Processize

Processize

Codify a process AFTER you have validated it manually. Never automate a workflow that is still being discovered. Automation locks in whatever you point it at — including the bugs.

Mindset

"If you automate before validating, you scale the wrong thing. If you process-ize after validating 3+ times, you've earned the right to compress." — paraphrased from Sahil Lavingia, The Minimalist Entrepreneur

Three-stage maturity:

1. Manual — first 1–3 times. Pure improvisation. Goal: learn what the process actually is.
2. Documented — 3rd to ~10th time. Same human, written checklist, no automation. Goal: prove the steps are stable.
3. Processized — 10th+ time. Checklist + AI assist + selective automation. Goal: compress the predictable parts, leave judgment with the human.

Skipping straight from 1 to 3 is the most expensive bug in vibe coding.

When to invoke

• User says "tô fazendo isso toda semana" / "isso já cansou" / "deve ter um jeito mais rápido".
• User says "vamos automatizar isso?" — invoke FIRST to check it has been validated 3+ times.
• daily-standup skill flags the same manual task across 3+ sessions.
• Post-launch operations work that is repeating: weekly user outreach, content publishing, retention checks, billing reconciliation.
• Before introducing any cron, scheduled job, webhook, or AI loop that codifies a workflow.

When NOT to invoke

• The task has been done fewer than 3 times.
• The task changes shape every time it runs (still in discovery).
• The user wants to redesign the process, not codify it — use brainstorming or prototype-lab instead.
• The task is a one-off (e.g. "set up the Stripe account") — write a session log, not a process.

The 5-question gate

Before producing a process document, the skill MUST get answers to these. Ask them one at a time if missing:

1. How many times has this been done manually? (< 3 → stop, do it manually a few more times.)
2. Did it produce the desired outcome the last 3 times? (no → not yet ready; debug the manual version first.)
3. Which steps are identical every time? (these are the automation candidates.)
4. Which steps require human judgment? (these stay manual or become AI-assisted.)
5. What would happen if the automated version ran wrong for a week without anyone noticing? (defines the safety net.)

If question 5 has a P0 answer (sends emails to customers, charges cards, deletes data), the process MUST include a circuit breaker, a sample size limit, and a notification.

Output format

Generate docs/processes/<slug>.md with this structure:

# Process — <name>

**Status:** documented · partially-automated · fully-automated
**Owner:** <human name>
**Last validated manually:** <YYYY-MM-DD>
**Run frequency:** <weekly / per-customer / on-event>
**Failure blast radius:** <none / cosmetic / financial / customer-trust / P0>

## Trigger

What event or condition starts this process.

## Steps

| # | Step | Mode | Notes |
|---|------|------|-------|
| 1 | ... | manual / AI-assisted / automated | ... |
| 2 | ... | ... | ... |

## Inputs

What you need before starting.

## Outputs

What this process produces. Where it lives.

## Quality gates

What "done correctly" looks like. How to verify.

## Failure modes

The 2–3 most likely ways this breaks. Detection + recovery for each.

## Automation boundary

Which steps are safe to automate today. Which require more validation runs first. Why.

## Circuit breakers (if any step is automated)

- Daily/weekly cap on runs.
- Required notification on every Nth run or on any error.
- Manual approval gate for actions with non-trivial blast radius.

## Change log

| Date | Change | Why |
|------|--------|-----|

Hand-off rules

• The skill never builds the automation itself. It produces the document. Implementation is a sprint task.
• The first automated version of any process MUST run alongside the manual version for at least 2 cycles before the manual version is retired.
• Any step with Failure blast radius: financial or higher requires /multi-ai-review before automation.
• Processes that touch personal data trigger privacy-audit.

Examples

Example 1 — first-100-users outreach

• Manual runs: 12 times (week-over-week).
• Stable steps: pull list, write personalized DM, log response, update CRM.
• Judgment steps: choosing tone, deciding when to follow up.
• Automatable today: list pull, CRM logging.
• NOT automatable yet: the DM body. Save for v2 once tone is calibrated.

Example 2 — billing reconciliation

• Manual runs: 4 times.
• Stable steps: download Stripe CSV, match against Supabase user table, flag orphans.
• Failure blast radius: financial — circuit breaker required if automated.
• Recommended: AI-assist for the matching step, human signoff for the flag-and-resolve step.

Cross-references

• Principle: ETHOS.md item 13.
• Pre-condition: daily-standup flagged repetition.
• Often paired with: cost-watchdog (automation can quietly increase API spend), privacy-audit (if personal data is involved), /multi-ai-review (financial blast radius).
• Inspired by: docs/registry/packs/slavingia-skills.md.