autonomous-run
Prepare, start, inspect, resume, or stop a finite local overnight coding run after a human has accepted a Wayfinder terminal…
Orient a non-coder to this project's Agentsmith harness — fires on "what is this harness?", "what are my rules?", "which profile am I on?", "what do I type next?". Part of the Agentsmith harness; reads the canonical instructions and selected agent's safety configuration, then
$ npx -y skills add PromptPartner/agentsmith --skill harness-help --agent claude-codeHow it fires
How this skill gets triggered: by you, by Claude, or both.
/harness-helpContext preview
The summary Claude sees to decide when to auto-load this skill.
Orient a non-coder to this project's Agentsmith harness — fires on "what is this harness?", "what are my rules?", "which profile am I on?", "what do I type next?". Part of the Agentsmith harness; reads the canonical instructions and selected agent's safety configuration, then
name: harness-help description: Orient a non-coder to this project's Agentsmith harness — fires on "what is this harness?", "what are my rules?", "which profile am I on?", "what do I type next?". Part of the Agentsmith harness; reads the canonical instructions and selected agent's safety configuration, then reports the profile, rules, safety mode, and next step in plain language. compatibility: Requires an Agent Skills-compatible coding agent with filesystem access.
Explain, in plain language with no jargon, what this harness does for the operator and what to do next. This is the non-coder's front door.
"what is this harness / what does it do?" / "what are my rules?" / "which profile am I on?" / "what do I type next?" / a general "help me get started here".
Use the agent named by the operator or the explicit `AGENTSMITH_AGENT` value. Never infer the active agent from this skill's install path: portable canonical content may be copied into a required client adapter such as `.claude/skills`. Without a declared agent, explain the common rules and label agent-specific safety configuration as unknown.
Read the canonical managed block in `AGENTS.md` and report: 1. **Active profile(s)** — the work-type this project is tuned for (from the profile section). 2. **Your rules, in 3–5 bullets** — the load-bearing ones (verify before done, protected main, look before you delete, be honest about what failed). Summarize; don't dump the file. 3. **Safety mode** — when an agent is declared, read its managed native configuration and explain the effective approval and sandbox settings. Otherwise say that client-specific safety was not identified. 4. **What to type next** — take one small task end to end. Say “verify” or “handoff”; mention an invocation prefix only when the declared client documents one.
Point to `docs/01-harness-philosophy.md` and `docs/07-how-to-pick-a-profile.md` for the why and the profile choices. If `FIRST-STEPS.md` exists, point at it too.
A short, friendly orientation: the selected agent when known, profile, a few rules, safety mode, and the one thing to do next. No file dumps, no jargon.
Proof before done. Move from vibe coding to agentic engineering. Give Claude Code or Codex your project rules, a definition of done, and checks that produce evidence.
Prepare, start, inspect, resume, or stop a finite local overnight coding run after a human has accepted a Wayfinder terminal…
Template/example skill showing the minimal structure. Replace or delete it. Use as a starting point when you want to capture a…
Grill a fuzzy, single-session repository change into shared language, recorded decisions, and a handoff-ready plan. Use when the…
Checkpoint work for recovery — fires on "handoff", "wrap up", context pressure, a closed phase, or before a long external wait or…
Check whether this project's Agentsmith harness is installed correctly and healthy — fires on "is my harness set up right?",…
Log a harness lesson / post-incident so the SYSTEM improves, not just this one bug — fires on "log a harness lesson",…