Agentic Vault Kit

A personal-growth vault where the AI grows with you. Clone it, open it in your agent app, and you have a Markdown vault with a resident agent that learns alongside you — and evolves itself from your feedback. No terminal skills required.

You bought one (or several) coding-agent subscriptions — Claude Code, Codex, Kimi Code… — and you mostly use them through their desktop apps. You take notes in Obsidian or plain Markdown. This kit connects the two, but the point isn't note management. It's co-evolution: the vault is where you grow (learn faster with an AI study partner, push projects, build judgment), and where the agent grows too — every correction you give it becomes a rule, every session it scans its own performance, prunes what didn't work, and gets measurably better at being your partner. The longer you use it, the less generic it is.

Built on Codex first. This system was created and is maintained with Codex day to day — AGENTS.md project instructions, Agent Skills, and Git-checkpointed workflows are exactly the primitives it leans on — and it runs equally well with Claude Code, which we use in rotation when quotas alternate. Everything here is extracted from a real private vault maintained daily by both, across two machines, including a complete study system. Only personal identity data was removed; the working machinery is intact.

What you get out of the box

• A resident partner, not a Q&A tool. AGENTS.md gives any agent (any vendor) a standing identity: assetize every useful output to disk, fix problems at the root when you complain, sense your mood before pushing the plan.
• Self-evolution built in. The evolve skill scans each session for corrections, misfired skills, and new tactics, then writes improvements back into the rules — with a pruning discipline so the system converges instead of bloating. Your feedback literally rewrites your agent.
• An AI co-learning system, not just storage. The study domain shows the pattern: the agent diagnoses how you got things wrong, builds spaced-review assets from your actual confusion patterns, and adapts its teaching method per subject — fast personal growth with the AI in the loop.
• Battle-tested skills, pre-installed for both Codex (.agents/skills/) and Claude Code (.claude/skills/): design-before-build brainstorming, evidence-before-claims verification, conversation closeout, self-evolution, model routing, quota handoff.
• Multi-agent / multi-device discipline: checkpoint commits, WIP markers, handoff and interrupted-work recovery — so two agents and two machines don't trample each other.
• Templates for decisions, reviews, project pages, daily notes, and AI-log summaries.

Quick start (GUI, zero setup)

See SETUP.md for the click-by-click version. The short form:

1. Click Use this template (or Download ZIP) on GitHub — you get your own private copy.
2. Open the folder in Obsidian (optional but nice) and in your agent app (Claude Code desktop, Codex app, or any agent that reads AGENTS.md).
3. Say: "Read AGENTS.md and take over as the resident maintainer of this vault."
4. Start talking — take notes, study, plan. The agent files everything where it belongs and commits as it goes.

What's inside

Path	What it is
AGENTS.md	The behavioral contract: identity, always-on obligations, loading layers, routing
SETUP.md	Zero-terminal, click-by-click onboarding
.agents/skills/ + .claude/skills/	Portable Agent Skills, pre-installed for both Codex and Claude Code (mirrored): brainstorm, evolve, verify-claims, dialogue-closeout, model-router, quota-handoff
docs/	Field notes: running several agents on one vault without them trampling each other
domains/study/	Example domain contract: a study system with spaced review and "generation ≠ completion"
templates/	Decision record, review, project page, daily note, AI-log summary
examples/	Fictional worked examples of the feedback loop and decision records
SECURITY.md / PRIVACY.md	Hard data boundaries the agent enforces

Design principles

1. Plain Markdown + Git is the database. Switch agent vendors anytime; your system survives.
2. Rules live in layers — global contract → domain contracts → on-demand runbooks → state files — so agents load only what the task needs.
3. Generation ≠ completion. The agent never marks your learning or decisions done for you.
4. Evidence before claims. "Should be fine" is banned.
5. The user is not the only QA. The agent self-checks, cross-validates files, and fixes inconsistencies before you notice them.

Status & roadmap

v0.1 — core contract, four skills, study-domain example, templates. Planned: one-click bootstrap script, more domain packs (project ops, reading workflow), plugin packaging. See issues.

License

MIT. skills/verify-claims and skills/brainstorm are adapted from obra/superpowers (MIT); the evolution loop adapts ideas from claude-evolve, Self-Evolving-Skill, and reflect-system (see file headers).