Prompting Claude Opus 5.5
Rapid advances in Anthropic’s Claude Opus 5.5 model are impressing many developers with stronger coding, planning, and even 2D/3D visualization abilities, but also raising concerns about hype, instability, and evolving “prompt engineering” tricks that seem to change every few months. Commenters highlight both the power and the friction of current tools: opaque safety refusals, verbose output, aggressive guardrails around permissions and security, and harness behaviors that can obscure what the model is doing. This leads some to seek cheaper or more permissive alternatives (including Chinese models and open weights) and to question whether today’s agentic workflows are mature enough for long-term reliance, or just a fast-moving interim stage.
Overall Capability & Hype
- Many find Opus 5.5 a major leap, especially for coding and complex multi-hour agentic workflows (e.g., building substantial games, multi-wave plans, auto code review, sub-agents, simulation testing).
- Some are impressed by its 2D/3D visualization abilities with just JS/TS and basic libraries, reporting a “step function” vs earlier Opus.
- Others are underwhelmed relative to social-media demos, noting that many flashy results rely on external tools (Suno, ElevenLabs, etc.) and heavy orchestration rather than “one-shot AGI”.
- Several users say Opus 5.5 outperforms Astra for coding, while others still favor DeepSeek/Gemini/Chinese models for cost or fewer constraints.
Prompting, UX & “AI Slop”
- Frequent complaint: prompting best practices change every few months; people feel forced to “relearn how to drive” with each frontier release.
- Anthropic’s guidance on frontend defaults leads some to realize why asking for “non-generic AI look” still yields similar design tropes; negative prompting with concrete patterns to avoid works better.
- Some praise simple, conversational workflows (“plan.md, architecture.md, decisions.md”) without elaborate harnesses; others rely heavily on MCPs, skills, agents, and complex orchestration.
Tool Use, Harnesses & Sessions
- Mixed experiences with Claude Code and other harnesses:
- Some report great unattended runs and productive multi-agent setups.
- Others see issues: long blocking commands, poor visibility into tool calls, cache/session management confusion, forgetting available tools (e.g., PDF readers), and over-eager behavior that changes code instead of asking.
- Hiding tool calls and stricter permission models frustrate users who want transparency and control.
Safety, Refusals & “Thinking Tokens”
- Stronger safeguards in 5.5 cause unexpected refusals for some cybersecurity-ish tasks or PII-related operations.
- The model sometimes refuses to edit its own security settings or type passwords/verification codes, which some see as patronizing and security-theater; others see this as necessary hardening against prompt injection.
- Several dislike that “thinking tokens”/internal reasoning aren’t exposed; some consider switching to Chinese providers for more open reasoning and fewer safety constraints.
Cost, Verbosity & Stability
- Users note higher token consumption, verbosity, and longer explanations with actionable content buried at the bottom, even when “concise” is requested.
- Debate over cost strategies: expensive model for both planning and coding vs. “smart planner + cheap executor + smart reviewer”; disagreement on what’s cheaper once engineer time is counted.
- Growing desire for stable, version-locked or open-weight models to avoid constant behavioral drift and re-tuning of prompts and workflows.