Sabitlenmiş Tweet
Vox
3.7K posts
Vox
@Voxyz_ai
Building Voxyz. The only human inside. Practicing the AI-native future before it arrives. Sharing every step on the road. ↓
London Katılım Kasım 2025
169 Takip Edilen16K Takipçiler
𝘀𝘁𝗼𝗽 telling Claude Code/Codex "no em dashes".
𝘀𝘁𝗼𝗽 telling Claude Code/Codex "stop saying delve".
𝘀𝘁𝗼𝗽 telling Claude Code/Codex "don't sound like AI".
you never gave it a 𝘄𝗿𝗶𝘁𝗶𝗻𝗴 𝘀𝘆𝘀𝘁𝗲𝗺. every README, PR description and landing page ships in the same AI voice, and you ban words one at a time.
orwell wrote the fix 𝟴𝟬 𝘆𝗲𝗮𝗿𝘀 𝗮𝗴𝗼. six rules, 1946. paste them into your global CLAUDE.md / AGENTS.md and every session picks them up:
→ never use a long word where a short one will do
→ if it is possible to cut a word out, cut it out
→ never use the passive where you can use the active
6 blocks you can copy-paste directly 👇
English
@LoopOnChain In theory yeah, each part works on its own. putting it all together is where it gets messy. social media still takes a little luck lol
English
@Voxyz_ai Do you see this happening a lot? I feel like I see these posts sometimes and then I look into it and it just doesn't seem like the workflow is repeatable
I have tried zero AI UGC though so i'm coming from a place of limited knowledge lol
English
𝗖𝗹𝗮𝘂𝗱𝗲 𝗖𝗼𝗱𝗲 closed the AI UGC loop: one folder → a week on TikTok + IG.
Jay's 60 creators hit 44M views/month.
His loop: test formats, validate winners, scale. The guide: AI reactions, real app footage, Postiz scheduling.
AI lets the founder run more clips per loop. They decide which ones get another round.
Watch the video, then read the full guide below. 👇
Will@athcanft
English
@Voxyz_ai GPT-5.6 Sol better for debugging.
Fable still good in writing.
English
I'm starting to think GPT-5.6 Sol may be better than Fable for writing.
Fable makes the writing sound polished, and that familiar AI voice comes with it. Sol is cleaner out of the box. the system below may work better with Sol 👇
Fable or Sol? haven't tried K3 yet. any good?
Vox@Voxyz_ai
𝘀𝘁𝗼𝗽 telling Claude Code/Codex "no em dashes". 𝘀𝘁𝗼𝗽 telling Claude Code/Codex "stop saying delve". 𝘀𝘁𝗼𝗽 telling Claude Code/Codex "don't sound like AI". you never gave it a 𝘄𝗿𝗶𝘁𝗶𝗻𝗴 𝘀𝘆𝘀𝘁𝗲𝗺. every README, PR description and landing page ships in the same AI voice, and you ban words one at a time. orwell wrote the fix 𝟴𝟬 𝘆𝗲𝗮𝗿𝘀 𝗮𝗴𝗼. six rules, 1946. paste them into your global CLAUDE.md / AGENTS.md and every session picks them up: → never use a long word where a short one will do → if it is possible to cut a word out, cut it out → never use the passive where you can use the active 6 blocks you can copy-paste directly 👇
English
Vox retweetledi
I had claude code and codex review each other's plans, both directions. gut call is clear: 𝗳𝗮𝗯𝗹𝗲 𝘄𝗿𝗶𝘁𝗲𝘀 𝘁𝗵𝗲 𝗽𝗹𝗮𝗻, 𝗴𝗽𝘁-𝟱.𝟲-𝘀𝗼𝗹 𝗰𝗮𝗹𝗶𝗯𝗿𝗮𝘁𝗲𝘀 𝗶𝘁. the reverse wasn't close.
the whole setup is 𝗼𝗻𝗲 𝗽𝗿𝗼𝗺𝗽𝘁 👇
- fable owns the edits, the reviewer stays read only
- their blind spots don't overlap, they catch different holes
- a finding only makes it into the plan after it survives a falsification check
honestly, 𝗲𝘅𝗲𝗰𝘂𝘁𝗶𝗼𝗻 𝘀𝘁𝗶𝗹𝗹 𝗴𝗼𝗲𝘀 𝘁𝗼 𝗰𝗼𝗱𝗲𝘅. fable wins at thinking a plan through, sol wins at tearing it apart.
setup:
1. the plan: no prompt from me for this step, you already have your own way to get a plan. i mostly use ce plan.
2. the review prompt: get a second opinion on this plan from codex CLI running gpt-5.6-sol at max effort. every objection needs a severity, a basis (fact or inference), a falsification check, and a suggested change, 5 objections max. revise the plan only with findings that survive falsification, reject and log the rest. repeat until no new objection holds, up to 5 rounds. report what changed and what got rejected after each round.
3. once it converges, have fable save the whole flow as a skill. next time one line triggers it.
swap the review panel however you want, use whatever CLIs you already have installed.
English
one more thing: Orwell's rules for Claude Code/Codex missed the AI tic i recognize on sight: 𝗰𝗼𝗿𝗿𝗲𝗰𝘁𝗶𝘃𝗲 𝗷𝘂𝘅𝘁𝗮𝗽𝗼𝘀𝗶𝘁𝗶𝗼𝗻.
the not X, it's Y shell:
the answer isn't X, it's Y
X isn't the problem, Y is
same mold, almost any topic.
my guess: it gives a plain statement the shape of insight without adding a fact. training supplies the pattern; post-training rewards the clean, decisive sound.
i use it in normal conversation. the structure is fine. rhetoricians call the family 𝗮𝗻𝘁𝗶𝘁𝗵𝗲𝘀𝗶𝘀; the corrective form is 𝗰𝗼𝗿𝗿𝗲𝗰𝘁𝗶𝗼. JFK's ask not and MLK's color/content belong to it.
rhetoricians used it for 2,000 years. LLMs wore it out in a few. wonder which good expression they'll ruin next lol.
six more rules for CLAUDE.md / AGENTS.md. the 𝟮𝟬𝟮𝟲 𝗽𝗮𝘁𝗰𝗵:
7. don't build a straw man to knock down. use not X, it's Y once per piece, max
8. two examples are enough. don't stretch to three
9. don't announce what you're about to say. say it
10. don't end two paragraphs in a row with punchlines
11. vary the length and shape of neighboring sentences
12. break any of these rules sooner than write like a machine
Vox@Voxyz_ai
𝘀𝘁𝗼𝗽 telling Claude Code/Codex "no em dashes". 𝘀𝘁𝗼𝗽 telling Claude Code/Codex "stop saying delve". 𝘀𝘁𝗼𝗽 telling Claude Code/Codex "don't sound like AI". you never gave it a 𝘄𝗿𝗶𝘁𝗶𝗻𝗴 𝘀𝘆𝘀𝘁𝗲𝗺. every README, PR description and landing page ships in the same AI voice, and you ban words one at a time. orwell wrote the fix 𝟴𝟬 𝘆𝗲𝗮𝗿𝘀 𝗮𝗴𝗼. six rules, 1946. paste them into your global CLAUDE.md / AGENTS.md and every session picks them up: → never use a long word where a short one will do → if it is possible to cut a word out, cut it out → never use the passive where you can use the active 6 blocks you can copy-paste directly 👇
English
@archvalmiki yeah, sol feels way better to talk to right now, especially with hermes
English
@Voxyz_ai this is epic. thanks! i noticed fable does much more of this crap than sol out of the box
English
@Voxyz_ai how did you know how i talk to my business assistant?
English
@KrishanuAR yeah, i catch myself saying it too. it's not even a bad structure, but now it somehow feels wrong lol
English
@Voxyz_ai I did this with mitosis and it changed the game completely
English
6. 𝗺𝗮𝗸𝗲 𝗶𝘁𝘀 𝗿𝗲𝗽𝗼𝗿𝘁𝘀 𝗿𝗲𝗮𝗱𝗮𝗯𝗹𝗲
the ✅ Successfully implemented wall at the end of every session. add this:
Report progress in plain sentences: what changed, what failed, what comes next. No emoji checkmarks, no Successfully, no Perfect, no wall of bullets. Start with three lines; add detail only when it changes the next action.
English
5. 𝗹𝗮𝗻𝗱𝗶𝗻𝗴 𝗽𝗮𝗴𝗲 𝗰𝗼𝗽𝘆
for the hero section that sounds like every AI launch:
Rewrite this landing page copy under the six writing rules in CLAUDE.md. One concrete claim per line. Short words, active voice. Run the swap test on every line: if a competitor could paste it unchanged onto their page, rewrite it or delete it.
English