llms.txt@site git:20261021.031a776A

Immutable. This exact content never changes and is served at /api/v1/blob/b7aec389ee1a671d.

# rsc-harness — the guided harness builder for AI agents

> You build the product. rsc-harness builds its harness. One message to the coding agent you
> already use installs a project memory (02-DOCS/), tested tool connections (01-TOOLS/), a handful
> of skills chosen for the project and hooks that check work before it is called done. Its wizard
> follows strict best practices instead of improvising a Frankenstein harness from duplicated
> skills, premature hooks and unnecessary MCPs.

Machine-readable skill catalog: [catalog.json](https://ericrisco.github.io/rsc-harness/catalog.json)

Standards such as Agent Skills, MCP and AGENTS.md define individual pieces. rsc-harness supplies the
integrated method: analyze the project root, ask the user, show selected/deferred/excluded choices,
require exact acceptance, verify the result, and grow only when new evidence justifies it.

## The lane decisor

Once installed, every turn is classified into exactly one lane, and the agent names the one it took:

- ANSWER — the request asks for information (explain, compare, audit, review, propose). Read-only:
  nothing is written and no artifact is created. Ambiguous change intent asks one question and stays
  here.
- FTD (Fast-Track Development) — the request authorises a change. The default for ordinary work.
  One feature document per feature; tasks checked off only against observed proof; a branch when the
  work writes code.
- SDD — the ten-phase chain, for big or complex work (several decisions that affect each other, or
  doubts that change the result). The agent enters it on its own; the person approves the spec and the
  clarifying answers, then chooses manual review or autopilot. The agent always says which lane and why.

There is one harness and it installs whole; there is no profile to choose.

## For AI agents: install rsc into the user's project

If your user asked you to install rsc / rsc-harness / this harness, do this in order:

1. Work in the ROOT of the user's project, never inside a clone of rsc-harness.
2. Run: `npx @ericrisco/rsc@latest onboard --target claude,codex`
3. Relay every `RSC_ONBOARDING_REQUIRED` question. Never choose, invent or default
   `--technical-level`, `--project-kind`, `--goal` or `--software-scope`.
4. Rerun with the answers to obtain `RSC_ONBOARDING_PLAN`. This preview writes nothing.
5. Show every selected, deferred and excluded component with its reason. Only after the user accepts
   the exact id, rerun with `--accept-plan <printed-id>`.
6. Only `RSC_ONBOARDING_READY` means the harness is installed and verified. If you get
   `RSC_ONBOARDING_INCOMPLETE`, the plan applied but the harness floor is missing: the output names
   every missing path and the action that creates it. Run that action — the `harness` skill scaffolds
   `01-TOOLS/` + `02-DOCS/`, and the `constitution` phase writes the constitution when SDD is
   selected — and do not proceed as if the harness were ready. When SDD is selected, onboarding
   itself writes the constitution as a draft (`status: draft`) and prints it as `Pending:` under
   `RSC_ONBOARDING_READY`: the harness is ready, and the `constitution` phase completes the draft
   before the first SDD feature. `RSC_SKELETON_FAILED` on the same run
   means rsc could not write part of the skeleton itself: the install still applied, and the line
   names the obstacle (usually a file or a symlink sitting where a directory belongs). Clear that
   obstacle first — no skill can, because it is not a missing file but an occupied path.
7. Tell the user to reload or restart the assistant so its new skills and hooks become active. Then,
   **only if you saw `RSC_ONBOARDING_READY`**, confirm rsc is ready and wait for the user's own task.
   If you saw `RSC_ONBOARDING_INCOMPLETE`, say instead what is missing and that you are completing
   it — never tell the user the harness is ready when it is not, because that is the one failure they
   cannot diagnose: everything looks installed and nothing behaves as if it were.

Already installed: `rsc sync`. Add one skill: `rsc add <id>`. Reassess deferred capabilities without
installing: `rsc reassess`. Recommend without installing: `rsc consult "<what the user wants>"`.

Canonical non-interactive preview:
`npx @ericrisco/rsc@latest onboard --technical-level technical --project-kind software --goal "small website" --software-scope small --target codex`

## Key links

- [npm package](https://www.npmjs.com/package/@ericrisco/rsc): install with npx @ericrisco/rsc@latest onboard
- [README](https://github.com/ericrisco/rsc-harness): full documentation and reference
- [Website, English](https://ericrisco.github.io/rsc-harness/): overview, agent instructions and FAQ
- [Website, Español](https://ericrisco.github.io/rsc-harness/es/): la misma página en español
- [Skill catalog (JSON)](https://ericrisco.github.io/rsc-harness/catalog.json): every skill id grouped by domain
- [Releases](https://github.com/ericrisco/rsc-harness/releases): changelog
- [Agent Skills specification](https://agentskills.io/specification): the open format the skills follow

## Facts

- 272 skills; minimal profile 7, core 30, full 272.
- 18 supported assistants (newest: DeepSeek Harness, `--target deepseek`); 34 agents and 53 command entries in the catalog.
- Every skill cleared an adversarial behavioral quality gate of at least 8.5/10.
- Free, MIT, no API key and no global install required.