Insights
Notes from building Spectacle and the systems around it: computation paths, boundaries, and what we learned along the way.
Start here for law-firm AI tools: How German attorney service-provider rules apply to AI tools in a law firm.
Firm-wide AI or a service for one matter: how to document the classification
Is an AI tool part of the firm's general infrastructure, or does it serve one specific matter? How law firms can document that classification for counsel, with examples and a template outline. Not legal advice.
ReadHow to map where an AI tool's prompts and logs actually go
A documentation method for law firms: trace every party that receives prompts, files, logs and backups from an AI tool, and where, before counsel reviews it. Not legal advice.
ReadHow the German secrecy offence and attorney service-provider rules work together for AI tools
Two German provisions apply when a law firm uses an AI tool: the criminal secrecy offence (StGB §203) and the professional rules on service providers (BRAO §43e). A short operational reading of how they connect. Not legal advice.
ReadHow to remove client and matter details from AI prompts
A worked method for law firms: strip names and identifiers, replace people with roles, generalise dates, places and amounts, and test for recognisability before a prompt leaves the firm. Not legal advice.
ReadHow to connect AI to your law firm software through an interface
What changes when AI reaches your files through an API instead of a chat window: read-only by default, scoping by object class, logging, and human approval before any write. Not legal advice.
ReadRAG with human release for sensitive work.
RAG with human release for sensitive work: retrieval and generation stay behind approval before answers touch operational or matter data.
ReadHow to set up local inference under German data rules.
How to set up local inference under German data rules: open-weight or local (plus German cloud where needed) with access control, no-train, deletion, and approval of which objects may enter.
ReadWhat an agent harness does in production.
What an agent harness does in production: tools, memory, permissions, eval, and release gates around the model — so proposals wait for human approval.
ReadHow German attorney service-provider rules apply to AI tools in a law firm.
Name the service-provider chain for law-firm AI tools: selection, text-form programme, config, human review. Not legal advice — a working map for counsel and ops.
ReadGraph RAG versus vector RAG for business files.
Vector RAG finds similar chunks. Graph RAG follows entities and relations in matter files. Choose the shape that matches the question — then keep human approval.
ReadNamed next action. Not another report.
A report that never names the next action is still a report. Agents on Spectacle propose a concrete next step; you hold Freigabe.
ReadSimulate before you bind labour and material.
Commit labour and material only after a simulation on what the operation already knows. Agents on Spectacle propose; you hold approval.
ReadKill the drag work. Don't add another dashboard.
Operators lose mornings to hand-drag: redeciding, stitching systems, assembling lists. Agents on Spectacle take that work — not another cockpit of tiles.
ReadNever write back into the sealed till.
Legally sealed systems of record are never written to. Agents on Spectacle propose; approval stays yours. The stack sits above ERP — not inside the sealed till.
ReadCapacity is already committed.
Capacity is already committed before demand clears. Agents on Spectacle propose inside that limit; approval stays yours. No capacity theatre.
ReadModels on STACKIT are not the operation.
Private models on STACKIT bless the metal. The product gap is the operation graph — ontology, connectors, approval, retainer. Not a GPU cloud.
ReadObjects from the Betrieb. Not a chat demo.
Spectacle binds real Betrieb objects — Schicht, Auftrag, Freigabe — not a chat paste. Agents propose; you keep the switch. Ontology, not a demo.
ReadSwitchboard, chat skin, or war-room OS. Mid-market needs the object model.
Switchboards route. Chat skins paste PDFs. War-room OS draws the picture. Mid-market and legal win on one living object model on German metal — with a retainer.
ReadSensitive data needs a case graph. Not another chat skin.
Client matters and files do not belong in consumer chat. Residency, citations, a deletable case graph — lane A on STACKIT; US frontier only labelled non-sensitive.
ReadPain spikes at the second site.
The second site is where the three-part pattern bites. Agents on Spectacle take the redecide loop — Freigabe stays human. Pattern, not a beachhead.
ReadWithout Freigabe the agent sleeps.
On Spectacle, agents propose. Without Freigabe nothing runs. Proposal is not release — you keep the switch.
ReadTime back. Not ROI theatre.
The commercial want is time back in the operation — not invented hours, ROI, or accuracy bands. Agents on Spectacle take Schlepperei with Freigabe.
ReadThe three-part pattern. Not a sector SKU.
Spectacle fits where demand is uncertain, capacity is fixed in the short run, and labour and material must be committed before demand is known. Pattern, not a sector. Agents on Spectacle — Freigabe.
ReadAbove the systems. Not instead.
Spectacle sits above ERP, Warenwirtschaft, and Zeitwirtschaft. It does not replace them. Agents propose; Freigabe releases.
ReadFreigabe holds the switch. You keep it.
On Spectacle, actions that matter wait on a person. Agents propose; Freigabe releases. You keep the switch.
ReadOne core. Every kind of business.
One core on Spectacle. Six equal bands. No beachhead product. Agents take operational work; you get the time back.
ReadWareneinsatz is a decision. Not a formula page.
Committing Wareneinsatz and labour before demand is known is a decision you release — not a calculator. Agents on Spectacle propose; Freigabe decides.
ReadYou need time back. Not another Kennzahlen wall.
The want is time, not more Kennzahlen. Agents on Spectacle take operational work the Inhaber still does by hand — with Freigabe.
ReadWhen an AI workshop is worth it
When an AI workshop is worth it: map the handwork first, then AI agents on tools you already use — computer use, skills, workflows, prompts, routines. Freigabe stays human. Not ChatGPT 101.
ReadAgents take the work. You get the time back.
Agents on Spectacle take operational work the Inhaber still does by hand. Time back. Human Freigabe. Not another dashboard.
ReadThe beachhead is not the market
Venue ops can be first where the loop is real and measurable. That is a beachhead. It is not the market, and it is not the product category.
ReadTonight's Schicht does not move
The Dienstplan is posted. The numbers arrived. If tonight's shift is unchanged, you are still reporting.
ReadA number without a next action is still reporting
Operators already have numbers. A weekly file is still reporting until someone can release a named action that changes the floor.
ReadA dashboard is not a decision
Operators already have numbers. What is missing is the step after the chart: a recommended action, a release, and a change on the floor.
Read