Engineering Workspace

Coding Agent Skills

Install Stimulir operator skills into Codex or Claude Code, then drive the application lifecycle with short, safe prompts from the adopter repository.

Install

bash
npx skills add stimulir/skills

The installer discovers the repository's SKILL.md packages and places the selected skills in the coding agent's skill directory. Operator skills use the installed Stimulir CLI or SDK; they are not copied into the workspace sandbox. The public source catalog is stimulir/skills.

Lifecycle

0connectEstablish authenticated, cost-visible access.
1migrate-inferenceRewire an existing application onto the Stimulir gateway.
2privacy-layerDe-identify sensitive text before persistence or forwarding.
2capture-tracesTurn live traffic into curated, immutable data assets.
3prompt-versioningMove prompts into durable versions and labels.
3eval-runMeasure one prompt, model, or adapter change.
3rsi (legacy skill for Improve)Diagnose and hill-climb behavior through a server-owned controller.
3eval-promoteApply one reviewed proposal with explicit human authorization.
∞usage-auditKeep cost-per-task visible throughout the lifecycle.

Operator versus managed

Operator skills

Run on the engineer's machine inside Codex or Claude Code. They inspect the adopter repository and call Stimulir through its public CLI or SDK. Improve is an operator interface over a durable server controller: the coding agent submits one action and returns instead of holding an improvement loop open.

Managed skills

Are imported into a Stimulir workspace and materialized into an isolated runtime sandbox. They use the workspace Vault and storage contracts. They now live in the separate public stimulir/managed-skills catalog. See Managed Skills.

Project-local context

Run lifecycle skills from the adopter repository. For application workflows, the nearest relevant dotenv file is the authority; a stale saved human login must not silently replace its production API, key, or project.

.env
STIMULIR_API_BASE=https://api.stimulir.com
STIMULIR_API_KEY=hyb_...
STIMULIR_PROJECT_ID=...

Skills should read recognized Stimulir keys without sourcing the entire dotenv file or printing secret values. A hyb_* key is already workspace-bound; no separate workspace export is required.

Short prompts

text
Connect this application to Stimulir and verify one real inference call.

Capture today's production traces into a privacy-safe immutable snapshot.

Run Improve on this project using the last 30 days of production traces tagged assessment. Do not promote.

Review the completed eval proposal and show me the evidence; do not apply it yet.

The skill carries the workflow contract. The prompt supplies intent and constraints, not a hand-written sequence of capture, eval, diagnosis, prompt-file, polling, and promotion commands.

Boundaries

  • Do not print credentials or silently cross production and staging contexts.
  • Do not ask users to manually reproduce mechanics already owned by the SDK or server.
  • Detached operations return durable ids and console links rather than polling inside agent context.
  • Promotion remains a distinct, explicit, human-authorized action.

Continue with Improve for the diagnostic controller contract or CLI Installation for terminal setup.