Generate with Claude, keep the record: AI orchestration with a paper trail
Claude can now run your generation tools - Higgsfield, Comfy, and more - from a chat. The Archibal connector asks Claude to log each of those jobs into your project: tool, model, prompt and, when Claude passes its link, the file. Here is how it works and what it honestly can and cannot record.
In one sentence: the Archibal connector asks Claude to record each job it runs - tool, model, prompt, job id and, when Claude passes its link, the file - into your project, so the session's work persists as a production record instead of dissolving into a chat transcript.
A new kind of production session has appeared in the last year: you sit in a chat with Claude, and Claude runs the tools. It calls Higgsfield for a key visual, sends a workflow to Comfy, iterates on a prompt until the shot lands. It is fast, it is genuinely useful - and by default it is the least documented workflow in AI production.
Think about what survives a session like that. The files end up in a downloads folder. The prompts live inside a chat transcript you will never find again. The model that made each image - was the good take Nano Banana or Seedance? Which seed? Nothing in the file will tell you, and in a month, nothing in your memory will either.
The Archibal connector narrows that gap at the source: it asks Claude itself to file the record while you work.
What connecting actually does
Setup is one URL. In Claude's settings (claude.ai, the desktop app, or mobile), open Connectors, add a custom connector, and paste your workspace's connector URL - Archibal shows it on the capture page. Sign in with your Archibal account, click Allow, done.
From that point Claude has a small set of Archibal tools, and - more importantly - standing instructions that arrive with the connection itself:
- Attach a project. "Use the Spring Campaign project" pins this connection session to a workspace project, and the jobs logged in it file there. A new chat asks again; anything logged without a project waits in your workspace Inbox.
- Log every job. When Claude runs a job through another connector - Higgsfield, fal, Runway, Kling, ElevenLabs, Magnific, Adobe, Gemini, Replicate - it logs it twice: once when the vendor accepts the job (tool, model, prompt, job id), and again when the render resolves, repeating the job id, tool, model and prompt with the result URL. The model is attributed to the route it actually came through. You do not have to ask - the instructions do.
- Pull the result files. Where the generating service returns a result URL, Archibal downloads the actual file and runs it through the same pipeline as every other capture: scanning, fingerprinting, model attribution, and linking near-identical files.
- Use project files as inputs. Claude can list your project's assets and fetch them through short-lived links - so "take the packaging shot from the project and generate a blue variant" works, and the output is recorded with a lineage edge pointing back at the input it came from.
The result reads like a session log because it is one: timestamped entries, each with a model, a prompt, and a file - the same production record every other Archibal capture channel feeds. And unlike the chat transcript, which is only useful as long as you can find and re-read it, the record is structured and independent of the session: searchable by model or prompt, attached to the actual files, and still there when the conversation is long gone.
The honest part: what this channel can and cannot claim
Archibal grades every capture by how the evidence arrived, and a Claude-orchestrated job is graded like what it is: a studio log - an orchestration system reporting work as it happens. That is strong evidence of process. It is not the same as a file arriving with an embedded workflow graph, and the record never pretends otherwise. (The four evidence grades are explained in full on the capture page.)
Two more honest limits worth knowing:
Instructions steer, they do not compel. The connector tells Claude to log everything; it cannot force a tool call. In practice the automatic logging is reliable, but the right habit in your first sessions is to glance at the project's capture feed and confirm the session is landing. If a job was logged without its file - some services withhold result URLs from certain tool paths - asking Claude to log the completed job again, with the same job id, tool, model and prompt plus its result URL, attaches the file to the existing evidence row rather than creating a duplicate.
Failed and blocked attempts are part of the record. A generation that errored, or was refused by the provider's content filter, gets logged file-less. That is deliberate. A record containing only the successes is a highlight reel, not evidence - and when someone later asks what was attempted, the ledger has the honest answer.
Why agencies end up caring
For a solo creator this is convenience: your work stops evaporating. For an agency it compounds into something bigger.
Every Claude session any team member runs feeds the same searchable library. The junior who generated forty takes on a Tuesday leaves a record the senior can search on Friday - by prompt fragment, by model, by project. When the client asks "which model made the hero image?", the answer is a lookup, not an investigation. And when the work ships, the delivery carries the whole chain: signed files, a provenance record (PDF), the models on record, and a public verify link the client's procurement team can check without trusting anyone's memory.
The chat-driven workflow is not going away - it is too productive. The only question is whether it leaves a record. One pasted URL decides that.