Epictetus
3a42f53333
- docs/vision/VISION_LOCK_V1.md: permanent editorial mission, audience (builders/tinkerers/
local-AI users/home-lab/agent devs), Editorial Test ('what would an enthusiast DO?'),
architectual constraint (editorial vision is primary system), priority order.
- docs/vision/ADR-0007_EDITORIAL_PIVOT.md: accepted pivot record, before/after table,
Sprint 1 evidence (high-signal = actionable, low-signal = spectator content).
- docs/vision/EDITORIAL_GUIDELINES.md: operating rules; score is draft, taste overrides;
hand-before-machine; memory must store 'what can be done' not 'what happened'.
- docs/vision/SECTION_DEFINITIONS.md: 5 sections (What Shipped / Run It Locally /
Benchmarks & Builds / Problem Solved / Worth Trying Tonight).
- docs/vision/SPRINT_1_FINDINGS.md: reference evidence, 45/200 UNCATEGORIZED taxonomy gap.
- Issue_001.md: handcrafted prototype edition from Sprint 1 stories; benchmark for all future automation.
This is the architectural directive. No scoring/narrative/memory work proceeds except in
service of the Vision Lock. Sprint 2 (Lens) stays blocked pending manual review tallies.
2.1 KiB
EDITORIAL GUIDELINES — AI Enthusiast Daily
Derived from VISION_LOCK_V1.md and ADR-0007. These are the operating rules for every edition, human- or machine-assisted.
The One Test
What would an AI enthusiast DO after reading this?
If the honest answer is Nothing, the story does not belong in the publication. No exceptions for "it's a big deal" or "everyone's talking about it."
Publish Lean
Default toward stories where the reader walks away able to:
- install something (open-source tool, model, agent)
- benchmark something (real numbers, reproducible method)
- run something locally (Ollama, llama.cpp, GGUF, self-host)
- replicate a result (code, weights, config)
- learn a technique (production war story, failure post-mortem)
- build with it (agent infra, framework, pipeline)
Reject Default
Reject by default when the story is:
- A lawsuit or legal filing (unless it changes what a builder can ship)
- A funding round or valuation (money moved, no capability changed)
- A CEO / exec opinion or take (someone said something)
- A celebrity or consumer AI culture piece (glasses, ads, vibes)
- A corporate feature announcement with no reproducible substance
- A pure market / stock / IPO item
Section Fit (see SECTION_DEFINITIONS.md)
Every published story must map to exactly one section. If it fits none, it's either UNCATEGORIZED-by-design (hold) or a reject.
Score Is Not Taste
The final_score from Sprint 1 is a draft signal only. It was built to be debugged, not trusted. Editorial judgment — the One Test — overrides any number. A 0.50 local-AI benchmark may ship; a 0.29 culture piece may not.
Hand-Before-Machine
No automation may publish. Automation may propose; the human editor confirms. The founder rule stands: "Publish? [Y/N] — Not algorithmic. Human review only."
Memory Rule
When future memory / knowledge-graph components land, they must record what a reader can do with a development — not a summary of the development. "Qwen3.6-27B runs on 4×5060Ti at INT8" is memory. "Qwen released a model" is not.