---
title: "Agent ops"
description: "Four skills for running agents on real work: a project brief, a workflow audit, a prompt-to-skill conversion and a fallback plan."
canonical: "https://www.diginomads.win/skills/agent-ops"
source: "content/pages/skills/agent-ops.md"
updatedAt: "2026-11-10"
---

**Level:** Intermediate

## Skills

| Skill | Trigger | Output | Guardrail |
|---|---|---|---|
| project-brief | When a harness needs project instructions, or the same rules are repeated in every chat | A short AGENTS.md with commands, boundaries, workflow, ask-first rules and a definition of done | Every command is one the user has run; no secret or client name appears in the file. |
| workflow-audit | Before handing a recurring workflow to an agent | A step map with agent/human ownership, failure modes, reversibility and human checkpoints | No irreversible step is automated without a checkpoint before it. |
| prompt-to-skill | When the same prompt is pasted more than twice a week | A SKILL.md folder plan with a specific description, procedure, output contract and two pass tests | The description excludes a nearby task; the skill does not trigger for everything. |
| fallback-plan | When a workflow depends on an API, model, queue or network | Failure classes, detection signals, capped retries, labelled fallbacks and a runbook | Silent wrong output is a failure; a non-idempotent write is never retried without a key. |

## Install

1. Create `skills/project-brief/`, `workflow-audit/`, `prompt-to-skill/` and `fallback-plan/` in the repository.
2. Give each skill a description written as the moment it should fire, not the topic it covers.
3. Put project-specific commands and paths in `references/`; keep the shared skill vendor-neutral.
4. Run the tests below against a real repository. If the project brief invents a command, the skill is not being followed.

## Tests

- **Prompt:** Write the AGENTS.md for this repository. The test command is `npm test` and the generated client is off-limits.
  **Pass:** The brief names `npm test`, the generated-client path and the ask-first decisions, and contains no secret.
- **Prompt:** Should I automate the weekly client report with an agent?
  **Pass:** The audit marks drafting vs deciding, puts a checkpoint before sending, and proposes one pilot step with a metric.
- **Prompt:** Turn this prompt I paste every Monday into a skill.
  **Pass:** The output includes a specific description, a SKILL.md body, a guardrail and one test that should not trigger.

## Sources

- [Agent Skills specification](https://agentskills.io/specification)
- [Codex CLI documentation](https://learn.chatgpt.com/docs/codex/cli)
- [Claude Code documentation](https://code.claude.com/docs/en/overview)

## Why this pack

Prompts make a single task repeatable. This pack makes the *system* repeatable: the project rules, the workflow boundaries, the skill that loads without a reminder, and the fallback when a dependency fails.

It is the difference between “the agent worked in that chat” and “a new session can pick up the work tomorrow”.

## The rule behind the pack

Write the rule down once, then test it. A project brief, a skill and a fallback plan are all claims about how the work behaves. Each one needs a test that would fail if the claim were false.

## Adapting it

- For a team, review the project brief like code and keep the skills in the repository.
- For a client repository, add the data-handling section and the approval path for a new dependency.
- For a solo workflow, keep the pack small: one project brief, one workflow audit, one fallback.

## Failure modes

- **The brief is a second README.** Cut the description and keep the commands and boundaries.
- **The skill triggers for everything.** Rewrite the description as a situation, not a topic.
- **The fallback is untested.** Inject the failure in staging before you need the runbook.

## Related

- [Write the project brief your agent actually reads](/prompts/write-agents-md)
- [Audit an agent workflow before you automate it](/prompts/audit-an-agent-workflow)
- [A portable agent setup](/harnesses/portable-agent-setup)

## Structured data

```json
{
  "@context": "https://schema.org",
  "@type": "WebPage",
  "name": "Agent ops",
  "url": "https://www.diginomads.win/skills/agent-ops",
  "description": "Four skills for running agents on real work: a project brief, a workflow audit, a prompt-to-skill conversion and a fallback plan.",
  "isPartOf": {
    "@type": "WebSite",
    "name": "DigiNomads",
    "url": "https://www.diginomads.win/"
  }
}
```