Most Capable Agent System Prompt
A copy-paste system prompt that makes your coding agent build a durable, self-improving agent system.
- Source repo
- fainir/most-capable-agent-system-prompt
- Stars
- ★ 882
- Last updated
- 1d ago
- License
- MIT
- FA score
- 44/100 · Major gaps
At a glance
- Works with
- Universal · cross-platformChatGPT · Codex · Claude CodeClaude.ai · OpenAI API · Claude API (Partial support)
- You'll need
- Typical use
- Engineers already using Claude Code or Codex daily who want to move from one-off chat coding to a project-level workflow with task graphs, verification, and persistent memory.
- Main limitation
- About 2,950 lines and 111,000 characters (~23,000 tokens measured with tiktoken); as a system prompt or instructions file it is sent on every request, adding cost and context pressure.
- Source review
- 44/100 · Major gaps 1 safety controls not found
What does this agent do, and when should you use it?
The repository is a single plain-text system prompt of about 2,950 lines and roughly 111,000 characters, not a packaged application. Paste it as the first message into an agent that can read and write files and run shell commands — Claude Code, OpenAI Codex, Cursor, Antigravity, OpenCode, OpenHands, the Claude Agent SDK, and similar — and it inspects the workspace, asks a minimal batch of questions, writes a runtime capability matrix and an implementation contract, and scaffolds a file-based project operating system (AGENTS.md, plan.md, tasks.md, knowledge.md, FAILURE.md). The prompt requires the agent to prove one closed loop end to end — goal, task graph, execution, verification, memory update, visibility, learning — as its first milestone, then proceeds through a BUILD ORDER toward a control plane, workers, layered memory, model routing, approvals and budgets, an eval harness, and a self-improvement loop. It supports two implementation postures: harness-wrapper mode around an existing strong runtime, or native runtime mode built directly on an agent SDK. Project state is canonical in the folder; chat history is an optional surface.
The prompt opens with a READER CONTRACT that fixes the order in which the agent must read core sections (NON-NEGOTIABLE DESIGN BETS, RELIABILITY MATH AND HARNESS ENGINEERING, RECOMMENDED DEFAULT IMPLEMENTATION CHOICES, BUILD ORDER, FIRST MILESTONE DEFINITION, NON-NEGOTIABLE RULES, INITIAL ACTIONS YOU MUST TAKE NOW) and requires a compact local operating summary. Phase 0 is runtime discovery and alignment: the agent classifies the runtime, inventories capabilities and constraints, and produces an implementation contract with mission, first milestone, non-goals, safety posture, and verification strategy. Phase 1 builds a capability matrix (repo read/write, shell, git, network, package install, browser control, screenshots, tool calling, sub-agents, background execution, scheduling, webhooks, dashboards) and decides for each gap whether to emulate, integrate, defer, or narrow scope. It then writes the canonical project file pack, maintains five live queues (now, next, blocked, improve, recurring), and implements the closed-loop milestone before advancing through the BUILD ORDER toward a REST + WebSocket control plane, hub-and-worker topology, SQLite WAL queues, per-task worktrees, approval tiers, budgets, evals, and a one-change self-improvement loop with rollback.
- Engineers already using Claude Code or Codex daily who want to move from one-off chat coding to a project-level workflow with task graphs, verification, and persistent memory.
- Solo developers or small teams on multi-month projects who need continuity across model swaps and runtime changes, resumable from the project folder alone.
- Researchers or operations leads who want an agent to handle planning, research, finance, and operations work, not just code.
- Teams building enterprise automation that need approval gates, budgets, and audit trails around long-running agents.
- Agent developers who want a written architecture specification as the starting point for a self-improvement loop (single change, eval, revert).
How do you install or deploy this agent?
There is no install step: the repository is a text prompt. Copy the prompt from the code block under The Prompt in the README, or fetch it from a terminal on macOS:
curl -fsSL https://raw.githubusercontent.com/fainir/most-capable-agent-system-prompt/main/README.md | awk '/^text$/{f=1;next} f&&/^On Linux, replace | LANG=en_US.UTF-8 pbcopy with | xclip -selection clipboard or | wl-copy; to save it to a file instead, use > prompt.md. You need an agent that can read and write files and run shell commands (Claude Code, Codex, Cursor, Antigravity, OpenClaw, OpenCode, OpenHands, Claude Agent SDK, or similar) and a dedicated, git-tracked working folder.
How do you use this agent?
The recommended path is to paste the prompt as the first message of a fresh session. From a saved file:
claude "$(cat prompt.md)"codex "$(cat prompt.md)"As an appended system prompt in Claude Code:
claude --append-system-prompt-file prompt.mdDo not use --system-prompt-file, which replaces Claude Code's built-in instructions. If you save it as AGENTS.md, Codex reads only the first 32 KiB by default; raise the limit in ~/.codex/config.toml:
project_doc_max_bytes = 131072To pick an implementation posture, append a line after the prompt, for example:
Use harness-wrapper mode around Claude Code.Or build natively on the Claude Agent SDK:
Use native runtime mode on the Claude Agent SDK, in TypeScript.What are this agent's strengths and limitations?
- Encodes architecture, reliability, approvals, budgets, evals, and self-improvement as concrete build instructions rather than a vague one-liner.
- Filesystem-first project state (AGENTS.md, plan.md, tasks.md, knowledge.md, FAILURE.md) lets work survive model swaps and runtime changes.
- Explicitly prioritizes proving the full closed loop before expanding breadth, which discourages premature multi-agent complexity.
- Supports both harness-wrapper mode around an existing agent and native runtime mode on an SDK, while staying neutral across Claude Code, Codex, Cursor, and OpenHands.
- The README names a specific set of core sections to keep and sections that are safe to trim, making shortening practical.
- About 2,950 lines and 111,000 characters (~23,000 tokens measured with tiktoken); as a system prompt or instructions file it is sent on every request, adding cost and context pressure.
- The repository ships only the prompt and an architecture diagram — no runnable software, tests, or eval data — so real output depends entirely on the host model's long-horizon execution.
- It instructs the agent to write files and run commands, so you need a dedicated folder, git, and conservative approval settings to avoid accidents.
- The README itself warns of many files, long sessions, and real token spend, which is a meaningful cost for individuals.
How does this agent compare with similar options?
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Most Capable Agent System Prompt This agent | 44 · Major gaps | — | ★ 882 | 1d ago | — | ChatGPT · Codex · Claude Code |
| CoCo Super Intelligence | 69 · Some gaps | Agent plugin / skillFree + model costs | ★ 797 | 3d ago | Python | Codex · Claude Code |
| Loop Library: Reusable AI Agent Workflows | 61 · Some gaps | Agent plugin / skillFree | ★ 3.2k | 1mo ago | JavaScript | Codex · Claude Code |
| HERO — Anti-OverDefense | Insufficient evidence | Agent plugin / skillFree | ★ 769 | 3d ago | Python | Codex · Claude Code |
How does FollowAgents rate this agent?
Why each dimension lost points
The prompt instructs sandboxed execution, approval for sensitive writes, idempotency keys and compensating/rollback paths (items 12, 13, 16, 17), so user_confirmation, external_effects and rollback score 2. Least privilege stays at principle level with no concrete permission list or default-deny policy, so 1. Data-flow transparency is only a slogan about a 'transparent state model' with no data-destination detail, so 1. Sensitive data handling has no dedicated section, so 1. No dependency manifest or supply-chain note exists, so dependency_security is 0. Source attribution is only a mention of a research appendix, not visible in the provided evidence, so 1.
The prompt is internally coherent, repeatedly stressing closed loops, verification and checkpoints, so self_consistency is 2. On dependency availability, the README claims ~108 KiB and ~2950 lines but the supplied file is truncated, so completeness and the stated scale cannot be confirmed, and no dependencies are declared, so 1. Failure handling has concepts like FAILURE.md and dead-letter queues but no concrete error-message spec, so 1.
Audience and scenarios are broad (coding, research, operations, etc.) with harness-wrapper and native modes, so 2. Capability boundaries include governance language on when not to act, ask or escalate, so 2. Trigger precision is weak: it is a long one-shot paste with no explicit trigger conditions or invocation contract, so 1. Environment fit lists multiple hosts and gives config advice (e.g. Codex project_doc_max_bytes), so 2.
Information architecture is clear with a table of contents, sections and a READER CONTRACT, so 2. Install notes are thorough with multiple usage modes and terminal commands, so 2. Naming stability is partial: it asks the agent to generate AGENTS.md etc., but the prompt itself has no version identifier, so 1. Examples and FAQ are substantial, so 2. Known limitations are stated ('it is long', 'run it somewhere safe'), so 2. License is MIT, so 2. No changelog, so versioning_changelog is 0. Maintenance responsibility is only a copyright line with an unverified publisher, so 1.
Output usability: the prompt demands files and a system, but real results depend heavily on model and runtime, and the provided evidence is truncated, so 1. Marginal value: it is an architecture blueprint overlapping the existing agent-harness, so 1. Cost-benefit: the README itself cites ~23000 tokens plus long sessions and real spend, high cost with uncertain payoff, so 1.
claim_traceability: the README's size and token claims cannot be checked against the truncated file, so 1. cross_source_corroboration: only a single README and LICENSE, no independent corroboration, so 1. fact_inference_separation: the prompt offers limited separation between asserted facts and inference, so 1.
- Not found in source: dependency securityPin versions and run a dependency audit (npm audit, pip-audit) before installing; prefer running it in a container.
- The prompt has the agent write files and run commands; run it in a dedicated git folder with normal approval settings and avoid unattended modes on machines with sensitive credentials.
- The supplied README is truncated, so the claimed ~2950 lines and ~23000 tokens cannot be verified; completeness and behavior are unconfirmed.
- Publisher identity is unverified and maintenance/update path is unclear; review the full prompt yourself before use.
- No dependency manifest or supply-chain note exists; if the agent pulls in third-party packages, assess their security independently.