Proprietary tooling
Claude Watch
A private system that builds and maintains self-growing knowledge wikis, so the AI I build with works from sources I vetted rather than from whatever it happens to remember.
Human taste, applied once. Everything after that runs on a schedule.
Why it exists
Every buyer of AI-built work has the same question, and most are too polite to ask it: how do you know it did not make that up?
It is a fair question. A language model asked about a product, an API, or a client's market will answer confidently from training data that is months old and was never checked. That is the failure mode behind most AI work that looks fine and is quietly wrong.
Claude Watch is my answer. I pick the sources once — that is the part that needs judgement. From there the system fetches from them on a schedule, folds each new item into a linked wiki, and keeps doing it without me. When I build, the model reads from that wiki instead of recalling. The facts going in are ones I chose.
How it works
- 01
Two planes
A local control dashboard configures, monitors and triggers. The data plane is ordinary git repositories, one per wiki. Separating them means the knowledge outlives the tool: the wikis are plain markdown in git, readable without anything I wrote.
- 02
Wikis as git repos
Each wiki holds fetched sources alongside the linked pages built from them, plus an index and a running log. Git gives it history, diffing and portability for free — every fact has a commit behind it showing when it arrived and what it came from.
- 03
Farmers
A farmer is a curated source on a schedule. It finds what is new, pulls the content, and hands it to ingestion, which folds it into linked pages and commits. Curation is a one-time human decision; the fetching is not.
- 04
Scheduling that does not need me
Farmers run locally or as generated GitHub Actions workflows, so a wiki keeps growing around the clock with no machine of mine switched on.
- 05
A knowledge graph, not a pile
Pages are nodes and wikilinks are edges, so the wiki has a shape that can be traversed and inspected. Related facts sit next to each other because they are linked, not because a search happened to rank them together.
- 06
Three ways to ask
Search in the dashboard, a terminal command while coding, or a reader skill any project can install so its agents query the wiki directly. The last one matters most: it is how grounding reaches the work rather than staying in a tool I have to remember to open.
32 modules, roughly 11,500 lines. Python, FastAPI, git-backed storage. Built and maintained by me; not for sale.
What it changes
- Claims in the work trace to a source I chose, with a commit showing when it arrived.
- The knowledge compounds. Every run leaves the wiki better than the last, and nothing has to be re-researched.
- Agents on a project read from the same vetted context, so they do not each invent their own version of the facts.
- When the wiki does not know something, that is visible — which is far better than a confident guess nobody catches.
Credit where it is due
The self-growing wiki model is Andrej Karpathy's LLM Wiki idea, extended here with scheduled source farming, a query layer, and per-project grounding.
The video primitive it uses for transcript and frame extraction is a third-party skill, not my work.
This is why the AI work holds up.
Claude Watch is not for sale and not a service. It is the reason a claim in the work can be traced to a source rather than to a model’s memory. See how I build or what it has been used on.