15 · Figure index

All twenty-one illustrations in the pack, with what each shows, where it appears, and whether it depicts Anthropic's material or ours. Every figure is a standalone file you can lift straight into a deck.

How to reuse these

Each figure is an independent .svg in site/assets/figures/. They carry no external references and no webfonts, so a single file is the whole illustration — drag one into Keynote, Figma, a doc or another web page and it works. Each embeds its own prefers-color-scheme block, so it follows the viewer's light or dark setting without any page-level CSS.

Attribution, as everywhere in this pack

documented means the figure depicts Anthropic's own model, with their wording. synthesis means the structure or framing is ours — see §14. Check the tag before you put one in front of an audience and attribute it to Anthropic.

Foundations

One shared agentic engine branching into Cowork and Claude Code, with a Dispatch bar spanning both.
Fig 2.1 · shared-engineOne engine, two products. The shared core, and the four things that differ. §02 · documented
A three-stage cycle of gather context, take action and verify results, with an interrupt arrow.
Fig 3.1 · agentic-loopThe agentic loop. Why an agentic task costs many model calls, not one. §03 · documented
Cowork executing in an Anthropic VM bridged to your machine, beside Claude Code executing locally.
Fig 3.2 · execution-modelWhere the work actually runs. The desktop product runs in the cloud; the terminal product runs on your laptop. §03 · documented
Five lanes showing trigger, where Claude runs, and product, for Dispatch, Remote Control, Channels, Slack and scheduled tasks.
Fig 4.1 · away-from-deskFive ways to work while you're elsewhere. Read with the reconciliation note in §04. documented

Configuration & control

A Cowork project beside a Claude Code directory, with a dashed barrier reading Cowork does not read tilde slash dot claude.
Fig 6.1 · config-unitsThe unit of configuration. Two answers to "how do I teach it about my organisation" — and the barrier between them. §06 · documented
Four tiers — always on, on demand, isolated and outside — plotted against when they load, beside a finite context window bar.
Fig 6.2 · context-loadingWhat enters working memory, and when. Loading strategy as the real classification axis. §06 · documented
Two columns: request mechanisms flowing into Claude interprets, versus guarantee mechanisms flowing into blocked before the model acts.
Fig 6.3 · request-vs-guaranteeRequest vs guarantee. The sharpest distinction in the corpus, and the one that explains the Cowork governance gap. §06 · documented
Two permission ladders over a shared autonomy axis, with Plan off-axis and two flag-only modes below.
Fig 7.1 · permission-modesThe autonomy dial. All six Claude Code modes, including the two flag-only ones. §07 · documented

Risk & decision

An AND gate combining untrusted content access and consequential action capability to yield prompt injection is possible.
Fig 8.1 · two-condition-riskThe two-condition threat model. An AND, not a checklist — break one leg and choose which. §08 · documented
A decision tree from what are you trying to produce, branching to Claude Code, Cowork, Dispatch and plain Chat.
Fig 10.1 · decision-treeWhich product, decided by the artefact. Choose by the output, not the person. §10 · synthesis
Three columns for Chat, Cowork and Code, each with the kind of task it suits and why.
Fig 10.2 · surface-selectionAnthropic's three-way selection. Use this when you want their authority rather than yours. §10 · documented

Adoption & fluency

A five-step ladder from Level 0 Ask to Level 4 Department plugin, with a parallel axis showing what survives the session.
Fig 12.1 · maturity-ladderThe five-level adoption ladder. Levels verbatim; the durability axis is ours. §12 · documented + synthesis
Three phases — Evaluate, Pilot and Scale — along a six-month timeline with actions and expected signals.
Fig 12.2 · adoption-roadmapThe six-month roadmap. All wording verbatim. §12 · documented
The four AI Fluency competencies drawn as an inner fast loop nested inside an outer slow loop.
Fig 13.2 · four-d-loopsThe 4Ds as nested loops. Definitions verbatim; the nesting depends on a login-gated source and is unsupported — see the flag on §13. documented + synthesis
Automation, Augmentation and Agency panels showing when the human marker appears on each timeline.
Fig 13.1 · three-modesAutomation, Augmentation, Agency. Definitions verbatim, but the axis and the closing line are §14's thesis, not Anthropic's. §13 · documented + synthesis

Synthesis

These five carry the pack's original argument. They are the ones to redraw first if you rebuild this in your own visual language.

A two-by-two of generality against enforceability, with one-off prompts, CLAUDE.md and skills, single-tool ask rules, and hooks with deny rules and managed settings.
Fig 14.1 · decision-gridGenerality × enforceability. The two questions that place any mechanism in either product. Replaces the withdrawn judgment-relocation figure. §14 · synthesis
Four columns showing who judges each action across the permission modes: a human, a classifier model, a rule decided in advance, or nobody.
Fig 14.2 · who-judgesWho judges. Permission modes vary on the actor, not the timing. §14 · synthesis
A five-stop graded spectrum from local file edit to external send, with recommended oversight beneath each.
Fig 14.3 · reversibility-spectrumThe reversibility spectrum. Principle documented in §07; the graded scale is ours. §14 · synthesis
Five failure modes in rows, each with what you would notice and a countermeasure.
Fig 14.4 · failure-modesHow agentic work goes wrong. Anthropic documents one of these five. §14 · synthesis
A two-by-two of capability maturity against verification maturity, with stalled, cautious, healthy and exposed quadrants.
Fig 14.5 · trust-vs-capabilityTrust vs capability. The Exposed quadrant is the one nobody measures. §14 · synthesis
Six concept nodes with their section tags, all connected to a central node labelled generality by enforceability.
Fig 14.6 · concept-mapConcept map. Where each load-bearing idea lives in the pack. §14 · synthesis

Manifest

FigFileHeightUsed inAttribution
2.1shared-engine.svg364§02documented
3.1agentic-loop.svg220§03documented
3.2execution-model.svg310§03documented
4.1away-from-desk.svg324§04documented
6.1config-units.svg296§06documented
6.2context-loading.svg340§06documented
6.3request-vs-guarantee.svg384§06documented
7.1permission-modes.svg404§07documented
8.1two-condition-risk.svg316§08documented
10.1decision-tree.svg340§10synthesis
10.2surface-selection.svg276§10documented
12.1maturity-ladder.svg422§12mixed
12.2adoption-roadmap.svg328§12documented
13.2four-d-loops.svg372§13mixed — nesting unsupported, see §13
13.1three-modes.svg262§13mixed
14.1decision-grid.svg454§14synthesis
14.2who-judges.svg266§14synthesis
14.3reversibility-spectrum.svg244§14synthesis
14.4failure-modes.svg344§14synthesis
14.5trust-vs-capability.svg414§14synthesis
14.6concept-map.svg324§14synthesis

Technical contract

All conform to the mechanical rules in site/assets/figures/SPEC.md, verified programmatically. The spec's legibility guidance ("~12 text elements") is deliberately relaxed for the text-dense figures, which carry class="dense" instead. If you add a figure, match the mechanical rules.

RuleValue
CanvasviewBox="0 0 720 H"; no width/height on <svg>, so page CSS sizes it
Accessibilityrole="img" plus <title> and <desc>, referenced by aria-labelledby; alt text supplied at the embed site
ThemingEmbedded stylesheet with a prefers-color-scheme: dark block; background always transparent
PaletteMatches style.css. Rust = Cowork or emphasis · green = Claude Code or safe · ink = shared · amber = partial · red = blocked or irreversible
TypeSystem font stack, px sizes only, manual line breaks (SVG does not wrap)
Self-containmentNo scripts, webfonts, external images or xlink:href. No gradients, shadows or filters
IDsEvery id namespaced with the filename slug, so several figures can coexist on one page
Narrow screensText-dense figures carry class="dense" on the <figure>: below 640px they scroll horizontally rather than shrinking type below legibility
If you redraw these

The five synthesis figures are the ones carrying an argument rather than reporting a fact, so they are where your own visual language will pay off most. The documented ones are closer to transcription — redraw them for house style if you like, but check the wording against §16 before changing a single label, because several are verbatim quotes.