angle-board

<essential_principles>

What this skill is for

One coding session, one research file, one book, one argument — is never one article. It's a pyramid. This skill builds the pyramid and leaves it on disk as a backlog you mine for months.

It sits in the middle of a pipeline:

RESEARCH                  THIS SKILL              WRITER
session write-ups      →                       →
saved web clips        →   angle-board         →  your writing skill
book highlights        →   (angle board)       →  (or you)
(or: any notes, any topic, nothing at all)

Research produces notes. This skill produces an angle board. A writer produces prose. Each stage stops at its own boundary. This skill does not write articles and does not mine sessions.

The one rule that makes this work

Altitude has an evidence requirement, and it is not negotiable.

A single session can only honestly support the bottom of the pyramid — the weeds, the tutorial, maybe the project story. It cannot support a thesis, because a thesis with one proof point is an anecdote wearing a suit.

High-altitude angles are unlocked by the corpus, not by the current file. That is why the board must read the whole evidence pool, not just the one file it was handed. Three sessions across two projects and two months is what makes "here's what everyone gets wrong about subagents" a real post instead of a bluff.

So: if the evidence pool doesn't carry it, say so. Never manufacture a thought-leader take to fill a slot in the ladder. An empty rung is information — it tells the author what to go do more of.

The four axes

Every angle on the board is a point in four dimensions. Full definitions, the altitude ladder, the lens bank, and the evidence gate live in reference/axes.md — read it before generating any angles.

Altitude is fixed so you always get top-to-bottom coverage and the pyramid shape is guaranteed. Lens is sampled so every board doesn't come out as the same seven articles with the nouns swapped.

Structure/form is not an axis here — that's a downstream decision, once a brief is picked.

The setup — three things to resolve before generating anything

The four axes are voice-agnostic. The outlet is a parameter, not a law. The same generation works for your own blog or for articles you're pitching to someone else. Before generating any angle, resolve these three inputs and say what you resolved:

Optionally there's a writer handoff: the skill (or person) that turns a brief into prose. If the project has none, the skill stops at the brief and says so.

What the setup never changes: the four axes, the altitude ladder, the evidence gate, the pillar/spoke map, the don't-restate constraint. Those are the skill. Voice affects how an angle is pitched, never whether it's honestly supported. </essential_principles>

Mode — from how the skill was invoked:

• A path to a research/notes file (or "use the notes from X", or invoked right after a miner skill ran) → Build Board, subject mode A
• A project or topic name (/angle-board Project Name) → resolve it against the corpus first, then Build Board, subject mode A
• A bare topic with no material ("angles on why subagents suck") → Build Board, corpus-only, subject mode A
• An existing tutorial, technique, or capability to fan out — "invert this", "what could you use this for", "N posts out of this one how-to", a vendor tutorial URL, "same technique, different use cases" → Build Board, subject mode B (see reference/axes.md Axis 1)
• "Write #2" / "let's do the pillar" / a pick against an existing board → Pick and Brief
• No args → list existing boards in the resolved output folder with their pending counts, ask whether to resume one (Pick and Brief) or build a new one (Build Board)

Modes A and B are subject-derivation modes, not separate workflows — both run build-board.md, which branches at Step 3. A board may legitimately use both (material you produced and a technique you haven't applied yet); say so if it does. </intake>

Both workflows depend on reference/axes.md. Read it first, every time. Read reference/board-format.md before writing a board file. </routing>

<success_criteria> Working when:

• The setup was resolved and stated: voice, evidence pool, and output path were all named before any angle was generated, and no other author's voice was borrowed
• The corpus was actually scanned — the board cites evidence from files other than the one it was handed, wherever the material supports it
• In subject mode B, every subject is a payoff/actor/signal triple and no subject's signal is "look at the data"
• All five altitude rungs were considered, and empty rungs are stated as empty with a reason, not padded
• Every angle names a specific reader in a specific moment — never "developers"
• Every angle carries an evidence rating and concrete pointers (session beat, commit, error, file, number)
• Lenses vary across the board and at least one is not from the seed bank
• One pillar is designated; spokes are linked to it and carry a "don't restate" constraint
• The board is written as a markdown file with real checkbox task lines, so any task tool (or a plain editor) can track it
• Nothing was written as prose — handoff to a writer is the only path to an article </success_criteria>