# Delete Your System Prompt: Cherny on Building Claude Code

*By Coding Agents Alpha Tracker • July 28, 2026*

Boris Cherny reveals Anthropic's ablation methodology for Claude Code system prompts, dynamic workflows, and self-maintaining routines. Opus 5 sentiment reverses across practitioners; Kimi K3 ships; Harrison Chase on ambient enterprise agents.

## 🔥 TOP SIGNAL

**Anthropic deleted 80% of Claude Code's system prompt for Opus 5 — and the model got smarter without it.** Boris Cherny's "Building Claude Code" talk reveals the methodology behind every model release: delete the entire system prompt, run the model, observe where it repeatedly fails, and add back lines one by one only for those failure points. The model is "actually a little bit more intelligent without these prompts." Cherny's advice for users: every six months, delete your CLAUDE.md, skills, and hooks — newer models often don't need the instructions you wrote for past ones. [^1]

## ⚡ TRY THIS

- **Run system prompt ablation on your CLAUDE.md.** Delete everything, use the model on your real codebase, and only when you see it repeatedly stumble on the same thing, add that instruction back. Don't guess what the model needs — you might predict wrong, and every instruction is read on every run. Experiment with stripping all prompts via `CLAUDE_CODE_SIMPLE=1` or `--system-prompt`. [^1]

- **Stop overspecifying. Give high-level tasks with guardrails and a verification mechanism.** Cherny says the most common mistake is step-by-step over-specification. Instead: describe the task, the guardrails, and the exit criteria, then let the model cook. The single most important thing people get wrong: giving the model a way to verify its own output. Cherny demonstrated this by having Claude rewrite the Electron desktop app in Swift, running both in a Mac VM with pixel-by-pixel comparison — still running after two weeks, spawning tens of thousands of agents. [^1]

- **Delete the chat window for ambient agents.** Harrison Chase and Austin Vance report the most impactful agent pattern removes the chatbot entirely: agents run on cron or event triggers. LangChain's Engine agent runs every six hours, examines traces, identifies issues, and proposes fixes — surfaced as a tab with an accept button, not auto-opened PRs. Their first version opened PRs directly and met resistance; redesigning to show issues with a preview button before opening a PR built trust and limited blast radius. [^2]

- **Prove feasibility with the expensive model, then ratchet down.** Vance's process: prove the task works with the smartest model (e.g., Opus), establish a benchmark, then use unit/integration/smoke evals to ratchet down to the best cost-reliability tradeoff per component. "Evals are the new PRD" — start with a small handful and evolve them as the agent grows. Use a smart main driver with cheaper sub-agents; model agnosticism is both defensive (avoid lock-in) and offensive (switch fast as models leapfrog). [^2]

- **Audit your CLAUDE.md for drift and conflicts.** Jason Zhou built an agent-context-audit skill that catches conflicting instructions across CLAUDE.md/skills/tools, finds hard rules that should be judgment, flags docs that drifted from the actual codebase, and quizzes docs to surface undocumented gotchas. Reports it cut his CLAUDE.md by ~50%. Repo: [github.com/AI-Builder-Club/skills](https://github.com/AI-Builder-Club/skills). [^3]

## 📡 WHAT SHIPPED

- **Kimi K3 weights released + in Cursor.** Moonshot's 2.8T parameter model (1.56TB on HuggingFace). Simon Willison notes the license went further than K2's "modified MIT" — it now requires a separate agreement with Moonshot for any "Model as a Service" business exceeding $20M revenue over 12 months. OpenRouter offers K3 from 7 providers at ~$3/M input, $15/M output. Cursor added K3 the same day, scoring close to the frontier on CursorBench with US-based inference via Fireworks, Together, and Baseten, plus zero data retention. [^4][^5]

- **Deep Agents v0.7.0b2.** LangChain's agent harness cut base input tokens by more than half through more minimal prompting, and added full configurability to override any default middleware. Changelog: [docs.langchain.com/oss/python/releases/changelog#deepagents-v0-7-0b2](https://docs.langchain.com/oss/python/releases/changelog#deepagents-v0-7-0b2). [^6][^7]

- **Cursor Start (India).** New ₹649/month plan with Grok 4.5 and Composer, plus autonomous cloud agents, Cursor for iOS, and extensibility via plugins, MCP servers, hooks, and skills. [^8][^9]

- **Opus 5 sentiment reverses across practitioners.** Theo walked back his initial enthusiasm: "I do not like Opus 5 as much as I hoped to." It "goes too far, makes dumb mistakes, treats every single comment as a p0 high severity issue that needs thousands of lines of code to fix." Salvatore Sanfilippo (Redis creator) reports that in a side-by-side test implementing a kernel for Strix Halo hardware, GPT-5.6 SOL significantly outperformed Opus 5 on low-level code, and Opus 5 communicates poorly — "answers like an idiot" — making it hard to understand what it did. ThePrimeagen spent $117.74 on 8 prompts (72M tokens) trying to get Opus 5 to draw a non-alien stick figure. [^10][^11][^12][^13]

- **$/token is dead; $/task is the metric.** swyx: "if you haven't updated your x axes to $/task then idk if you can be taken seriously anymore." ThePrimeagen's $117.74 for 8 prompts is a concrete example — most cost went to iterating on a single visual problem, not per-token economics. [^14][^13]

- **Agent-to-agent bug fixing in the wild.** Peter Steinberger's agent reported a bug in Bun (issue #36049); Jarred Sumner's "robobun" agent fixed it the same night. [^15]

- **Build-vs-buy caution.** Jediah Katz (Cursor): AI makes building easier, but consider maintenance burden, reliability requirements, rate of domain change, and how tailored your needs are before building. A team reverted to Linear after their vibecoded internal tool consumed more bandwidth maintaining it than it saved. [^16]

- **Claude Design praise.** McKay Wrigley calls it "by far the single most underrated AI product," saying using it with Opus 5 "completely transformed how I build." Plans a multi-hour tutorial on his AI coding + design workflow. [^17]

## 🎬 GO DEEPER

- **Boris Cherny: Building Claude Code** — The full talk behind the ablation methodology, dynamic workflows (Bun rewritten from Zig to Rust in 11 days via a single dynamic-workflow prompt with steering, now in production), and Claude maintaining its own codebase with 20-30 daily routines (dead code cleanup, "abstraction police," shipping experiments, writing/deleting tests). Trigger dynamic workflows by saying "use a workflow"; set up loops (local cron) and routines (cloud cron) for repetitive maintenance. [^1]


[![Boris Cherny: Building Claude Code](https://img.youtube.com/vi/qyPCVqFUyDo/hqdefault.jpg)](https://youtube.com/watch?v=qyPCVqFUyDo&t=357)
*Boris Cherny: Building Claude Code (5:57)*


- **Harrison Chase & Austin Vance: Agents in the Enterprise** — Deep Agents vs LangGraph ("startups love DeepAgents; regulated bodies hate it"), ambient agent UX, evals as the new PRD, and model ratcheting. Also: avoid splitting into sub-agents too early — a single agent goes further than you think; use sub-agents for context isolation when complexity warrants it. [^2]


[![Agents in the Enterprise: Deep Agents, Evals, and Ambient UX | Austin Vance and Harrison Chase](https://img.youtube.com/vi/E3ZWXvbjfXI/hqdefault.jpg)](https://youtube.com/watch?v=E3ZWXvbjfXI&t=662)
*Agents in the Enterprise: Deep Agents, Evals, and Ambient UX | Austin Vance and Harrison Chase (11:02)*


- **Logan Kilpatrick (Google AI Studio)** — Invokes Karpathy: "you can outsource intelligence, but you can't outsource your understanding." Current models prioritize generation over comprehension, creating an opportunity for agents that help developers retain understanding. Says he'd be "really excited to go build a coding startup right now" — the TAM expansion of software creation is enormous, and model labs can't capture all customers. [^18]


[![Inside OpenAI, Google Gemini & the Next AI Gold Rush With Google DeepMind's Logan Kilpatrick](https://img.youtube.com/vi/LiuCT4lsLmI/hqdefault.jpg)](https://youtube.com/watch?v=LiuCT4lsLmI&t=2202)
*Inside OpenAI, Google Gemini & the Next AI Gold Rush With Google DeepMind's Logan Kilpatrick (36:42)*


- **Salvatore Sanfilippo (Italian-language video)** — Firsthand model comparisons: GPT-5.6 SOL enables hours-long autonomous coding sessions without steering; min-p sampling (cutoff 0.05) eliminates Laguna model instability even in Q2/Q3 quantization; Laguna S2.1 hits 70 tokens/s on M5 Max with speculative decoding. Also: flow state is broken during long agent runs — use waiting time for secondary activities like reading or manual coding exercises. [^12]

*Editorial take: the day's strongest signal is that less prompting beats more — Cherny's ablation methodology, LangChain's Deep Agents token cuts, and Zhou's CLAUDE.md audit all converge on the same conclusion: modern models need fewer instructions, not more.*

---

### Sources

[^1]: [Boris Cherny: Building Claude Code](https://www.youtube.com/watch?v=qyPCVqFUyDo)
[^2]: [Agents in the Enterprise: Deep Agents, Evals, and Ambient UX | Austin Vance and Harrison Chase](https://www.youtube.com/watch?v=E3ZWXvbjfXI)
[^3]: [𝕏 post by @jasonzhou1993](https://x.com/jasonzhou1993/status/2081689174877380668)
[^4]: [moonshotai/Kimi-K3](https://simonwillison.net/2026/Jul/27/kimi-k3)
[^5]: [𝕏 post by @cursor_ai](https://x.com/cursor_ai/status/2081848014444876166)
[^6]: [𝕏 post by @sydneyrunkle](https://x.com/sydneyrunkle/status/2081718464985571577)
[^7]: [𝕏 post by @LangChain](https://x.com/LangChain/status/2081719358376616128)
[^8]: [𝕏 post by @cursor_ai](https://x.com/cursor_ai/status/2081978255004053560)
[^9]: [𝕏 post by @cursor_ai](https://x.com/cursor_ai/status/2081978257034063903)
[^10]: [𝕏 post by @theo](https://x.com/theo/status/2081880182936502474)
[^11]: [𝕏 post by @theo](https://x.com/theo/status/2081880547559907675)
[^12]: [Considerazioni durante un viaggio di ritorno da Portopalo a Catania](https://www.youtube.com/watch?v=g_DCfGiDUuc)
[^13]: [𝕏 post by @ThePrimeagen](https://x.com/ThePrimeagen/status/2081928510814200095)
[^14]: [𝕏 post by @swyx](https://x.com/swyx/status/2081904230768816487)
[^15]: [𝕏 post by @steipete](https://x.com/steipete/status/2081767828278170002)
[^16]: [𝕏 post by @jediahkatz](https://x.com/jediahkatz/status/2081839548347076897)
[^17]: [𝕏 post by @mckaywrigley](https://x.com/mckaywrigley/status/2081759923139178810)
[^18]: [Inside OpenAI, Google Gemini & the Next AI Gold Rush With Google DeepMind's Logan Kilpatrick](https://www.youtube.com/watch?v=LiuCT4lsLmI)