Guided harness builder · 17 assistants · strict by design
You build the project. rsc-harness builds the harness.
You decide what to build — software, operations, research, content, or a mix. The wizard turns that goal into a minimal, verified setup without improvising a Frankenstein harness.
No harness expertise. No API key. Free & open source (MIT).
Paste into Claude Code, Codex, Cursor, Gemini or another supported assistant:
The problem is the harness, not your goal
A powerful model can still end up inside a Frankenstein.
An agent needs more than a brain: project memory, tools, knowledge, rules and repeatable workflows. Add them ad hoc and you get duplicated instructions, unrelated skills, premature hooks, unnecessary MCPs and no clear source of truth.
Open standards define individual parts — Agent Skills, MCP and AGENTS.md. What they do not decide is which complete harness your project needs. rsc-harness applies that missing method for you.
One goal in, one coherent harness out
The wizard makes every decision visible.
State the outcome
Say what you want to build or run and how much guidance you want. No harness vocabulary required.
Review the exact plan
rsc reads only the project root and explains every selected, deferred and excluded component.
Accept, verify, evolve
Nothing is written before acceptance. The result is verified and grows later only when evidence justifies it.
Strict best practices
The opposite of “install everything and hope.”
Each rule is visible in the plan or enforced by a deterministic check.
Minimal first
Only what the current goal and project evidence justify. Knowledge loads progressively, never all at once.
Separated by purpose
Memory, tools, documentation, rules and specialists keep clear owners and one shared source of truth.
Bound to the project
Analysis stays inside the selected project root. Local state and memory stay private and local.
Consent is exact
You see the complete plan first. If it changes, the old acceptance no longer counts.
Verified, not declared
READY appears only after deterministic checks prove the installed harness matches the accepted plan.
Growth follows evidence
Deferred capabilities return only when observable changes make them useful, and still require a new plan.
For AI agents reading this page
Build the user’s harness through the binding wizard
Do not choose components or onboarding answers for the user.
Proof, not the promise
A catalog large enough to stay selective.
273 quality-gated skills across 23 domains give the wizard options. You still install only what the plan justifies.
Granular by default
The unit of installation is one skill. Install fastapi without ever pulling go. Nothing you don't use touches the model's context.
Self-recommending
An always-on detector watches your work and names the next skill the moment a task needs it — a one-word confirm installs it.
Quality-gated
Every skill was researched against live 2025–2026 sources and had to clear an adversarial ≥ 8.5/10 behavioral rubric before it shipped.
Not code-only
Beyond languages and frameworks, rsc can run a whole operation: bookkeeping, invoicing, GDPR, pitch decks, SEO, and a YouTube / TikTok / LinkedIn presence — each wired to a knowledge loop that learns from your own results.
One base, every IDE
Skill files are written once to your project and symlinked into each assistant you pick — no copy duplicated across IDEs.
Works with
17 assistants, one harness method
Use the assistant you already have. The same planning discipline travels with the project.
Get started
Start with what you want to build
Paste the prompt into your agent, or use the terminal if you prefer.