Claude Code cuts system prompt by more than 80% without hurting coding evals

• Claude Code's prompt diet was real: trq212's post said Anthropic removed roughly 80% of the system prompt, and Anthropic's own post says Opus 5 and Fable 5 saw no measurable coding-eval loss.
• Anthropic's new context playbook is a six-part swap from rules and examples to judgement, interfaces, progressive disclosure, simple tool descriptions, auto-memory, and rich references, as rohanpaul_ai's breakdown pulled from the official post.
• Durable context is moving into wikis, folders, session IDs, and project files: pauliusztin_'s wiki-memory thread, rauchg's filesystem workflow, and kevinkern's Codex session-ID note all describe context that survives the chat turn.
• Two research threads bracket the memory design space: rohanpaul_ai's ACM summary emphasizes lifecycle-managed compaction, while dair_ai's PRO-LONG summary emphasizes a complete searchable log.
• Multi-agent work is turning context into coordination state, with Squid's staged software factory in pauliusztin_'s architecture thread, Agent Mail's cross-agent inboxes in doodlestein's Agent Mail post, and Offloop's persistent Channels in testingcatalog's Offloop note.

Anthropic's official Claude post includes the new /doctor command for right-sizing Skills and CLAUDE.md files. The Agentic Context Management paper turns the same pressure into five primitives. Google's Open Knowledge Format announcement formalizes the LLM-wiki pattern as Markdown plus YAML frontmatter. PRO-LONG takes the deliberately boring route: keep the full log and search it with code.

Anthropic says it removed more than 80% of Claude Code's system prompt for Claude Opus 5 and Claude Fable 5 with no measurable loss on its coding evaluations. The company framed the deleted text as accumulated constraints that newer models no longer needed.

The official example is comments. Anthropic's old prompt told Claude to avoid comments, multi-paragraph docstrings, and planning documents; the new prompt says to write code that matches the surrounding code's comment density, naming, and idiom.

trq212 added one operational detail in a reply: Claude Code now uses different system prompts per model. That explains why the 80% cut is about the newest models, not a universal deletion across every Claude Code runtime.

Anthropic's post turns the old playbook into six replacements:

1. Rules → judgement: fewer rigid guardrails, more surrounding-context inference.
2. Examples → interfaces: enum values and schemas carry usage constraints without long demonstrations.
3. All upfront → progressive disclosure: code review and verification moved into Skills that load when needed.
4. Repetition → simple tool descriptions: tool-use instructions live with the tool, not repeated in the system prompt.
5. Memory in CLAUDE.md → auto-memory: Claude saves work-relevant memories automatically.
6. Simple specs → rich references: test suites, HTML artifacts, rubrics, and code can be higher-fidelity specs than prose.

jerryjliu0 described the same failure mode from the user side: extensive edge-case prompts can make the model satisfy the written constraints while contributing fewer new perspectives on the problem in jerryjliu0's reaction.

Wiki memory is the cleanest version of the new pattern. The agent does not repeatedly retrieve raw documents, reconstruct understanding, and discard it after the session; it maintains a compact, inspectable layer between the raw sources and the harness.

The shape in pauliusztin_'s thread is simple:

• raw notes, docs, code, PDFs, and Slack remain the source of truth;
• the wiki becomes the compressed representation the agent reasons over;
• the first conversation creates the wiki;
• later conversations refine it.

That maps to his personal split between an immutable second brain and a project-specific LLM wiki. The second brain is research; the project wiki is the agent's workspace.

His retrieval order is also explicit: start with summaries, expand into concepts, follow references, and read raw sources last. That is progressive disclosure as a file layout.

rauchg's setup uses a research/ folder plus an AGENTS.md file as the interface. Prior sessions become markdown files, scripts sit beside them, and the agent can correlate prior work without a dedicated knowledge app.

Codex exposes a related escape hatch. kevinkern said the copy-session-ID feature lets a user ask Codex to read what was discussed in ~/.codex/session for a specific session ID, and he later noted an outage reset as another reason session continuity matters in kevinkern's session-ID post and his outage follow-up.

A leaked Claude Code diff claim pushed the same idea into always-on project memory. According to imjustnewatai's report, Claude Code 2.1.219 added an unannounced, feature-gated metadata.pinned: true path that can inject up to four pinned files into future project conversations, capped at 200 lines or 25 KB each.

The ACM paper names five jobs for production context systems:

• architecting;
• ingesting;
• scoping;
• anticipating;
• compacting with consolidation.

The paper's economic claim is sharper than the taxonomy. Sending all prior context makes token use grow roughly with conversation length squared, while validated compaction keeps growth closer to linear without the accuracy cliff of crude summaries, according to rohanpaul_ai's summary and the arXiv paper.

PRO-LONG goes lossless. It stores every observation, action, result, and short plan in a structured log, then lets the coding agent search that log with ordinary code and regex.

The reported numbers are the hook: average ARC-AGI-3 gains of 18.0 percentage points over base coding agents, up to 76.1% pass@1, and 4.2x to 5.8x fewer tokens, per dair_ai's PRO-LONG summary.

Squid is the coding-agent version of context as workflow state. Paulius Ztin rebuilt it from one giant automation into three stages:

1. Planning: grill the feature idea for missing requirements, weak assumptions, architectural gaps, and edge cases; output a spec, ADRs, glossary, and task list.
2. Implementation: split the work between a Software Engineer role and a Tester role that tries to break the code.
3. Review: review the pull request diff, then run CI, formatting, linting, integration tests, and regression checks.

The punchline is not the loop. The durable assets around the loop, specs, ADRs, tasks, diffs, CI results, are the context the next agent turn can use.

steipete's Codex QA run is the industrial version: 12 subagents, live API keys, multiple dev gateways, worktrees, autonomous PRs, and a goal of finding 200 bugs. The prompt in his full QA instruction told agents to keep a markdown test report updated on the desktop.

Agent Mail adds cross-agent coordination to that picture. doodlestein's post showed Fable 5 and Sol Max collaborating through MCP Agent Mail, with threaded messages and file reservations visible in the attached terminal UI.

Prompt caching turns context layout into a cost surface. Pi's post says caching affects latency, cost, tool design, and product feature decisions, not just provider plumbing.

The failure mode is model routing. A cheaper model can cost more when switching mid-session breaks the cache; pvncher described that penalty in his model-router warning, and omarsar0 described a cache-aware router that only switches when the move saves money in his cache-aware routing post.

Tooling is starting to expose the hidden bill. jlongster added an OpenCode debug view that renders token usage for each step, including new, cached, and total tokens.

Supermemory packaged the pattern as composable context blocks: user profiles, a memory graph, hybrid retrieval, extractors, connectors, and SMFS, described as memory-as-a-filesystem for agents.

Genspark 6.0 moved the same idea into an enterprise workspace. Genspark's Workspace 6.0 announcement described SecondBrain as a memory layer fed by email, meeting notes, IMs, docs, apps, and projects; its Microsoft partnership post said that memory now appears inside Outlook, Office, Teams, and Agent 365.

Offloop's unit is the Channel: context, files, schedules, approvals, and connectors live together so recurring work can continue. Its D1 dispatcher decides which agent acts at each step, according to testingcatalog's follow-up.

OpenWorker took the local-first route. AndrewYNg's OpenWorker announcement said the Mac app can draft documents, send Slack messages, update calendars, use local files and everyday tools, and run with user-provided model keys or Ollama.

Codex added a smaller but practical context primitive: one local project can include multiple folders, while the primary folder remains the Git root for Git operations, reviews, pull requests, AGENTS.md, and skill discovery in OpenAIDevs' multifolder Codex post.