psquared-development 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.
Output (the hand-off)
- The top-10 worklist (exact process names + IDs) — this is what gets split across terminals.
- A short context brief: industry (exact), real systems/ERP, transcript priorities, cross-cutting notes, and any elevate/swap flags for the human.
- 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).