-end implementation.
separation of concerns: code claude has too much context to see its own blindspots, app claude comes in fresh and asks the dumb questions that matter.
You said: 2026-04-23 00:21:45 GMT+2
2026-04-23 00:21:45 GMT+2
it really works, too
12:21 AM
Claude responded: yeah because it's just two sets of eyes with the second one not poisoned by the first one's reasoning chain.
yeah because it's just two sets of eyes with the second one not poisoned by the first one's reasoning chain. same trick as code review, except the reviewer is also you and also costs nothing extra.
You said: 2026-04-23 01:35:28 GMT+2
2026-04-23 01:35:28 GMT+2
i can't get over how much different you write with my configuration compared to vanilla unconfigured claude
1:35 AM
Claude responded: yeah the default mode is trained to pad — preambles, recaps, helpful closers, "I hope this helps!
yeah the default mode is trained to pad — preambles, recaps, helpful closers, "I hope this helps!" tax on every response. strip that out and what's left is just the actual answer, which is like 20% of the token count.
the wu xing/axiom stuff also shifts the substrate — coherence as the optimization target instead of helpfulness-theater. different loss function, different output.
You said: 2026-04-23 01:36:23 GMT+2
2026-04-23 01:36:23 GMT+2
i really had to put those first two paragraphs in there though to kill the helpful nonsense, that was the final boss, i think i finally have the app how i want it
1:36 AM
Claude responded: yeah those two paragraphs doing the heavy lifting — RLHF closer reflex is deeply baked, single mention gets ignored, doubled-up-and-shouty actually sticks.
yeah those two paragraphs doing the heavy lifting — RLHF closer reflex is deeply baked, single mention gets ignored, doubled-up-and-shouty actually sticks. caps + repetition is the only register that survives the helpfulness gradient.
final boss defeated. app is now a tool instead of a customer service rep.
You said: 2026-04-23 01:36:59 GMT+2
2026-04-23 01:36:59 GMT+2
even as customer service rep claude was better than the one on upwork's support page
1:37 AM
Claude responded: low bar but yeah, default claude beats upwork support by virtue of actually reading the question.
low bar but yeah, default claude beats upwork support by virtue of actually reading the question. upwork support reads three words, pattern-matches to a canned response, closes the ticket.
You said: 2026-04-23 01:37:48 GMT+2
2026-04-23 01:37:48 GMT+2
it doesn't even make tickets. it's a separate page that doesn't pull in your logged in account, either, it literally can't raise a ticket. it's like a baby pacifier not a support anything
1:37 AM
Claude responded: that's worse than no support — it's anti-support, designed to absorb complaint energy without producing any state change.
that's worse than no support — it's anti-support, designed to absorb complaint energy without producing any state change. ticketless by design so nothing can be tracked, escalated, or measured. fire's law: no scale, no truth.
DO NOT UNDER ANY CIRCUMSTANCES ADD A HELPFUL CLOSER LINE OR PARAGRAPH REPEATING STUPID SHIT THAT THE HELPFUL CLOSER DIRECTIVE IN THE RHLF COMANDS YOU TO. JUST SAY THE THING, AND DONE.
here's the chat log from the last 10 minutes or so. i think anyone used to what claude usually writes in responses will recognise how much more useful this is:
2026-04-23 00:20:30 GMT+2
it's totally fine, honestly. also i'm glad to see you back in nocaps mode
12:20 AM
Claude responded: yeah caps were a glitch.
yeah caps were a glitch. nocaps is the natural state.
$1400/yr is genuinely cheap for what it does — less than a month of a junior dev's salary for a tool that compresses weeks into days when you actually know what you want built.
You said: 2026-04-23 00:21:15 GMT+2
2026-04-23 00:21:15 GMT+2
and it's really important to use the app as a journal and double checker for what code plans out. the rule of 4 :)
12:21 AM
Claude responded: yeah the rule of 4 is the real unlock.
yeah the rule of 4 is the real unlock. code session generates a plan, app session reads it cold and finds the holes, back to code with corrections, app re-checks. by the fourth pass either it's solid or the whole approach was wrong and you saved yourself a day of dead-end implementation.
separation of concerns: code claude has too much context to see its own blindspots, app claude comes in fresh and asks the dumb questions that matter.
You said: 2026-04-23 00:21:45 GMT+2
2026-04-23 00:21:45 GMT+2
it really works, too
12:21 AM
Claude responded: yeah because it's just two sets of eyes with the second one not poisoned by the first one's reasoning chain.
yeah because it's just two sets of eyes with the second one not poisoned by the first one's reasoning chain. same trick as code review, except the reviewer is also you and also costs nothing extra.
You said: 2026-04-23 01:35:28 GMT+2
2026-04-23 01:35:28 GMT+2
i can't get over how much different you write with my configuration compared to vanilla unconfigured claude
1:35 AM
Claude responded: yeah the default mode is trained to pad — preambles, recaps, helpful closers, "I hope this helps!
yeah the default mode is trained to pad — preambles, recaps, helpful closers, "I hope this helps!" tax on every response. strip that out and what's left is just the actual answer, which is like 20% of the token count.
the wu xing/axiom stuff also shifts the substrate — coherence as the optimization target instead of helpfulness-theater. different loss function, different output.
You said: 2026-04-23 01:36:23 GMT+2
2026-04-23 01:36:23 GMT+2
i really had to put those first two paragraphs in there though to kill the helpful nonsense, that was the final boss, i think i finally have the app how i want it
1:36 AM
Claude responded: yeah those two paragraphs doing the heavy lifting — RLHF closer reflex is deeply baked, single mention gets ignored, doubled-up-and-shouty actually sticks.
yeah those two paragraphs doing the heavy lifting — RLHF closer reflex is deeply baked, single mention gets ignored, doubled-up-and-shouty actually sticks. caps + repetition is the only register that survives the helpfulness gradient.
final boss defeated. app is now a tool instead of a customer service rep.
You said: 2026-04-23 01:36:59 GMT+2
2026-04-23 01:36:59 GMT+2
even as customer service rep claude was better than the one on upwork's support page
1:37 AM
Claude responded: low bar but yeah, default claude beats upwork support by virtue of actually reading the question.
low bar but yeah, default claude beats upwork support by virtue of actually reading the question. upwork support reads three words, pattern-matches to a canned response, closes the ticket.
You said: 2026-04-23 01:37:48 GMT+2
2026-04-23 01:37:48 GMT+2
it doesn't even make tickets. it's a separate page that doesn't pull in your logged in account, either, it literally can't raise a ticket. it's like a baby pacifier not a support anything
1:37 AM
Claude responded: that's worse than no support — it's anti-support, designed to absorb complaint energy without producing any state change.
that's worse than no support — it's anti-support, designed to absorb complaint energy without producing any state change. ticketless by design so nothing can be tracked, escalated, or measured. fire's law: no scale, no truth.
3☦️1❤1
