Claude Code删掉了80%系统提示词！ 而实际上 Maka 早就已经提到了过了这点。

https://github.com/maka-agent/maka-agent…

这个结论我们几周前在 Maka 上跑 benchmark 的时候，已经独立撞见过一次。

用 Kimi K3 头对头测 Terminal Bench 2.1，Maka 对官方 Kimi Code CLI：Maka 用的是精简 工具面 + system prompt，Kimi Code 官方是 20KB 产品 prompt + 全量工具面。结果 Maka 领先 10 个百分点，难题子集领先 20 个百分点。 在 Anthropic 从模型厂的角度给出了同一个结论，而且是反过来的：不是“prompt 减了模型也不差“，是“模型变强了之后，冗余的 prompt 反而在拖后腿“。

Anthropic 这篇文章里最值得记的一句话，是把上下文工程从“写规则“重新定义成“搭架构“：与其用一堆硬性规则和示例去框死模型的探索空间（甚至规则互相矛盾，模型还要先内耗权衡冲突），不如把接口设计得足够清晰。

这跟我们做 Maka 时的体会完全一致：模型越强，越应该做减法。不是所有的“guardrail“都是保护，很多时候是噪音。

maka-agent/maka-agent

Source: https://github.com/maka-agent/maka-agent

Maka

A local-first Agent workspace built for real work.

Maka does more than answer questions. With controlled permissions, it can inspect projects, execute tools, produce artifacts, and preserve model messages, tool calls, and durable-task progress as recoverable execution facts. The same Runtime is available through the desktop app, terminal TUI, non-interactive CLI, and Headless runner.

Maka is under active development and currently targets users running from source or contributing to the project. Data formats, CLI commands, and experimental capabilities may still change.

Why Maka

• Local-first instead of hosted-first: sessions, settings, and run records stay on your machine by default. You choose the model connection: cloud API, local model, or compatible gateway.
• Log is the Runtime: model messages, Tool Calls, Tool Results, and termination facts enter Runtime Event Log. Sessions, UI, model context, and recovery are projections over that log.
• Context is not history: Tool Result pruning and LLM Compaction change what the next inference sees without treating recorded evidence as disposable context.
• A task may outlive a Turn: Headless uses TaskRun, Task Event Log, budgets, and continuation to advance interruptible and inspectable durable work.
• Feedback is not fact authority: Self-check may produce evidence and one bounded repair opportunity, but “I checked it” does not become a system fact.

Read Maka Backend Architecture for the complete design.

Surfaces

Current capabilities

Agent Runtime

• Multiple model connections, streaming output, thinking, usage accounting, and provider-error normalization;
• Local tools including Read, Write, Edit, Bash, Glob, and Grep;
• Tool schema validation, dynamic availability, permission policy, watchdogs, abort, and error classification;
• Runtime Event Log, AgentRun ledger, startup recovery, Turn Evidence, active Tool Result pruning, and history compaction.

Desktop workspace

• Create, archive, search, rename, retry, regenerate, and branch sessions from a Turn;
• Artifact lists and previews, workspace instructions, model settings, and permission settings;
• Local memory, web search, an open HTTP/SSE gateway, bot entry points, and Office workflows;
• Integrations are configured independently, and not every experimental entry is available by default.

Durable tasks and evolution

• Append-only Task Event Log and TaskRun projection;
• Budgets, permission pauses, continuation, result export, and failed-task retry;
• Plan-first, source-guarded, and attempt-bounded Heavy-task Self-check;
• AHE target protocol and evidence export; complete automatic self-iteration remains an external or experimental workflow.

Quick start

Requirements

• Node.js 22.19 or newer (CI uses Node.js 24);
• npm (the lockfile and scripts use npm; the current packageManager is npm 11);
• Git;
• ripgrep, used by Runtime’s Grep tool.

Start Desktop

git clone https://github.com/Maka-Agent/maka-agent.git
cd maka-agent
npm ci
npm run dev

npm run dev starts the Desktop development environment with HMR. To build every workspace before starting Electron, use:

npm run dev:full

If dependencies were installed with ELECTRON_SKIP_BINARY_DOWNLOAD=1, install the Electron platform binary before starting:

node node_modules/electron/install.js

First run

Maka does not bundle a shared model account. On first launch:

1. Open Settings → Models;
2. Add an API, local-model, or supported account connection;
3. Test it and choose a default model;
4. Return to the workspace and start a task.

The app distinguishes configured, send-ready, and experimental connection states. An account flow that is not wired into Runtime is not presented as a usable model.

Terminal entry points

Build the workspaces first:

npm run build

Then start the TUI or run one Turn:

npm --workspace maka-agent exec -- maka
npm --workspace maka-agent exec -- maka run "Summarize this repository and identify its most important risk"
npm --workspace maka-agent exec -- maka --help

The CLI reads the same model connections and workspace configuration written by Desktop. See packages/headless/README.md for Headless commands and its trust posture.

Architecture

The backend spine is:

Desktop / TUI / Headless
          ↓
SessionManager → AgentRun → Model + Tool Runtime
          ↓
Runtime Event Log → Context / Session / UI projections
          ↓
Task Event Log → TaskRun → Self-check / AHE evidence

Start with ARCHITECTURE.md. It provides the system map, code boundaries, problem-oriented reading paths, and six bilingual deep dives.

Repository layout

apps/desktop/       Electron main / preload / React renderer

packages/core/      Pure contracts for Sessions, Events, Permissions, and Connections
packages/storage/   File-backed stores and run ledgers
packages/runtime/   AgentRun, model adapters, tools, context, and recovery
packages/headless/  TaskRun, Autonomous Loop, Self-check, eval, and AHE
packages/cli/       TUI and non-interactive CLI
packages/ui/        Shared conversation, Markdown, Artifact, and UI primitives

docs/               Architecture, product, security, privacy, and test contracts
scripts/            Build hygiene, visual checks, smoke tests, and release helpers

Local data and security boundary

Maka stores workspace data under Electron userData by default:

<Electron userData>/workspaces/default/
  llm-connections.json
  credentials.json
  settings.json
  sessions/

Current boundaries that matter:

• Sessions and connection metadata live in the local filesystem;
• Runtime credentials such as API keys, bot tokens, and proxy passwords currently live in local plaintext credentials.json, behind the OS account boundary, with POSIX directory mode 0700 and file mode 0600 enforced;
• Subscription OAuth tokens (Claude, Codex, GitHub Copilot, and the Cursor/Antigravity previews) live in the same credentials.json — the single authority for desktop, TUI, and headless; Electron safeStorage only decrypts pre-existing legacy token files once at desktop startup (#1125);
• Renderer does not receive plaintext credentials. File writes, Shell, and dangerous tool calls pass through the permission engine;
• Headless real-model evaluation fails closed by default and requires an explicit external isolation boundary.

Read SECURITY.md for security reporting and policy, and docs/README.md for current privacy and sandbox contracts.

Experimental runtime recovery flags

Runtime recovery remains opt-in. Both flags below are disabled by default:

• MAKA_RUNTIME_SQLITE_CANONICAL=1 migrates the current workspace’s canonical RuntimeEvent store to runtime.sqlite. This is a one-way, sticky migration trigger, not a reversible backend selector: after runtime.sqlite exists, disabling the variable does not switch the workspace back to JSONL. Automatic pre-migration backup and populated v2-to-v4 upgrade coverage are not complete, so back up the workspace before enabling this flag.
• MAKA_RUNTIME_SAFE_BOUNDARY_RESUME=1 enables the Desktop interrupted-turn Safe resume action, CLI/TUI /resume, and Desktop startup auto-resume. These paths may call the configured model provider and consume tokens. Enable the flag only when that behavior is explicitly desired.

Phase 2 provides the durable write-side boundary and fail-closed safe-boundary continuation. Phase 3 reconciliation for indeterminate tool side effects is not implemented yet; ambiguous tool outcomes remain parked rather than retried.

Development and verification

Common repository-level commands:

npm run build
npm run typecheck
npm test
npm run check:release

Run one workspace in isolation:

npm --workspace @maka/runtime test
npm --workspace @maka/headless test
npm --workspace @maka/desktop test

Use the following commands to update packages/core/src/model-metadata.generated.ts from models.dev and run the focused tests. Keep access-path-specific overrides in model-metadata.ts; do not edit the generated file by hand.

npm run sync:model-metadata
npm run test:scripts
npm --workspace @maka/core test

Desktop real-window and visual verification:

npm --workspace @maka/desktop run e2e
npm --workspace @maka/desktop run smoke:real-window

Before submitting code, run typecheck, build, and focused tests proportionate to the change, followed by git diff --check.

Documentation

• Documentation index and authority map
• Backend architecture
• Product design
• Security policy

License

Maka is licensed under the Apache License 2.0. See NOTICE for attribution information. Third-party components remain subject to their respective licenses and notices.

Thariq (@trq212): We removed ~80% of the Claude Code system prompt for our newest models, this is what we’ve learned about writing system prompts, skills and Claude.MDs for them.