harness-kit

A better base for working with your AI agent. Clone it once and it becomes yours — the agent runs the structure, you just talk and work. A way of working, refined over years, made cloneable and de-anonymized: a canon of hard rules, self-learning, knowledge that survives sessions, and a concierge stance — built in from the first minute.

What it is, in plain language → https://share.s1lash.com/harness-kit

• Solo, cross-agent (Claude Code / Codex / Cursor), domain-agnostic.
• Fork-and-own: no upstream, no updates to chase. From the clone on, it's yours to grow.
• The structure is for the agent. You never file, sort, or organize anything.

---

Quickstart (for a person)

1. Clone it into a folder you name — pick a place and a name you'll recognize; you'll launch your AI agent from there.

   git clone <this-repo-url> my-base    # name it whatever you like

2. Tell your agent: "install this." It asks a few plain questions — where, your language, which agents you use, whether to use git — and wires everything up. Then you just work.

That's the whole onboarding. You don't need to understand the structure below — the agent does.

---

For the agent receiving "install this"

You have been handed this repository by a person who said "install" (or similar). Do the setup for them — do not make them learn the structure. Steps:

1. Read AGENTS.md (and CLAUDE.md if you are Claude Code) — that is your operating contract. The canon in rules/ is your standard and minimum quality bar; load it.
2. Run the installer for the person's platform — ./install.sh (macOS / Linux / Git-Bash) or install.ps1 (Windows) — OR perform its steps conversationally if a shell is unavailable. The installer:
   • asks where to place the base and what to name it (this clone is harness/; a sibling projects/ is created beside it for their apps);
   • asks the person's language and records it in profile.md — from then on you converse with them in that language, while all base content, code, and commits stay in English;
   • asks which agents they use and wires this canon into each one's global entry point (Claude ~/.claude/, Codex, Cursor) so it is hot from any folder;
   • asks about git (default yes; a remote is a separate opt-in) — if declined, touch nothing git until asked;
   • runs a health check (/harness:doctor).
3. Confirm in plain language what you set up. Then work normally — you now own this structure: route facts to their homes, keep it healthy, grow it as the person works. Never push the structural burden back onto them.

New session later? CLAUDE.md / AGENTS.md load the canon automatically; consult knowledge/_index.md on demand and activities/_index.md only on narrow "we did / last time / continue" signals — never load history by default.

---

What's inside

Path	What it is
rules/	The canon — 11 hard-rule files, loaded hot every session. Your standard.
doctrine/	On-demand authoring meta — deep-knowledge, edit-checklist, skill-creation gate, knowledge/activities discipline, tool-vs-instrument.
knowledge/	Durable understanding. Ships empty with the routing & growth discipline.
activities/	Work that survives sessions. Agent-maintained index + strict anti-bias rule.
tools/	Executable automation (scripts, CLIs, the MCP wrapper). Ships empty.
plugins/frontend-crafter/	Bundled: an anti-slop frontend-design skill, ready to use.
profile.md	The one personal file — who you are, grown carefully by the agent. Not a habit tracker.
install.sh · install.ps1	The conversational, cross-platform installer.
AGENTS.md · CLAUDE.md	Cross-agent and Claude entry points.

---

License

See LICENSE. Bundled plugins/frontend-crafter/ carries its own license and attribution.