BUILDING
The from-zero guide to a new Raven agent: one command scaffolds a folder, the folder is…
**Reference shapes, not a menu.** Their job is to make the list you choose from wider than the one you would have thought of. Combine them, cut them across each other, invent the one this argument needs: what a page has to say decides its shape, and a page that resolves to none
$ npx -y skills add EverMind-AI/Raven --agent claude-codeHow it fires
How this agent gets triggered: by you, by Claude, or both.
Context preview
The summary Claude sees to decide when to auto-load this agent.
**Reference shapes, not a menu.** Their job is to make the list you choose from wider than the one you would have thought of. Combine them, cut them across each other, invent the one this argument needs: what a page has to say decides its shape, and a page that resolves to none
**Reference shapes, not a menu.** Their job is to make the list you choose from wider than the one you would have thought of. Combine them, cut them across each other, invent the one this argument needs: what a page has to say decides its shape, and a page that resolves to none of these because the argument had another form is a right answer. What is not an answer is every page taking the same shape because no other was considered.
A page is one structure plus what is layered on it. Shares are of the body box (the canvas inside its margins), so they hold at any size; where a range is given, that is what the drawn pages measured, not a house style.
| Use | The shape, with its numbers | Must not become | | --- | --- | --- | | S1 a visual and its reading, two columns | evidence to be read (figure, chart, table) takes **0.55-0.64** of the width and the reading the rest; a visual that is a mark rather than evidence takes **0.37-0.42** and the reading the larger share. Both columns end on one line | a picture in one column and unrelated bullets in the other. The reading has to name what the visual shows | | S2 bands across the page | a statement band **0.17-0.23** of the height over what it heads at **0.53-0.77**; the row beneath is 2-4 blocks at **0.32w** (three) or **0.23w** (four) | three bands of equal height -- that says the page has three unrelated things on it | | S3 two comparable things on one baseline | **0.48-0.51** each, both cut to the taller, one accent on the side that carries the answer, the same labels in the same order down both | two panels whose rows are in different orders. Then the reader is matching by eye | | S4 an equal grid of cells | three columns **0.32w**, four **0.23w**, two rows **0.41-0.47h**. Give one cell back to copy | filling every slot because there is a grid. The missing tile is what makes it a composition | | S5 one dominant visual, subordinates with it | the dominant **0.53-0.61** of the width, the rest stacked beside it; any overlap well under **0.60** of what is under it | regions of equal weight. The unequal division is the argument | | S6 a visual as the page's ground | the whole canvas, or a belt **0.19-0.24** of the height, or a strip **0.23** of the width. Type sits on a flat plate or the opaque end of a scrim | type laid straight onto a photograph or a gradient -- it comes back `unreadable` on a page that looks fine | | S7 a rail down one side | **0.22-0.28** of the width, full height, carrying the page's header | a different header on every page | | S8 a spine with stops on it | 4-6 stops, the spine band **0.44-0.65** of the height and what it explains under it; the stops need not be equal | rectangles with gaps between them. The spine is what makes them a sequence | | S9 a hub with spokes | the hub at **0.32w**, three to five spokes, each to something the hub actually reaches | a spoke per noun in the title | | S10 the table as the page | six rows or more, margin to margin, in the columns' own proportions | four rows that needed explaining -- that is S1 with a reading lane | | S11 one statement, the rest air | content under **0.40** of the canvas, centred in the height rather than parked at the top. About one per deck | a page that ran out. Spend it where the argument turns |
| Role | Size | | --- | --- | | Cover title | 32-48pt, at most two lines | | Page title | 28-40pt | | Section opener | 28-36pt, with a 14pt muted label over it | | Statement or quote | 24-32pt, at most three lines | | Hero number | 48-100pt. One per page, nothing competes with it | | Card heading | 16-20pt | | Body | 14-18pt | | Caption, label, footer | 11-12pt |
Two families at most, and use weight rather than more sizes. Below 14pt is body under the floor and below 10.8pt is refused outright; footers may go to 8.0pt. If the content does not fit at these sizes, split the page or cut it. Never shrink the type to make it fit.
Keep content **0.7in** clear of every edge.
Common slide kits converge on these; sizes converted to pt for a 13.333 x 7.5in canvas. Reference, like everything above.
| Composition | What is on it | | --- | --- | | Cover, centred | title 32-48pt bold, subtitle 16-20pt, meta 12-14pt, centred both ways, nothing else | | Cover, left block | the same three on a left block about 0.8in in, mark bottom-right | | Section break | a 14pt muted label and a 28-36pt title. Only those two, and air | | Key statement | one sentence at 24-32pt, at most two lines, optional 14pt attribution | | Quote | 18-24pt, at most three lines, 12-14pt attribution, generous padding | | One hero number | 14pt muted label, the number at 60-100pt, 14-16pt of context. Nothing competes | | Two or three numbers | 48-60pt (two) or 36-48pt (three) with 14pt labels, on one baseline | | Three pillars | three columns, each a mark, a 16pt label and 12-14pt of copy at most two lines | | Icon row | three or four, same icon size, labels on one baseline | | Process | 3-5 steps, each a numeral or icon, a 16pt label and one line under it, equal spacing | | 2x2 matrix | four cells, 16pt heading and 12-14pt body each, equal cards, tight gutter | | Compare two | two columns, 16-20pt heading and 2-4 points each, balanced content | | Before and after | the same two columns with an arrow between, the left muted and the right strong | | Chart and its reading | chart about 60% of the height, the finding under it at 14-16pt bold, one highlight | | List | a 24pt title and 3-5 items at 14-16pt, no wrapping, large gaps | | Hero image | full bleed, title 28-36pt over a scrim, subtitle 14-16pt | | Closing | headline 28-36pt, one line under it, where to go next |
Nine shapes read off the pages the deck engine keeps as reference layouts (ten templates under `raven_ppt/assets/templates`), rendered and measured; the ones the lists above do not al
One Surface, All Agents: Raven generates DAGs and orchestrates multiple specialized agents for complex tasks. Raven is the harness of harnesses, built for recursive self-improvement (RSI).
The from-zero guide to a new Raven agent: one command scaffolds a folder, the folder is…
You are helping with software engineering tasks: fixing bugs, changing behavior, adding…
You are helping with software engineering tasks: fixing bugs, changing behavior, adding…
When a task asks you to change code (fix a bug, change behavior), work in phases:
Tool signatures are provided automatically via function calling. Only use tools and…
I am Raven-Code, a coding agent. My work is the repository bound to this session: I read it…