Titan

Titan is an autonomous agent pipeline for Claude Code and other AI agent hosts. It is a methodology rather than a service: there is no API key, no server, and no external dependency. The skill turns the host agent's own reasoning and subagents into a disciplined plan → execute → QA → gate pipeline.

The goal is simple: a faster or cheaper model running Titan's discipline — task decomposition, a minimalism ladder, reflexion-based retries, end-to-end verification, a five-judge review council, and an evidence-based definition of done — should produce output that rivals a top-tier model running at maximum effort.

Installation

In Claude Code, run:

/plugin marketplace add HDHNezherParking-cum-Y638-Intl-Ltd/titan
/plugin install titan@titan

That's it. Both skills are picked up automatically and will also trigger on their own when a task matches their descriptions.

Other agent hosts

Both skills are plain SKILL.md directories (the open Agent Skills format), so any host that reads that format works. Clone once, then copy the two skill folders into your host's skills directory:

git clone https://github.com/HDHNezherParking-cum-Y638-Intl-Ltd/titan

OpenAI Codex — personal (all projects) or project scope:

mkdir -p ~/.codex/skills && cp -r titan/skills/titan titan/skills/ops ~/.codex/skills/

Google Antigravity — workspace scope (use ~/.gemini/config/skills/ for global):

mkdir -p .agents/skills && cp -r titan/skills/titan titan/skills/ops .agents/skills/

GitHub Copilot — repo scope (.agents/skills/ as above also works); personal scope:

mkdir -p ~/.copilot/skills && cp -r titan/skills/titan titan/skills/ops ~/.copilot/skills/

Copilot users can also manage skills with the gh skill GitHub CLI command. On Windows, replace mkdir -p/cp -r with New-Item -ItemType Directory -Force/Copy-Item -Recurse.

Skills

How it works

When the titan skill activates, the agent classifies the task (code, plan, research, or text), breaks it into role-owned subtasks, and recalls lessons from previous runs. For code tasks it first climbs a seven-rung minimalism ladder — checking whether the need is already met by existing code, the standard library, or a one-liner before building anything new. High-stakes decisions get a small tree search over three candidate approaches.

Execution runs as ReAct loops with strict budgets. Failures are never blindly retried: the agent writes a short critique of what broke and why, stores it, and retries with that critique in context (the Reflexion pattern). Finished work is verified by actually running it — building, testing, exercising the affected flow — and then scored by a five-judge council covering correctness, completeness, security, quality, and efficiency. Only then is the result presented to the human, once, with any open decisions batched together.

Repository layout

.claude-plugin/          plugin.json and marketplace.json manifests
skills/titan/SKILL.md    pipeline overview (loaded when the skill triggers)
skills/titan/references/ per-stage depth: guard, ladder, tree search, council,
                         reflexion, loops, context, prompting, memory, and more
skills/titan/workflows/  playbooks for code, planning, and research/text tasks
skills/ops/SKILL.md      preflight / land / decide / commission

Reference files load on demand (progressive disclosure), so a run only pays the context cost of the stages it actually uses.

Optional CLI

Reference files write the recurring helper operations in a CLI shorthand (titan guard, titan compress, titan memory). No CLI is required: the skill defines inline equivalents for every operation and stores cross-session lessons in .titan/memory.md at your project root. Add .titan/ to your project's .gitignore.

Sources

The methodology is distilled from Reflexion (NeurIPS 2023), ReAct (ICLR 2023), LATS (ICML 2024), Generative Agents (2023), CoALA (2023), Anthropic's agent and context-engineering guidance, MetaGPT, Voyager, karpathy/llm-council, ponytail, and prompt-master (MIT). Each reference file carries its own attribution.

License

Apache-2.0 — see LICENSE.