deusXmachina-dev/automation-instructions

Automation Instructions

Turn one workflow into clear, build-ready instructions for a specific automation tool. Assumes it runs after /discover-patterns or /process-analyst-new, so it reuses their output as the source and does NOT re-mine the timeline unless nothing is supplied.

Step 1. Get the workflow (prefer prior output)

In priority order, do NOT re-derive what you already have:

1. A pattern from /discover-patterns (a selected pattern + its details) -> use it as-is.
2. A process from /process-analyst-new (a process-analysis*.json ledger, or a process the user points to) -> read that process's steps, apps, frequency, and automatable split. Glob process-analysis*.json in the working dir; if found, use it.
3. A workflow the user describes in chat -> use the description as-is.

Only if NONE of the above is available, fall back to mining: search_context(query) across ~30 days for instances, browse_timeline around the strongest matches for the full sequence, and get_activity_details(ids) on 2 to 3 clear instances for exact steps. Keep it bounded.

Step 2. Ask three things (one batched AskUserQuestion)

1. Target tool - Claude, n8n, or Zapier. Drives the whole output style.
2. Constraints - anything it must NOT do (no writes to prod, no emails to customers, human approval required, leave step X manual, etc.).
3. Output format - text, markdown, or JSON. Assume markdown if skipped.

Respect the constraints everywhere in Step 3, and call out any step the constraints force to stay human.

Step 3. Write the instructions (tailored to the tool + format)

Reconstruct the flow from trigger to done, separating variables (change each run, the inputs) from constants (fixed, hardcoded). Then write it for the chosen tool:

• Claude (Cowork / Skill) - frame as Connect, Create, Run. Which apps to connect, the Skill/agent brief and the steps Claude runs, what it drafts vs writes, and the human approval gate before anything is sent or saved.
• n8n - the trigger node, the ordered nodes with the app/action and the data mapped between them, and error/retry handling. Deterministic, so any judgment step stays a human or AI-assist node.
• Zapier - the Trigger, then one Action per step (app + action + field mapping), plus any Filter or Path for branches. Note steps Zapier cannot do that need a human or code.

Render in the chosen format: markdown = the skeleton below; text = the same content as plain prose, no markdown; JSON = one object { name, tool, trigger, variables[], constants[], steps[{n, app, action, input, output, on_error}], human_steps[], constraints[] }.

Markdown skeleton (keep it tight, drop empty sections):

# [Workflow] - Automation instructions ([tool])

**Does:** [one line] **Trigger:** [what starts it] **Frequency:** [from the source]
**Constraints:** [what it must not do]

## Variables (inputs that change each run)

- [name]: [example]

## Constants (hardcoded)

- [name]: [value]

## Steps

1. [Action] in [app] - input [x] -> output [y]. On error: [fallback].
2. ...

## Stays human

- [step + why, incl. anything the constraints forced manual]

## Build notes ([tool])

- [trigger + key apps/nodes/actions + effort: easy/medium/hard]

Step 4. Save and present

Ask where to save (default ~/Desktop/automation/[workflow-slug].md, or .txt / .json to match the format). Create the folder if needed. Then show the path and a 2-line summary (tool, step count, key apps). Offer to do another workflow.

Notes

• Privacy - never put passwords, API keys, or personal-message content in the output; keep only process-relevant details (app names, field labels, URLs).
• Gaps - if the source evidence is thin, say which steps are assumed and what the user should confirm before building.