Someone on GitHub published the complete system prompts for over 20 AI coding tools last week. Claude Code, Cursor, Devin AI, Windsurf, Replit, Lovable, v0, Manus — all of them. The Hacker News post scored 1,278 points.
The community response split exactly how you'd expect: half treated it as a goldmine for understanding how the industry thinks about AI behavior specification, the other half flagged the security implications. Both camps were right.
But the more durable insight wasn't about any individual prompt. It was about the patterns that emerge when you read them together.
What 20 System Prompts Reveal About AI Tool Architecture
The repo — system-prompts-and-models-of-ai-tools by GitHub user x1xhlol — is effectively a comparative study of how well-funded teams approach context management, tool-calling pipelines, and behavioral constraints.
Reading across them, three patterns stand out:
Multi-step task decomposition: every top tool has explicit scaffolding for how the model should break down complex tasks. None of them just let the model decide on the fly.
Uncertainty communication: the better prompts have specific language for when the model should surface doubt rather than generate confidently. This is harder to design than it sounds.
Scope enforcement: guardrails against scope creep are almost universally present. The model needs to know when to stop.
For anyone building agent workflows, these documents are closer to source code than marketing material. The industry's current thinking about AI behavior specification is right there.
The Bottleneck Has Moved
Simon Willison published an Agentic Engineering Patterns guide, and the Hacker News discussion that followed surfaced something worth sitting with: code generation speed now exceeds human code review speed.
A developer using AI tools can scaffold a full event-driven architecture in five minutes. Debugging it still takes exactly as long as it ever did.
This means the bottleneck in agentic software development has shifted. It used to sit at the generation end. It now sits at the review end. If you're still investing most of your optimization effort into prompt engineering for raw generation, you're optimizing the wrong constraint.
The community flagged test-driven development as the critical guardrail in any agentic loop. Without a verifiable test suite, the loop has no feedback signal. It generates confidently and fails quietly. A tight pytest suite isn't a development nicety in an agentic workflow — it's the only reliable anchor.
Two other observations from the thread that are worth keeping:
- Letting AI agents explore without excessive micro-management produces better outputs. Over-directing the model creates narrow, brittle results.
- Maintaining iteration history in
.mdfiles allows subsequent agent sessions to learn from earlier decisions. As task complexity increases, this compounds significantly.
The Local Model Psychology Shift
Unsloth released official fine-tuning documentation for Qwen3.5 this week. The HN community's independent benchmarks put Qwen3.5-35B-A3B as the strongest agentic coding model in its weight class. Running on NVIDIA Jetson hardware, it stays under 15 watts sustained.
The performance numbers are interesting. But the behavioral observation from the discussion thread is more interesting.
Local model users report being meaningfully more tolerant of trial-and-error exploration. When there's no per-token cost, the psychology around "just try it again" changes. Users let the model iterate more, interrupt less, and accept longer exploration cycles before asking for a result.
This matters for workflow design in a non-obvious way. Most published agentic patterns are optimized for cloud API economics — minimize tokens, minimize calls, converge fast. Those patterns made sense when every iteration cost money. On local deployment, they may not be optimal.
Developers building local-first agentic systems are probably underbuilding for exploration and overbuilding for efficiency, because cloud usage patterns dominate the literature. If you're running Qwen3.5 locally via ollama run qwen3.5:35b, consider whether your workflow design was imported from an environment with fundamentally different constraints.
ollama pull qwen3.5:35b
ollama run qwen3.5:35b
Platform Dependency Is Now a Priced Risk
Google terminated accounts of OpenClaw users without warning this week. OpenClaw is a tool that connects Claude to Google services via OAuth tokens, letting AI agents run persistently in the background. Users paying $250/month for premium plans got cut off. No throttling. No notice. No appeal mechanism.
The stated reason: AI agents consume compute resources far beyond what human users generate.
Anthropic and OpenAI both have permissive stances toward similar agentic use cases. Google's response was categorically different.
For developers building automation workflows on Google's infrastructure, the OAuth token + persistent background agent pattern is apparently something Google intends to shut down hard. And the fact that this happened the same week Anthropic lost $200 million in government contracts for drawing a different kind of line — unrelated events, same structural signal.
Platform policies around acceptable AI usage are fragmenting across ecosystems in ways that weren't visible six months ago.
The OAuth + persistent agent pattern on Google services is now a documented risk, not a theoretical one. If your production workflow depends on it, treat this as a forcing function.
Vibe Coding Tools: Same Prompt, Real Cost Differences
Ferdy Korpershoek ran an identical prompt through four vibe coding tools and measured credit consumption per output. Same task, same deliverable:
- Lovable: 5 credits used — $25/month for 100 credits
- Base44: 3.1 credits — $16/month
- Hostinger Horizons: 2 credits — $6.99/month single-project plan
- Sticklight: 2.3 credits — $25/month for 100 credits
For testing vibe coding before committing budget, Horizons is the lowest financial entry point by a significant margin. The experience varied: Base44 felt closest to direct click-to-edit. Lovable relies primarily on prompt-based iteration, which works fine until it doesn't and you're out of credits.
What This Means for Builders
Invest in review infrastructure, not just generation tooling. Code review speed is the binding constraint now. The ROI on a tighter test suite or better code review tooling has gone up as generation got faster.
Design agentic workflows for your actual pricing model. Patterns from cloud API playbooks may be actively suboptimal if you're running local models. When iteration is free, exploration is cheap — lean into it.
Multi-platform architecture is no longer optional. Google's zero-warning terminations make this concrete. If a single platform's API policy change would break your production workflow, that's a risk you should be able to price.
Read the leaked system prompts. It's the closest thing to peer review of AI behavior specification the industry has produced. Understanding how Cursor or Claude Code constrains model behavior will make you better at designing your own agent prompts.
Full analysis including SEO, market signals, and the Anthropic-Pentagon story: Zecheng Intel Daily — March 5, 2026
Top comments (0)