[RFC] What should human-context actually become? Positioning + v1 direction

[zzgiabaozzbui]

I built human-context as a single-file tool to read large markdown knowledge bases for AI workflows.

It started as a personal utility.

Now I'm trying to figure out what it should actually become — and I genuinely don't know yet.

So instead of pretending I have the roadmap figured out, I'm asking the people who might actually use it.

---

A workflow it already supports — and the gap it exposes:

1. Open your project docs folder
2. See which files are token-heavy at a glance
3. Select only the docs relevant to your task
4. Copy optimized context into Claude / Cursor

The gap: most AI workflows have no idea how many tokens they're consuming until the model cuts off — or until the quality degrades silently. Large docs are the silent killer of prompt quality.

---

1. What should this tool be?

Option A — AI Context Workspace
A tool that helps you inspect, budget, and package your markdown docs specifically for AI agents. Core focus: token awareness and context quality before you hit send.

Option B — Local-first Markdown Viewer
The best offline markdown reader that works with zero install, zero sync, zero friction. Focus on UX, search, and rendering. One HTML file, forever.

Option C — Something else entirely
I may be looking at this wrong. Tell me what you see.

All options are genuinely on the table. I'm not attached to any of them.

---

2. V1 features — what gets built first?

One hard constraint: this tool must stay a single HTML file with no server and no install. That's not negotiable for v1 — it's the whole distribution model.

Within that constraint, here's what I'm considering:

A — Token heatmap per file
Color-coded file tree: red = heavy, green = light. Large docs silently destroy prompt quality — this makes the problem visible.

B — Context flow builder
Select files + set reading order → get exact token count for that sequence. "If Claude reads these 4 files in this order, it costs 42k tokens." Know your cost before you pay it.

C — Total budget bar
Running total against a target (e.g. 200k). Live as you select files. Simple, but changes how you think about a folder.

D — Stale file detection
Flag files not updated in 3+ months. Outdated docs are worse than no docs when fed to an AI.

E — Prompt Pack Export
Select files → export as a formatted, ready-to-paste context block for Claude / Cursor / Gemini. Structured, not raw.

Feel free to suggest features not on this list.

---

3. The question I actually want answered

Not "which feature should I build" — but:

When do you lose track of what you're feeding your AI? What does that moment look like?

I'm still exploring the direction, so raw opinions are genuinely helpful.

[richardchen874-sys]

The RFC] What should human-context actually become? Positioning + v1 direction is a better routing anchor than raw model price, because the real cost is whether the agent completes the run cleanly.

I would log route choice, provider/model, tool-call support, latency, usage, and retry/fallback reason together. Agent failures are much easier to debug when the route decision is visible.

This is close to what I am experimenting with: official Chinese models behind an OpenAI-compatible multi-model layer, with an emphasis on predictable usage and routing behavior. For human-context, is the harder problem provider compatibility, routing quality, or keeping per-run cost predictable?