Does it repeat enough?
High-frequency work creates more value from automation than a one-off task.
High score when weekly or more frequent
Playbook live
Playbook 03
Updated 18 Jul 2026
Automation · Data pipelines · AI collaboration
Automation & AI System
A practical method for deciding what should be automated, building reliable data and deployment pipelines, using AI as a structured collaborator, keeping failures visible and preserving human control over consequential outputs.
This public playbook explains the operating method. Private prompts, credentials, account identifiers, client data and personal memory remain outside the public layer.
01 · When to use
Automate repeated decisions only after the process is understandable.
Automation works best when the inputs, rules, outputs and failure states can be described clearly. It should remove repetitive work without hiding the logic that makes the result trustworthy.
Use this playbook for scheduled data pulls, report generation, file synchronization, deployment workflows, content transformation and AI-assisted drafting or analysis where a human can define acceptable inputs and review the final result.
Avoid automation when the process is still changing daily, the cost of a silent error is high, or the workflow depends on judgment that has not yet been translated into explicit checks.
02 · Automation decision test
Score the workflow before writing the script.
A useful automation saves meaningful time without creating more operational risk than it removes.
Are the rules stable?
The workflow should have a repeatable shape even when its data changes.
Low score when the process is still being inventedAre the inputs accessible?
Files, APIs, sheets or events must be available in a format the system can validate.
High score with structured sourcesWhat happens when it fails?
A failed run should be recoverable without damaging the last known good state.
Low score when errors are irreversibleCan a human verify the result?
Important outputs should remain inspectable before they affect customers, money or public content.
High score with a clear review gateDoes it improve the operating system?
Automate to reduce delay, inconsistency or cognitive load—not simply because automation is possible.
Proceed when benefit exceeds maintenance03 · System architecture
Separate triggers, sources, transformation, validation and delivery.
A reliable workflow makes every stage inspectable instead of collapsing the entire process into one opaque script.
Schedule or event
Manual run, hourly schedule, daily schedule or a verified repository event starts the workflow.
Explicit startSource of truth
Google Sheets, platform APIs, repository files or approved structured inputs provide the data.
Controlled inputTransformation layer
Apps Script, Node.js, GitHub Actions or other logic converts the source into a usable output.
Repeatable logicQuality gate
Required fields, row counts, output paths, data freshness and content integrity are checked before release.
Protect good stateOutput and notification
Reports, generated pages, updated sheets or deployment artifacts are published with a clear result status.
Visible outcome
Trigger→Source→Transform→Validate→Human review→Release
04 · Build workflow
Design the operating contract before implementation.
Map the current process
Write the manual inputs, decisions, output, owner and exceptions before changing anything.
UnderstandDefine the source of truth
Choose which sheet, file, API or system is authoritative when sources disagree.
ControlSpecify success and failure
Define expected row counts, files, timestamps, status messages and abort conditions.
ObserveBuild the smallest working path
Automate one stable slice before adding fallbacks, multiple sources or AI assistance.
SimplifyAdd validation and retries
Check inputs before transformation, retry temporary failures and refuse unsafe output.
ProtectAdd human review
Keep approval where the output can affect public content, customers, spend or irreversible actions.
GovernDocument the runbook
Record triggers, owners, dependencies, common failures and recovery steps.
Maintain05 · AI collaboration
Use AI as a structured collaborator, not an invisible authority.
AI is most effective when it receives a clear role, grounded context, explicit constraints, a defined output and a review standard. The quality of collaboration improves when those elements are reusable rather than repeatedly improvised.
What should the AI do?
Research assistant, analyst, editor, developer, reviewer or another clearly scoped role.
What does it need to know?
Only the relevant project state, data, constraints and prior decisions required for the task.
What must it not do?
Privacy boundaries, unsupported assumptions, prohibited changes and quality requirements.
What artifact should exist?
Patch, report, recommendation, checklist, draft, code review or another inspectable result.
How should claims be grounded?
Files, source data, official documentation, tests or clearly labeled assumptions.
How is quality accepted?
Definition of done, validation steps and a human owner responsible for final approval.
06 · Observability
A workflow should explain what happened without requiring code archaeology.
Trigger and timestamp
Show when the workflow started, what initiated it and which configuration it used.
Sources selected
Record file names, sheet tabs, date windows, row counts or API scopes used in the run.
Steps completed
Make major stages visible instead of returning one generic success message.
Artifacts changed
List created, updated, skipped or deleted outputs with enough detail for review.
Non-blocking issues
Separate recoverable warnings from failures that must stop the workflow.
Clear final status
End with success, partial success or failed status and the next required action.
07 · Human review gates
Place approval where a mistake becomes expensive.
Safe to execute automatically
- Reading approved structured sources.
- Calculating repeatable metrics.
- Generating draft reports or pages.
- Running validation and non-destructive tests.
- Skipping no-change commits.
Keep a human approval step
- Sending messages to customers or guests.
- Publishing sensitive or high-visibility content.
- Changing spend, bids or financial settings.
- Deleting data or replacing the last known good output.
- Acting on uncertain AI recommendations.
08 · Automation runbook
Every recurring workflow should have an operating card.
The runbook makes ownership and recovery visible even when the original builder is unavailable.
Purpose
What repetitive problem does the workflow solve?
Trigger
Manual, scheduled or event-based—and how often?
Sources
Which files, sheets, APIs or repositories are authoritative?
Outputs
What files, reports, records or deployments should be created?
Validation
Which checks must pass before output is accepted?
Owner
Who reviews warnings, approves outputs and maintains the workflow?
Failure recovery
How is the last good state restored or the run repeated safely?
Change history
Where are important modifications and decisions recorded?
09 · Public / private boundary
Publish the method, protect the operating secrets.
Safe to document
- Workflow architecture and validation principles.
- General examples of Apps Script and GitHub Actions use.
- Human-review rules and observability patterns.
- AI role, context, evidence and output standards.
- Failure modes and recovery principles.
Never publish
- API keys, secrets, tokens or authentication details.
- Private prompts containing personal or client-sensitive context.
- Internal account identifiers and unpublished datasets.
- Customer, guest, lead or employee-level information.
- Private repository contents or memory-vault material.
10 · Common failure modes
Most automation problems begin with hidden assumptions.
Automating an unstable process
The workflow changes faster than the automation can be maintained.
No source-of-truth rule
Multiple sheets or files disagree and the system chooses silently.
Generic success messages
The workflow reports success without explaining what changed or what was skipped.
AI output accepted without evidence
A convincing recommendation is treated as correct before files, data or tests support it.
Failure destroys the last good state
Invalid or empty input overwrites working reports, pages or datasets.
No operational owner
The workflow runs regularly, but no person is responsible for warnings, updates or recovery.
Playbook status
Third playbook complete.
The Automation & AI System now provides a reusable method for building assisted workflows that remain observable, reviewable and safe to operate.