SuperAPI AI Directory

processflow-overview

General & other

View repository

Updated 10-16 04:53

About this skill

ProcessFlow — Overview / Kickoff (from zero)

You are starting a fresh company. Build the map, not the buildings.

What this engagement is (context you need at 0)

neverlost (Dominik Rockenschaub) runs paid workshops; an AI fills a canvas per process in the ProcessFlow app (processflow.neverlost.at, Supabase backend) and proposes a Lösungskonzept + score + diagrams. psquared (Martin) red-teams AND refines those analyses — paid deliverable, per customer. The AI is often wrong (wrong industry, invented tools, rebuilds what the ERP already does, wrong AI technique, mis-priced). Our job: catch that and produce presentation-ready concepts. Load the shared knowledge: processflow-review (guardrails = Definition of Done) and processflow-app (how to drive it: helper app + Chrome-MCP). Standard = Dominik's SOP.

Phase 1 — Setup & context (SOP, once per project)

Do the SOP setup check (setup_verified in processflow-review/references/guardrails.md): customer briefing/industry/size, tool preferences, project IT-infra, project-tool types (Bestand vs. Neu), context & transcripts. Missing → ask the workshop lead, don't guess.

Phase 2 — Read the transcript (from the folder)

The workshop transcript is stored in the company folder, typically technical-analysis/<company>/meeting-notes.txt. Read it fully first — it's ground truth. Note: industry/products (exact!), real systems/ERP (exact names), the customer's stated priorities and pains, who said what. This is what every later transcript-check compares against.

Phase 3 — Ingest the processes

Chrome is logged into ProcessFlow as the user → pull the session token from localStorage (sb-…-auth-token → access_token) + anon apikey from any request; you do NOT need a pasted curl. Use the helper app (processflow-tool/, skill processflow-app) to ingest processes+scores+reports into data/<company>/. Filter by project_id (multiple customers share one org → RLS won't separate them). Save the transcript into the app too.

Phase 4 — Triage by Analyse-Reihenfolge (rule/filter/priority)

Run processflow-review/scripts/select.py <processes.json> --scores <scores.json> — it applies the 11-tier order (Kundenfavoriten·QuickWin·Foto → … → Rest). Take the top 10. Score/quadrant live in process_scores (NOT on the process row); photo = source_image_url. Surface anything near the cut. Question the ranking against the transcript: if the customer's #1 stated priority scores low (the AI does this), flag it and consider elevating — but that's a human decision; present it.

Phase 5 — Overview only (no deep work)

For each of the 10, one line: what it is (per transcript), cluster, score/quadrant, obvious concerns (no foto → AI-synthesized; not in transcript → suspect). Note cross-cutting themes (shared tools, InboxMate as intake, an OS-consolidation angle, industry-hallucination risk). Do not rework concepts, edit canvases, or touch diagrams here.

Postfach-Signal (one check, evidence only)

Scan the transcript and ALL processes (not only the top 10) for the shared-inbox pattern:

  • one address (office@/info@/service@/rechnung@) read by several people or teams,
  • mail sorted, forwarded or answered by hand; "wer ist zuständig?" delays,
  • misdirected mail (wrong person, wrong department, wrong company),
  • invoices, orders or requests that arrive by mail and are typed or re-filed manually. If found: list it as Postfach-Signal in the context brief, with the exact transcript quote(s) and the process name(s). Do not add InboxMate to any concept, canvas or score here — the human decides whether it is a lead (then CRM tagging per agenthub/docs/gtm/neverlost-crm-tagging.md). If there is no evidence, write "Postfach-Signal: keins" — never infer one from the industry alone.

Output (the hand-off)

  1. The top-10 worklist (exact process names + IDs) — this is what gets split across terminals.
  2. A short context brief: industry (exact), real systems/ERP, transcript priorities, cross-cutting notes, the Postfach-Signal (quotes or "keins"), and any elevate/swap flags for the human.
  3. Tell the user: next, open one terminal per process and run processflow-run <process name> (works the process to guardrail-completion + self-critique + fix). Overview does not build.

Seed the Team-Update-Feed (for the parallel crew)

The per-process terminals coordinate through the helper app's shared update feed (company-wide). Post your context brief + cross-cutting findings there so every worker starts informed: POST /api/update {company, agent:"overview", tags:[…], text:"…"} — e.g. exact industry, real ERP/systems (and any invented ones to avoid), stated priorities, shared modules/intake (InboxMate, article-matching), OS-consolidation angle, known pitfalls. Each processflow-run reads this feed at start and keeps posting its own cross-cutting findings.

Never set approved/HI-freigegeben (human) or name the presentation date (Dominik).