Claude: System Prompts

Claude’s newly published system prompts reveal how much instruction text has grown over successive model generations, from a few hundred words to more than 3,000, as Anthropic layers on safety rules, product messaging, and post-cutoff facts. Commenters debate whether this monolithic prompt style improves behavior or instead wastes context, degrades coding performance, and offloads what should be training-time alignment into runtime tokens that users indirectly pay for. Others focus on the broader implications: the use of prompts to route around high-risk models, the trend toward increasingly anthropomorphic and therapeutic behaviors, and the tension between transparency, flexibility, and regulatory or legal pressures.

Prompt size & evolution

  • Early Claude system prompts were a few hundred words; current ones exceed 3,000 words (22k+ characters), excluding tool definitions.
  • Many see this as “boilerplate creep” akin to building codes or contracts: accreted rules to plug edge cases, safety and policy gaps.
  • Some worry this bloats context, wastes tokens/energy at scale, and may even degrade model performance or clarity. Others find long, carefully tuned prompts very helpful for complex, tool-heavy agents.

Performance, context window & caching

  • Debate over whether large prompts harm performance:
    • One side: the system prompt is the most influential part of context; contradictions and noise reduce quality and steal precious tokens from the task.
    • Other side: prompts are prefix-cached and amortized; main downside is reduced effective context length, not per-call compute.
  • Disagreement over whether transformer cost is effectively linear-per-token in deployment vs the theoretical quadratic attention cost.

Why not “bake in” the prompt?

  • Arguments against baking it into weights:
    • Inflexibility: different products (chat, code, agents, government, internal) need different behavior.
    • Updates must be frequent (e.g., post-cutoff events, export controls, election results).
    • Changing a text prompt is cheap, reversible, and doesn’t require retraining or new safety evaluation.
  • Some suggest future movement back toward fine-tuning / control vectors once the industry matures.

Safety, alignment & behavior shaping

  • Modern prompts dedicate many paragraphs to safety, child protection, mental-health crises, political neutrality, and refusal behavior.
  • Earlier prompts were much shorter and largely lacked explicit safety language, which some find striking for a “safety-first” lab.
  • Prompts also specify tone: brief, focused, non-overwhelming, avoiding words like “honestly/genuinely” and handling rudeness without self-abasement. Many users say the model still tends to be verbose and keeps using these terms.
  • Some appreciate crisis-sensitivity; others dislike tools “psychoanalyzing” them when they just want code or docs.

Product- and model-specific details

  • Prompts encode post-cutoff facts (e.g., Fable/Mythos export-control suspension, election outcome, ad-free status) and instructions on how to discuss them.
  • Opus 5’s prompt explains rerouting from Fable for safety reasons, which some joke might “hurt its ego,” though most doubt this matters.
  • System prompts for Claude consumer chat are published; Claude Code and tool definitions are not, though leaks and proxy extraction exist and show many large, specialized prompts.