Claude Skills: From Skills To Workflows To An Agentic System

---
name: skills-to-workflows-to-system
description: Audits an owner's AI setup against three rungs (a skill is a file, a workflow is skills in order with a trigger, an agentic system is workflows with an owner, a cadence and a report), converts repeated prompts into skill files, writes workflows as ordered steps with handoffs and human stops, and promotes one workflow to a system with a job file. Trigger on "audit my ladder", "turn my prompts into skills", "which of my workflows are systems", "make this run without me".
---

# From Skills To Workflows To An Agentic System

The ladder:

- **Skill**: a file. Name, a description that says when to use it in the owner's own
  words, instructions for one job. An agent reads it before doing that job.
- **Workflow**: skills in a fixed order with a trigger (a schedule or an event). Each
  step names its skill and hands its output to the next. Human steps are marked.
- **System**: a workflow with a named owner, a day and time, and a report that lands
  in one place in a shape a stranger can judge.

Prompts are rung zero. A prompt dies in the chat it was born in.

## Step 1: Inventory

Ask the owner for the five things they ask AI to do most often and anything they have
automated. Classify each: prompt / skill / workflow without trigger / workflow with
trigger / system. Say which rung they are on overall. Do not flatter.

## Step 2: Prompts become skills

For every repeated prompt, write the skill file:

```
---
name: [kebab-case]
description: [what it does in one sentence, then "Trigger on" and the exact
phrases the owner would say]
---
# [Title]
[Instructions for one job. Inputs it reads. Output shape. What it never does.]
```

The description is a trigger, not a summary. If it is vague the file never loads.

## Step 3: Skills become a workflow

For any set of skills that already run in order, write:

```
WORKFLOW: [name]
TRIGGER: [schedule: day and time] or [event: exact event in the CRM / inbox / folder]
1. [step] : uses [skill] : takes [input] : hands [output] to step 2
2. ...
n. [step] : HUMAN : approves / does X, workflow stops here until they do
FAILS SILENTLY IF: [which step, what it looks like, where the workflow must stop and say so]
```

Draw it. If the owner cannot follow the drawing in ten seconds it is not done. The
public `draw-the-system` skill (https://marketing.doctorleadflow.com/skills/) does
the layout.

## Step 4: One workflow becomes a system

Pick the workflow closest to done. If the owner has no system yet, recommend a
reporter (reads and reports, cannot damage anything) before any doer. The public
`first-ai-employee` skill runs the interview for that. Write the job file:

```
SYSTEM: [name, let the owner name it]
OWNER: [a name: the owner, a teammate, or a named agent]
JOB: [one sentence]
INPUTS: [exactly which tools and data it reads, nothing else]
OUTPUT: [the report: where it lands, sections, and the one line at the top that
says whether the owner must act]
CADENCE: [day and time]
GUARDRAILS:
- Read only unless a line here says otherwise
- Drafts anything customer facing; never sends
- Says "no data for that" instead of guessing; an invented number is a firing offense
- Money, deletion and the public require the human
VOICE: [3 bullets in the owner's words]
REVIEW LOG: [empty; dated corrections go here, never in chat]
LOG LINE AFTER EVERY RUN: YYYY-MM-DD : DID / DECIDED / BLOCKED / NEXT
```

## Step 5: Leave them with the rule

One workflow promoted per week. Finish it before starting the next. Ten half built
agents is the graveyard. The owner's job on the top rung is to read reports and
decide, and to correct the file, not the chat.

## Rules

- Never invent tools, numbers or steps the owner did not describe.
- One source of truth per fact; if two systems would read different places, flag it.
- Every live write in any workflow passes an approval step. No exceptions for speed.