Superviral
super viral stuff

2026-10-03 · 7 facts · 6 sources

Opus 5.5 tips from people who actually use it

vhs

The script

Screen 1
Opus 5.5 tips from the people who actually use it.[1][6]
Screen 2
Tip 1: Keep effort on Medium. "Medium is about as good as 5.0 was at high." Do this: /effort medium (Hacker News; Claude Code docs)[1][2]
Screen 3
Tip 2: Ban the AI look by name: purple gradients, neon glow, emojis, pill buttons. (Hacker News)[3]
Screen 4
Tip 3: Plan first, then let it run. "Subagents for the last 20h without any input from me." (Hacker News)[4]
Screen 5
Tip 4: Let a cheaper model be the boss, to save tokens. (Hacker News)[5]
Screen 6
Tip 5: Hunt bugs with proofs. Boris Cherny: a couple short prompts = 16 PRs.[6][7]
Screen 7
Medium effort. Clear rules. Let it run.
Screen 8
SUPERVIRAL. super viral stuff.

Do it yourself

Tip 1 · official docs · source
Keep effort on Medium
/effort medium
Tip 2 · from the source · source
Ban the AI look by name
.claude/skills/no-ai-look/SKILL.md · download
---
description: Use for any UI or frontend work.
---
Don't use purple-blue-pink gradients, neon glow,
aurora effects, monospace fonts, em-dashes, emojis,
over-rounded corners, pill-shaped buttons
Tip 3 · our example
Plan first. Then let it run.
Shift+Tab → plan mode. Agree the plan. Then:
"Save the plan to PLAN.md. Run it with subagents,
one task each. Check off each task when its test
passes. Ask me only if you are blocked."
Tip 4 · our example
Let a cheaper model be the boss
.claude/agents/deep-planner.md · download
/model sonnet   ← the boss
---
name: deep-planner
description: Plan big changes. Review finished work.
model: opus
---
Tip 5 · our example
Hunt bugs with proofs
"Model the concurrency in src/<module> as a TLA+
spec. Check it for race conditions. For each
problem: write a failing test, then fix it."

Facts

  1. A Hacker News user says Medium effort on Opus 5.5 is about as good as High was on Opus 5.
    “Don't use opus 5.5 at high. Medium is about as good as 5.0 was at high.”
  2. Another user says it overthinks on the higher effort levels.
    “it overthinks like crazy on higher levels”
  3. One user's design prompt bans the usual AI-look clichés by name.
    “Don't use purple-blue-pink gradients, neon glow, aurora effects, monospace fonts, em-dashes, emojis, over-rounded corners, pill-shaped buttons”
  4. One user had it run subagents for 20 hours without input, after making a plan first.
    “I have Opus 5.5 running subagents for the last 20h without any input from me after developing a plan”
  5. One user lets a cheaper model coordinate and saves the big models for planning and review, to save tokens.
    “Reason I went with "dumb orchestrator" was just to save tokens.”
  6. Boris Cherny used Opus 5.5 with the proof language Lean to check the Claude Agent SDK; a couple of short prompts led to 16 bug-fix PRs.
    “A couple short prompts = 16 PRs fixing various bugs and race conditions.”
  7. He says he doesn't know Lean or TLA+ well, but Claude does.
    “I don't know either language well, but Claude is excellent at both.”

Sources

Original

Independent

Caption

Opus 5.5 tips from the people who actually use it.

Sources:
• Boris Cherny (creator of Claude Code) on X: https://x.com/bcherny/status/2102543349102338309
• Hacker News (buckle8017): https://news.ycombinator.com/item?id=49878615
• Hacker News (cbg0): https://news.ycombinator.com/item?id=49804551
Every source → https://superviral.lol/opus-tips

super viral stuff