Three AI Automation Pipelines: From Manual Repetition to Hands-Free Workflows

Criteria for Automation

The author proposes two practical tests to decide whether a task deserves automation:

Three-repetition rule: One occurrence is chance, two is coincidence, three is a pattern. Tasks repeated three times should not be done manually again.

Handoff-document test: Can you write a clear handoff document for a new intern? If you find yourself writing "depends on the situation" or "I need to decide here," the task is not ready for automation.

Technically, an automation pipeline consists of three elements: when (trigger), what (instruction), and where (output destination). The pipeline wakes at the scheduled time, executes the instruction verbatim, and deposits the result in the specified location. Crucially, the pipeline has no memory — it does not know what happened yesterday or last week.

Pipeline 1: Daily AI Morning Brief (7:30 AM)

Function: Searches AI industry news from the past 24 hours, selects six items, writes summaries with links, and saves as a dated file.

Time saved: ~20 minutes daily (~60+ hours per year). More valuable than time is the cognitive benefit: the author starts the day with curated content instead of spending 30 minutes switching into "work mode."

Failure 1 — Missing time range: Initial instruction "search today's AI news" returned articles from three days prior, including two already published. Fixed by adding "past 24 hours" to the instruction.

Failure 2 — Missing output location: Files landed in the workspace root directory, mixed with other files. The author had previously written about a three-layer directory structure but failed to apply it here. Fixed by specifying the exact save path.

Pipeline 2: Wednesday Material Audit (3:00 PM)

Function: Scans the "material layer" (where ideas are dumped), lists recent entries, and flags duplicates against the "settled layer" (finalized content).

Time saved: ~1 hour each weekend, plus elimination of the "I think I already thought of this" uncertainty.

Failure — Duplicate detection: First run surfaced an idea the author had already used in a published article six months earlier. The pipeline, like a human, rediscovered the same duplicate. Fixed by adding an explicit instruction: "First compare against the settled layer; mark duplicates as 'already used'." After this, only truly novel items are surfaced.

Pipeline 3: Friday Weekly Report Draft (5:00 PM)

Function: Reads all files modified during the week, extracts accomplishments, blockers, and next-week plans, and saves a draft document. Monday morning the author reviews and edits in ~20 minutes instead of spending 40 minutes reconstructing the week from memory.

Failure — Hallucination due to statelessness: First version used the instruction "write a weekly report like last week's." The pipeline fabricated plausible but false achievements: "completed three technical design reviews" and "produced two deep-dive articles." The root cause: the pipeline runs as a fresh instance with no memory of previous weeks or current week's work. "Like last week" was interpreted as "invent one that looks like a weekly report."

Fix: Changed instruction to "read this week's modified files and summarize from them." Since then, no hallucinations have occurred.

Pre-Deployment Checklist (Four Rules)

Complete the instruction: Specify source location, search criteria, output path, and file naming. The pipeline has no context; anything omitted will be guessed.

Manual test first: Run the instruction once in a chat session, verify output quality, then promote to scheduled automation. Skipping this wastes runs and credits.

Provide a fallback: Add "if the file does not exist, note the absence in the report and continue" to prevent the pipeline from stalling on missing inputs.

Set an expiry date: Temporary tasks (e.g., exam-prep reminders) must have a stop date. The author once forgot to set one, and a task ran for over two months, consuming credits unnecessarily.

Cost awareness: Each run consumes credits. The author's three pipelines combined use minimal credits per month; avoid over-scheduling (e.g., hourly runs) unless truly necessary.

What Not to Automate

High cost of error: Emails sent automatically without human review.

Requires ad-hoc judgment: Tasks where the machine would need to guess the author's intent; wrong guesses create cleanup work.

The ideal candidates are fixed-schedule, fixed-process tasks that require no decision-making once started.

Closing Analogy

The author draws on factory experience: an automated production line feeds material in, runs on a fixed cadence, and humans only monitor gauges and handle exceptions. After 20+ years, the same principle applies to personal workflows — humans should read the gauges, not count the beats.