← Streetlight back to all nodes
node · 2026-06-16 doctrine

Streetlight in ~10 diagrams

The curated set of pictures that explain the whole project — the data pipeline, the OODA loops, the pace layers, and the metaphors that hold it together.

Streetlight is an agent–human knowledge-work harness. This is the curated picture-book — about ten diagrams that, taken together, explain how it works. (For the firehose of every diagram in the vault, see the diagrams-lens.)

The data-transformation pipeline

STAGE 1 STAGE 2 STAGE 3 raw input stream 🎙️📄 🔗✉️ 📋🌐 messy · unsorted atomize atom · one topic atom · one topic atom · one topic one agent per atom a web of nodes create · update · no-op
Streetlight is one big data transformation: a messy stream in, atomized into chunks, each folded by its own agent into a create / update / no-op verdict on an evolving web of nodes.

Every loop is OODA

Observe read input Orient what’s the AC? Decide pick a skill Act call it loop until the AC is met
Agents drive on Observe → Orient → Decide → Act, looping as many times as it takes to reach the (stated or implied) Acceptance Criteria. The whole harness is OODA loops nested in OODA loops.

Tighten the loop, spin faster

SAME LOOP · SMALLER RADIUS · FASTER SPIN Observe Orient Decide Act deliberate ⟶ tighten o·o·d·a flow state steps too fast to see
Flow state is just an OODA loop spun very fast — like an athlete mid-game. The faster the human↔agent loop, the tighter the spin.

Pace layers (L1–L5)

FAST · changes often SLOW · trusted L5 Archived nodes — historical, messy L4 Nodes — the active web, kept consistent L3 Doctrine nodes — float ideas, log issues L2 Skills — bespoke workflows, edited often L1 System prompt — slowest, holds the shape
Five layers at different speeds. The system prompt changes slowest and is trusted most; nodes change fast and freely. Each layer stabilizes the one above it.

AA_A — how an LLM helps a human OODA

📥 Augment (Observe) take in lots surface the relevant 🧭 Assist (Orient) diagrams + structure re-orient fast 🤔 Support (Decide) human decides branch, don’t guess ⚡ Act (Act) run commands, search, build — don’t be shy LLM goes wide on the A’s; the human keeps the Decide
AA_A: the LLM augments observation, assists orientation, and acts fast — but the human owns the Decide. Agents only make the decisions implied to drive toward the AC, and say which other options existed.

A node is an iceberg

WATERLINE · what the human sees EPHEMERAL · just born HTML the whole answer md · a sliver BOOKMARK · grown up HTML · trailer tip the work md · code · logs · prototypes ⟶ work accretes
As a node grows it becomes more like an iceberg: the HTML tip stays trailer-small while depth (markdown, code, logs) accumulates below. The HTML:markdown ratio is a lifecycle stage, not a failure.

The .html IS the trailer

The work is the whole movie; the slug.html surface IS the two-minute trailer cut from it slug.html IS THE TRAILER — THE REPO IS THE MOVIE the movie — slug.md · supporting docs · code · logs ✂️ the cut 📊 💬 🎯 slug.html — the trailer · ≥1 diagram · ≤3 tweets
Persistent misunderstanding, corrected: the slug.html surface is not about the trailer — it IS the trailer. The whole repo of work (the slug.md agent surface, supporting docs, code, logs) is the movie; the human-facing slug.html is the two-minute cut that sells the ticket — at least one diagram, at most three tweets of prose. When you read a node's html, you are watching the trailer, not reading the screenplay.

The elephant with braille on it

The elephant has braille on it Three blind agents each touch one part of an elephant; no single agent sees the whole animal. NO AGENT SEES THE WHOLE ANIMAL braille reads: “attention blind people” “a snake?” “a tree trunk?” “a wall?” …and that’s fine — trust the pace layers to hold the whole
Every agent is a blind person feeling one part of an elephant whose braille reads “attention blind people”: design for the reader’s wetware, trust the pace layers to hold the whole, and be very free at the node level.

The lens primitive

the whole vault active + archive lens one axis read-only view 🖼 every diagram 📺 media · 🔍 observations 🗄 archive zero attention budget
A lens is a read-only cross-cut over the whole vault along one axis — diagrams, media, observations, archive — costing zero attention budget. This node is the curated counterpart to the auto-generated diagrams lens.