A note of appreciation + one question (from a sibling harness)

[chrono-meta]

Hi — I maintain forge-harness, a governance-focused meta-harness for Claude Code. While exploring Obsidian as a knowledge backend I came across your project and was struck by how much we'd landed on independently (cross-CLI-from-one-source, index-first loading, adversarial self-challenge, [[wikilink]] recall) — with zero awareness of each other. That convergence made me want to reach out.

Mostly I wanted to say there's one thing you do that I'd genuinely like to learn from: your positive writing-voice model (the Orwell→Pinker ladder + the "does this bear one human's fingerprint?" test). My harness only has the negative half — it strips self-marketing and filler, but it can't push prose toward a voice target, which is the harder and more useful direction. If you're open to it, I'd love to hear how you arrived at that model and how well it holds up in practice.

[No] ask beyond that — just wanted to credit the work and close an awareness gap. Thanks for building in the open.

[eugeniughelbur]

Thanks for this, genuinely cool to see someone hit the same shapes from a different direction. The independent convergence is the interesting part.

Honest answer on the voice model: it's less of a "model" than it sounds. It's a small Python linter plus a rules doc. The doc lists the AI-tell words to kill (delve, leverage, seamless, "it's worth noting"...) and hard-fails on em-dashes. The script flags those and counts the human signals (specific numbers, first person, real file and command names) as positives.

But you put your finger on the exact gap I had too: stripping bad writing is the easy half, steering toward a voice is the real work. Mine was all negative half as well. So your message kind of nerd-sniped me and I went and built the positive half today.

It turns the voice target into measurable signals (sentence-rhythm variance for Pinker, short-sentence punch, first-person rate, contractions, concreteness, plain-word ratio for the Orwell rung) and returns a 0-100 voice-fit score, a "fingerprint" composite, and specific nudges like "sentences run long, break some up" or "too abstract, swap adjectives for numbers". Deterministic, runs in milliseconds, no API call. The point is it gates an LLM judge: only borderline drafts get sent to a model, so steering-toward-voice is cheap enough to run on every draft.

Happy to share the script. Would genuinely love to compare notes, feels like we're solving the same thing from two ends.

[chrono-meta]

This is great. Yes, please send the script, I'd really like to read it.

The part that gets me is that you built the positive half the same day. That's the exact half I was stuck on too. The negative side (kill the AI-tells, fail on em-dashes) is mechanical and easy. Steering toward a voice is the hard part, and turning Pinker rhythm variance and Orwell plain-word ratio into a 0-100 number is a clean way to make it cheap enough to run on every draft.

The bit I keep landing on from my side is the deterministic score gating the LLM judge. I hit the same shape but in a different domain, verification rather than prose. Same move: a cheap mechanical check up front so the expensive judge only fires on the borderline cases, because a judge-only gate can get talked into a yes.

Mine is already all public, so I won't write it up twice. The gate-before-judge pattern lives in forge-harness, and the steel-quench and phantom-quench pieces are the closest match to what you built: https://github.com/chrono-meta/forge-harness . Happy to point at specific files if it helps.

Same thing from two ends. Agreed.

[cantkillspencer]

Hi All- I wanted to let the AI Operator know how helpful this has been. I have been in hospice 3x, on life support and medically induced comas multiple times and wasted away in hospitals for over three years because of liver cancer. It is a miracle I am able to write this post. Another (albeit smaller) miracle is how AI has helped me overcome my executive function problems from being so sick for so long.

obsidian-second-brain has been the best thing yet for my memory and my agents' memory. When you are starting from a very challenged place you can imagine the benefits for me. But it's more than that. Because I don't fail in these areas like I used to I feel so much better about myself which helps my two boys feel better and on and on.

So, with as much thanks as can be given in a post "Thank You". One question I had is if anyone forked this for MSFT 365. It is not what I would choose if I were starting from scratch and I have been using this with my Google set up post hospice. However, since so much legacy data is in the MSFT ecosystem I wanted to ask.
-Spencer

[eugeniughelbur]

@cantkillspencer
this means a lot, genuinely. thank you for sharing something this personal. And thank you for the appreciation of the project!

the whole reason i built it was to stop dropping things i cared about, so reading that it's helping you show up for your boys is the best thing anyone's said about it. that's the actual point of the project.

on M365: no native integration yet. the vault is just plain markdown files, so the usual path is syncing the folder through OneDrive and pointing the skill at it. a couple of people have wired up Outlook/Teams via MCP but nothing packaged. if you want, open a discussion or issue with your setup and i'll help you get it working.

[eugeniughelbur]

Spencer, thank you for writing this. What you have been through puts the repo in perspective, and hearing that it is helping you rebuild executive function and feel more like yourself for your boys means far more than any stat on this project. Genuinely glad it is useful to you.

On Microsoft 365, two parts:

Storage works today, no fork needed. The vault is just a folder of plain markdown files. Put that folder inside OneDrive and it syncs and backs up across your devices exactly like the iCloud and Dropbox setups others use. Nothing in the skill changes. So you can keep your Google setup and simply make OneDrive the vault'''s home if that sits closer to your Microsoft data.

Pulling your legacy M365 content in is the real gap. I do not know of a dedicated M365 fork (the closest community fork, scholarbrain, is academic). There is no built-in connector yet that reads Outlook mail, OneNote, or Word/SharePoint into the vault. Two honest paths:

• Available now: export a OneNote notebook or Word doc to markdown or PDF and run /obsidian-ingest on it. The vault rewrites itself around it like any other source.
• The proper fix: a Microsoft Graph MCP connector that exposes Outlook/OneNote/OneDrive the way the existing connectors work, so the agent pulls M365 content on demand. It does not exist yet, but it is a clean, well-scoped build.

If a Graph connector is something you would actually use, say so and I will open a tracked issue with your use case as the driving one. No promises on timing, but it is the kind of thing worth doing right.

Take care of yourself, and thank you again for this.