Purpose
Turn a raw process description or SOP into a scored, prioritized automation plan — showing which steps are worth automating, how risky each one is, the specific tooling approach to use, and the order to build in. Replaces gut-feel automation decisions with a repeatable scoring model that survives scrutiny from finance and engineering.
When to Use
Invoke this skill when the user asks to:
- Evaluate whether a specific process or SOP is worth automating
- Decide which steps in a workflow to automate first
- Prioritize an automation backlog across multiple candidate processes
- Build a business case or ROI justification for automating a task
- Choose the right approach (RPA, workflow tool, AI agent, or API) for a given step
- Assess the risk of automating a customer-facing or compliance-sensitive process
Inputs Required
Before starting, collect:
- Process description or SOP: Step-by-step account of how the work is done today (paste the SOP or describe the flow)
- Volume: How often the process runs — executions per day/week/month
- Time per run: Approximate minutes of human time per execution
- Who does it: Role(s) performing the work and their loaded hourly cost if known
- Systems involved: Tools, apps, or databases touched during the process (e.g., Salesforce, Zendesk, spreadsheets, email)
- Decision points: Any steps where a human judges, approves, or handles exceptions
- Sensitivity: Whether the process is customer-facing, touches money, or is subject to compliance/regulatory rules
If loaded cost is missing, assume $35/hr for support/ops roles and note the assumption. If time-per-run is missing, estimate from the step count and flag it.
Instructions
Step 1: Decompose the Process into Discrete Steps
Break the SOP into atomic steps — each a single action with one input and one output. Number them. For each step, tag it as one of:
- Trigger (what starts the step)
- Data movement (copy, transfer, enter, look up)
- Transformation (calculate, format, summarize, classify)
- Decision (approve, route, judge, choose)
- Communication (notify, respond, escalate)
A step that mixes multiple actions should be split — mixed steps hide automation boundaries.
Step 2: Score Each Step (Automation Opportunity Score, 0–100)
Score every step across four weighted dimensions:
- Frequency (30%): How often the step runs. High-volume, repetitive steps score highest.
- Rule clarity (30%): Can the step be described as explicit if/then rules? Deterministic = high; requires human judgment = low.
- Data structure (20%): Is the input structured (fields, forms, API data) or unstructured (free text, images, phone calls)? Structured scores higher — but note that AI now handles semi-structured well.
- Time saved (20%): Minutes of human time removed per run × frequency.
Compute a weighted 0–100 score per step. Band the result:
- 80–100 — Prime candidate: Automate now
- 60–79 — Strong candidate: Automate after prime candidates ship
- 40–59 — Conditional: Automate only if a decision point can be codified or a human stays in the loop
- 0–39 — Keep manual: Judgment-heavy or low-volume; not worth the build
Step 3: Assign a Risk Level to Each Step
Independent of the opportunity score, rate risk Low / Medium / High using:
- Error blast radius: What breaks if the automation does the wrong thing? (internal cleanup = low; wrong customer charge or message = high)
- Reversibility: Can a mistake be undone? Irreversible actions (payments, deletions, sends) raise risk
- Compliance exposure: Regulated data, financial actions, or contractual commitments raise risk
- Exception rate: If the step frequently hits edge cases, full automation is riskier
Flag any step that is High opportunity + High risk — these need human-in-the-loop design, not full automation.
Step 4: Recommend an Approach per Step
Match each automatable step to the lightest-weight tool that does the job:
- API / native integration: Best when both systems have APIs and the step is pure data movement — most reliable, lowest maintenance
- Workflow / iPaaS tool (Zapier, Make, n8n, Workato): Best for multi-app triggers and simple transformations without code
- RPA (UiPath, Automation Anywhere): Only when systems have no API and require screen-level interaction — highest maintenance burden, treat as last resort
- AI agent / LLM step: Best for transformation and semi-structured decisions — classifying tickets, drafting responses, extracting fields from free text, summarizing. Pair with human review when risk is Medium or High
- Hybrid (human-in-the-loop): AI drafts or recommends, human approves — the default for High-opportunity/High-risk steps
State the specific tool category and one concrete example tool per recommendation.
Step 5: Estimate Effort and Payback
For each Prime and Strong candidate:
- Build effort: S (under a day) / M (1–3 days) / L (1–2 weeks) / XL (multi-week project)
- Annual time saved: minutes/run × runs/year ÷ 60 = hours; × loaded cost = dollar value
- Payback: rough build cost vs. annual savings — flag anything that pays back in under 3 months as a quick win
Step 6: Sequence the Rollout
Order candidates into three waves:
- Wave 1 — Quick wins: High opportunity, Low risk, S/M effort
- Wave 2 — High value: High opportunity, any risk, but higher effort or needing human-in-the-loop design
- Wave 3 — Conditional: Steps that require codifying a decision rule or process change first
Note any dependencies (Step B can't automate until Step A produces structured output).
Output Format
- Structure: Numbered step decomposition → per-step scoring table → risk flags → approach recommendations → effort/payback → 3-wave rollout sequence
- Length: 800–1,200 words
- Format: A scoring table (Step | Type | Opportunity Score | Band | Risk | Recommended Approach | Effort | Annual Savings) followed by prose for risk flags and the rollout sequence
- Sections: Process Breakdown → Scoring Table → High-Priority Risk Flags → Approach Recommendations → Effort & Payback → Rollout Sequence (Waves 1–3)
- Tone: Operations lead — pragmatic, numbers-first, biased toward the lightest tool that works
Quality Checklist
Before delivering output, verify:
- [ ] Every process step has an opportunity score (0–100), a band, and a risk level — none skipped
- [ ] Opportunity score and risk are rated independently (a step can be high opportunity and high risk)
- [ ] Each automatable step names a specific tool category and one example tool — not just "automate it"
- [ ] Every High-opportunity/High-risk step is flagged for human-in-the-loop design
- [ ] Annual time-saved and payback figures show the math (runs × time × cost), with assumptions noted
- [ ] The rollout sequence puts quick wins in Wave 1 and calls out any inter-step dependencies
Notes
- Bias toward the lightest tool: an API integration beats an RPA bot, and a workflow tool beats custom code, whenever they can do the job. RPA is a last resort — it breaks whenever the underlying UI changes.
- The most common mistake is automating a step that should be eliminated. If a step exists only to patch a broken upstream process, recommend fixing the upstream process instead of automating the workaround.
- AI agent steps have shifted the "rule clarity" bar — semi-structured judgment (classifying, drafting, extracting) that was un-automatable five years ago is now Wave 1 territory, provided a human reviews outputs when risk is elevated.
- If the process description is too vague to decompose, ask one question: "Walk me through the last time you ran this, click by click."
How to use this skill
This file is a Claude Code skill in Anthropic's standard format — markdown with YAML frontmatter at the top.
Installing in Claude Code:
~/.claude/skills/<folder-name>/SKILL.md— auto-loads in every session anywhere on your machine (recommended for skills you'll use across multiple projects)<your-project>/.claude/skills/<folder-name>/SKILL.md— auto-loads only when working in that project (recommended for project-specific skills)
Restart Claude Code or open a new session to register.
Using in Claude.ai, ChatGPT, Gemini, or other AI tools:
- Drag this file into a new chat, or copy and paste its contents
- Send your task or request as your next message — the AI will apply the skill where it fits
For Claude Code skill details: https://docs.anthropic.com/en/docs/claude-code/skills