Skill

automation-pick

Before automating a task, decide whether to automate. Decision tree based on volume, frequency, error cost, and reversibility. Use when someone says 'we should automate this' or 'let's build a script for that'.

Category
core
Version
0.1.0
License
MIT
Status
active
Length
93 lines
Source
skills/core/automation-pick/SKILL.md

Install and discovery

npx --yes --package @youcisla/agent-foundry agent-foundry-install

After the installer runs, this record is available as automation-pick. See the docs for harness roots and verification steps.

Contents

Automation Pick

A framework for deciding which tasks are worth automating. Most tasks aren't.

When to Use

  • Someone says "we should automate this"
  • You're tempted to write a script for a one-off task
  • Reviewing a backlog of "automate X" requests
  • Building automation that turned out to be wrong

Procedure

The Decision Tree

Answer in order. Stop at the first "no".

1. Has this been done manually at least 3 times?

If no → don't automate yet. Do it manually 3 times. Patterns only emerge with repetition.

2. Will it be done ≥10 more times in the next year?

If no → don't automate. Even a 2-hour script doesn't pay back.

3. Does manual execution have a real error cost?

CostExamplesAutomate?
CatastrophicBilling, security, data loss✅ Yes, urgently
AnnoyingWrong formatting, missed deadlines✅ Yes, eventually
CosmeticLooks bad❌ Probably not
Zero"It would be nice"❌ No

4. Is the task fully understood and deterministic?

If the rules keep changing, or humans make judgement calls → don't automate (yet). Standardize first, automate second.

5. Is the task reversible?

If the automation runs and gets the wrong answer:

  • Easy to detect and revert → automate
  • Hard to detect, propagates → don't automate without safeguards

The 4-Quadrant Matrix

Low error costHigh error cost
High volumeAutomate freelyAutomate carefully (with monitoring)
Low volumeDon't automateDon't automate (do it manually + check)

Worked example

A rejected automation that stops at the first failed question.

Job: rename screenshots every week.
1. Manual runs: 2 -> stop here.
Decision: do not automate yet; log the next three manual runs.
Revisit when manual runs >= 3 and forecast >= 10/year.

Anti-patterns

  • Writing a script for a task done twice.
  • Automating a judgment call with no stable rules.
  • Automating an irreversible action with no safeguard.

References

  • details

Verification Checklist

  • [ ] Manual runs >=3 and forecast >=10/year are recorded. (step 1)
  • [ ] Error cost and volume place it in the matrix. (step 3)
  • [ ] The task has deterministic rules. (step 4)
  • [ ] Reversibility and detection are stated. (step 5)
  • [ ] A rejected automation has a reason. (step 1)