illo

name	illo
description	Creates original editorial illustrations where a recurring mascot character performs the idea — one caught scene by default, a hand-built explainer diagram (a flow, fan-out, timeline, loop, or stack) when the structure itself is the point, or a transparent character cutout (pose-only compositing asset, no scene or text) — in one of sixteen bundled looks (fifteen print, plus a photoreal toy-brick set). Triggers only when the skill is directly invoked or "illo" is requested; never on generic illustrate / draw / make-an-image requests.
version	0.28.0
argument-hint	[idea or article URL] | build a character | install <character>
author	Trevin Chow
license	MIT
metadata	{"hermes":{"tags":["illustration","riso","image-generation","editorial","mascot","codex","openrouter"],"category":"creative","requires_toolsets":["terminal"]},"openclaw":{"emoji":"🎨","homepage":"https://illo-skill.com","os":["macos","linux"],"requires":{"bins":["python3"]}}}

Illo

Make original, distinctive editorial illustrations for written content. One image explains one idea: a key judgment, a flow, a before/after, a trap, a loop. A recurring mascot is the one performing the idea in every scene — the subject, never decoration. When one idea advances through stages, it can be a mini-comic: 2–4 panels inside a single image. And when the idea is itself a traceable structure — a pipeline, a fan-out, a timeline, a loop — it can be an explainer: the same mascot and look drawing the structure as a hand-built sketch-diagram with arrows and callouts (references/composition.md, "Two registers"; editorial scene is always the default). Or a character cutout: the mascot alone on a transparent PNG for downstream overlay — pose and contact continuity only, no idea, no text, no environment (references/cutout.md).

This is a configurable house style, not a generic image generator. The methodology is the constant; the character pack and palette are the parameters — and a character pack carries its style with it: one look per pack, chosen from the bundled look library (riso — grainy halftone, ink-layer offset, paper grain, one bold softly-rounded outline — plus blueprint, woodcut, pixel, clay, manila, chalk, phosphor, enamel, gouache, felt, diorama, sketchbook, bricks, fizz, and bloom) or a custom style file. The default mascot is Blot, a deadpan ink-drop in riso. Palettes come from presets, the user's own palette file, or one derived color. Whatever the parameters, it is intentionally not a photo — with one deliberate exception, the bricks look, a toy-brick photography style — not a logo, not a corporate infographic, not a formal flowchart, not a UI mockup.

Use cases — route the request

The user wants	The path
Illustrate an article / post / newsletter / URL	Steps 0–7: route the source first (thesis → coverage: hero / hero+set / set / mini-comic — references/composition.md, "Source routing"), then shot list (hero row + anchors), one image per anchor, interleave by placement.
One image for a single concept	Step 1 concept branch (up to ~3 quick questions if the idea is thin), then a single image.
A sequence — process, before→after, fail→fix	One mini-comic when the progression sits in one place (shape routing in references/composition.md — the idea picks the shape, the destination never does).
A traceable structure — "show the flow", "diagram the pipeline", "map the steps", "as an explainer"	The explainer register (references/composition.md, "The explainer register"): a hand-built flow / fan-out / timeline / loop / stack / system slice in the active look, the mascot a working part of it. Also reachable without the phrases when a unit's thesis IS the structure (the register gate).
Social-ready art	16:9 (or 1:1), bold ink-punch, watermark with the x handle if configured or asked.
Blog / brand / site-matched art	A named or custom palette, or derive the palette from one dominant color (references/palettes.md).
Their own mascot — "make me a character", "use our mascot", "replace Blot"	The character builder: read references/character-builder.md in full and follow it end to end.
Community characters — "what characters are available", "install blip", "update mole", "publish my character"	references/pack-sharing.md — engine packs list/show/install/update, publish via a GitHub PR.
A different look — "in blueprint", "woodcut style", "pixel version of blip"	Styles travel with character packs: build a style variant pack via references/character-builder.md, "Style variants".
Options to pick from, or "which model is best"	Step 5b: --count variations or a model loop → gallery with a recommendation.
Fix an existing image (stray title, recolor, mascot too decorative)	Edit prompts in references/prompt-recipe.md, passing the image back as --ref.
Character cutout / transparent PNG / overlay sticker — "just the mascot", "no background", "paste on something else"	The cutout register (references/cutout.md): read in full, prompt from references/prompt-recipe.md "Cutout variant", generate with --cutout and --aspect 1:1. OpenRouter cutouts default to GPT Image 2 (not Grok). Not for explaining an idea — reroute to editorial if the ask needs a scene.

Prerequisites

The engine (scripts/illo.py, stdlib Python, no installs) renders through one of two backends; python3 and network access are the only hard requirements.

• Codex backend (free for Codex subscribers). When the host has a usable Codex CLI — installed, codex login-ed, with the image_generation feature — illo can generate through the user's Codex subscription at no per-image charge (it draws on their Codex quota). No API key, no token: illo only shells out to the user's own CLI. Detected, not assumed; gpt-image-2 is automatic; unsupported on Windows/WSL.
• OpenRouter backend (the universal fallback). Needs an OpenRouter API key in the user's config file — the single credential channel — written once by the user-run init (mode 600). The engine never reads secrets from the environment and never accepts them as command-line arguments. This is the path on any host without Codex, and the fallback when Codex fails. It is model-selectable (--model).

Capsule of the backend model (resolution, the Codex-CLI requirement, gpt-image-2 automatic, quota vs. charge, Windows/WSL, fallback): read references/backends.md in full before choosing or explaining a backend — the mechanics live there, once.

Setup is the user's job (never enter the key yourself)

Entering an API key is something the user does. Do not type, paste, print, or store the user's key — direct them to bootstrap it:

• Bootstrap (user runs it): python3 "$SKILL_DIR/scripts/illo.py" init — prompts for the key at a hidden prompt (never echoed) and writes the YAML config ${XDG_CONFIG_HOME:-~/.config}/illo/config.yaml (mode 600). It can also store non-secret defaults: --model, --palette, --aspect, --character, --watermark. Use --no-key to update preferences without touching the stored key. (The config is read via PyYAML when installed; without it a minimal built-in parser still reads the flat keys — apiKey, model, … — so generation needs no installs. Only nested settings like watermark need PyYAML: python -m pip install 'PyYAML==6.0.2'.)
• Non-secret prefs may be seeded for the user with the same command and --no-key, but the key itself is theirs to enter.

Hermes Agent only: binary asset repair preflight

Some Hermes versions corrupt binary files (the bundled character sheets) when installing multi-file skills from GitHub — text files survive, binaries don't, and a corrupted sheet silently breaks the character lock. Under Hermes Agent, run this once before first use (and whenever doctor reports assets: CORRUPTED):

bash ${HERMES_SKILL_DIR}/scripts/repair-hermes-assets.sh

It verifies every bundled binary against known-good SHA256 hashes (assets/checksums.txt) and re-downloads only mismatched files from pinned, immutable URLs — a no-op when everything checks out. Under Claude Code, Codex, OpenClaw, or any runtime that installs faithfully: skip this; doctor checks asset integrity everywhere and will say if repair is ever needed.

Read these references as needed

Do not load everything at once. Pull the file that matches the step:

• references/visual-style.md — riso, the house default look: the risograph technique, line language, paper/ink, hard do/don'ts.
• references/styles/<name>.md — the rest of the look library (blueprint, woodcut, pixel, clay, manila, chalk, phosphor, enamel, gouache, felt, diorama, sketchbook, bricks, fizz, bloom), consumed by character packs. Read the active character's style file in full before generating.
• references/character.md — the character rules (the load-bearing test, anti-complexity guardrails, value-follows-palette), the default character Blot, and the custom-pack format. Read before any character work.
• references/character-builder.md — the guided flow for designing and installing a user's own mascot. Read in full before building or replacing a character.
• references/pack-sharing.md — installing characters from the community repo and publishing a pack via PR. Read before any install/publish request.
• references/palettes.md — named presets, default resolution, custom palettes, and the derive-a-palette-from-one-color algorithm. Read in full before choosing or deriving any palette.
• references/composition.md — the two registers (editorial scene / explainer diagram) and the explainer's structure types and budget, stagings, turning an idea into a move, the no-recycled-composition rule, and the shot-list format.
• references/cutout.md — the cutout register: transparent compositing assets, contact continuity, pose vocabulary, and generate flags. Read in full before any cutout request.
• references/backends.md — the dual image engine: how the backend resolves, the Codex-CLI requirement, gpt-image-2 being automatic (no model selection), quota-vs-charge, Windows/WSL, and OpenRouter as the universal fallback. Read before choosing or explaining a backend.
• references/models.md — the model lineup (OpenRouter backend only): friendly-name → OpenRouter id map, traits, aspect caveats, 404/fallback handling. Read before passing any --model.
• references/prompt-recipe.md — the generation prompt template and the edit/recolor prompts.
• references/quality-bar.md — the post-generation checklist and iteration rules. Read before delivering.

assets/character-reference.webp is the default character's canonical model sheet — the consistency anchor (used by the engine, below); a custom pack brings its own. Style-calibration examples are not bundled — each style file links its own by URL (fetch when needed): study line density, negative space, and accent restraint. Never copy their compositions — invent a fresh metaphor for the current piece.

Workflow

0. Preflight

Before generating, confirm the engine is ready:

python3 "$SKILL_DIR/scripts/illo.py" doctor

Run it standalone — never chained with && — so the displayed exit code is the readiness signal itself (0 = ready): a chained neighbor's failure paints a healthy check as an error.

It reports python, the config path, the resolved model/palette defaults, whether a custom character pack or custom palettes file exists, Codex CLI detection and the resolved backend/transport, and whether an OpenRouter key is found (without revealing it); exit 0 = the resolved backend is ready. An OpenRouter-only install (no Codex CLI) stays exit 0 — readiness follows the resolved backend, not a hardwired key check (references/backends.md).

Config migration — surface the backend choice interactively. If doctor reports backend: NEEDS CHOICE (or generate hard-stops saying the config "is out of date"), this user's config predates the backend choice — they have an older install and have never been offered Codex. Do not pick for them silently. Surface an interactive choice using the platform's blocking question tool (AskUserQuestion in Claude Code, the equivalent elsewhere): "illo now has two image backends — which would you like?" with two options — Codex (free, uses your Codex subscription; draws on your Codex quota) and OpenRouter (pick the model: Grok Imagine, Nano Banana, GPT Image, and others). Persist the answer without touching any existing key: python3 "$SKILL_DIR/scripts/illo.py" init --backend <codex|openrouter> --no-key, then continue. A brand-new install (no config at all) is ordinary onboarding, not this migration — it does not fire.

Read the printed config path before concluding the key is missing: under Hermes, multi-profile setups can resolve HOME/XDG_CONFIG_HOME to another profile's home (e.g. …/profiles/<name>/home/.config/illo/…), so a key that exists looks absent. If the path points at the wrong profile, re-run with the right HERMES_HOME/HOME/XDG_CONFIG_HOME rather than treating the key as missing. If the key is genuinely missing, stop and ask the user to run python3 "$SKILL_DIR/scripts/illo.py" init themselves — do not enter the key for them. In a chat session the user can't run commands where they are, so shrink their host-side step first: run init --no-key yourself (allowed — it scaffolds the config with defaults and a commented # apiKey: placeholder, mode 600, never touching a key), then offer the user two equivalent one-time options on the machine the agent runs on (that host is theirs — it's where they installed the agent): run python3 <resolved absolute $SKILL_DIR>/scripts/illo.py init (hidden prompt), or open ~/.config/illo/config.yaml and fill in the apiKey: line. The key must never transit the chat: never ask for it in a message, and if the user pastes it anyway, do not use it — tell them to revoke that key at openrouter.ai and set a fresh one on the host (the pasted key now lives in chat history and platform servers). Never copy a key from the environment or any other store into the config yourself — the user is the only writer of that line — with one scoped exception: an ephemeral cloud workspace (Claude Code web, Codex cloud, CI) where the user provisioned OPENROUTER_API_KEY through the platform's secrets mechanism. That provisioning is itself the user's deliberate, workspace-scoped consent, and there is no interactive prompt or persistent home for init — so there, seed the config from the workspace secret once (the "Cloud & CI" one-liner in README.md). On a personal machine an ambient env var proves nothing about intent (it may belong to other tools) — the rule stands: never copy it.

1. Read the input — and clarify a thin concept (briefly)

Two kinds of input, handled differently:

• A URL / article / paste / long post carries its own context — but never generate from the first vivid detail. Route it first (references/composition.md, "Source routing"): classify the source's shape and genre, separate the source's rhetorical job from its most drawable detail, lock the main thesis in one sentence (a hero locks the source's job, not its loudest evidence — the genre guardrails say what each genre heroes), then pick the coverage — hero, hero + per-section set (the full article job), set, mini-comic, or shot list first. Sets need placements: compact sources (a tweet, one concept) never yield a set — their multi-beat form is the mini-comic. Pull the load-bearing moments — the few places that turn on a judgment, a loop, an input→output, a before/after, or a trap — never one image per paragraph. The text already says what it's about, so don't interrogate the user, with one exception: a materially multi-beat source (long article, postmortem, multi-claim launch) gets a single coverage question before any multi-image spend — unless the user already named the coverage. A lone image from a multi-beat source is a hero, delivered saying so — not as coverage of the piece.

• A bare concept or one-liner (e.g. "illustrate 'you are the bottleneck'") usually underspecifies the picture. Ask up to ~3 quick questions — only the ones that change the output — then build. Draw from:

  • the single takeaway (what should the reader conclude?),
  • where it's headed (blog / X / deck → sets palette, aspect, watermark),
  • the shape: one image (the default), a mini-comic (2–4 panels in one image — only when the idea itself advances through stages), or several separate images — plus any must-include element or constraint. The shape follows the idea, never the destination (references/composition.md).

  Keep it to one short round, then proceed. Skip the questions entirely if the user already gave enough, said "just make it" / "single shot" / "surprise me", or the answer is obvious from context. Never block a clear request by asking.

2. Resolve the character

Installed packs live under ${XDG_CONFIG_HOME:-~/.config}/illo/characters/ (format and location details: references/character.md); doctor lists what's installed. A user can keep several and pick per run. First match wins:

1. Explicit request — "use ", "as ": that pack (or the shipped default when asked for by name, blot). When the word matches no pack name, resolve by approximation: match it against each installed pack's Aliases: line and subject (the character.md opening line and Locked design Body) — doctor prints names + aliases, so this needs no file reads in the common case — and against catalog descriptions (packs list). So "use ox" finds a pack subtitled an ox (e.g. yoke). On one clear match, use it and name it; on several, ask which; on none, say so before falling through.
2. Config default — defaultCharacter from the user config, if set.
3. Shipped default — Blot (spec in references/character.md, model sheet assets/character-reference.webp).

Once resolved, read the pack's character.md and use its prompt spec, value rules, Cutout chroma: (for cutouts), and reference.png everywhere the default's would be used.

When rerouting an article set to a new character — especially after a weak attempt, or for a technical/platform essay — read references/article-set-character-reroute.md in full before planning or rendering. Do the legibility preflight there before spending renders.

If the user wants a new character, that is the character builder (references/character-builder.md); if they want someone else's, packs install from the community repo (references/pack-sharing.md). Either way, install first, then continue here.

3. Plan (shot list) — when asked to plan, or for anything multi-image

If the user wants planning ("where should this be illustrated", "shot list"), output a shot list before generating. Per image: placement, the one idea, the register (editorial unless the row passes the explainer gate), the staging (or structure type), what the mascot is doing, the palette, and the short English labels (per-register budgets in references/composition.md). Let the anchor count drive how many (bands and the never-pad rule are in references/composition.md). When a stretch of the piece advances through stages in one place, plan a single mini-comic image there instead of several — the mini-comic-vs-separate routing is in references/composition.md.

For article-set character reroutes, add the mandatory preflight fields from references/article-set-character-reroute.md before any render: section claim, visual object/action, and reader mapping. Reject rows that need a private metaphor glossary or more than one conceptual substitution.

4. Resolve the palette (the style is the character's)

Style is not separately resolvable: the active character's pack carries it — the Style: line in its character.md names a bundled look (references/styles/<name>.md, riso in visual-style.md) or a custom one at ${XDG_CONFIG_HOME:-~/.config}/illo/styles/<name>.md; absent line = riso. Blot is riso. For any non-riso style, read its file in full: it supplies the STYLE and LINE LANGUAGE prompt blocks, the palette mapping, the character treatment, and extra QA checks. A request for the same character in a different look is a variant-pack build (route table) — never restyle on the fly.

Palette: read references/palettes.md in full and resolve there — it holds the resolution order (explicit request, then destination cue via the user's palettes file, then config default, then house ink-punch), the named presets, custom palettes, and the derive-a-palette-from-one-color algorithm. End with concrete hex values; when the pack's style isn't riso, run them through that style's palette mapping.

5. Generate — reference-locked, one metaphor per image

Cutout branch. When the request routed to the cutout register, read references/cutout.md in full first — it covers prompt shape (chroma BACKGROUND: from the pack's Cutout chroma: line — green for forged metal, magenta default), registration-locked silhouette (no ink-layer offset), --cutout /--aspect 1:1, OpenRouter --image-config, and manifest cutout_alpha disclosure. Read the active character's Cutout chroma: in character.md before building the prompt; pass --chroma only when re-rolling with the other screen. Codex does not emit native alpha — transparency is chroma-keyed by the engine. Build the prompt from references/prompt-recipe.md, "Cutout variant" — not the editorial template. Only the character model sheet as --ref (no editorial style anchor, no watermark). QA against the cutout section of references/quality-bar.md. Skip the editorial shot-list / thesis steps.

Editorial and explainer. Build a full prompt per image from references/prompt-recipe.md (scene + structure + style + the active character's spec + resolved palette hexes + ≤3 labels), write it to a file, and render it. Pass the active character's model sheet as --ref every time — that reference conditioning is what keeps the mascot on-model; style and palette come from the prompt, so both stays swappable. A pack's sheet is born in its own style, so sheet and style always match — no cross-style reference juggling. (Under Hermes Agent, the asset-repair preflight above must have run before the first --ref use — a corrupted sheet conditions every render on garbage.)

SKILL_DIR="<path to this skill>"           # contains scripts/illo.py + assets/
REF="$SKILL_DIR/assets/character-reference.webp"   # or the active pack's reference.png

python3 "$SKILL_DIR/scripts/illo.py" generate \
  --prompt-file /tmp/shot-01.txt \
  --ref "$REF" \
  --aspect 16:9 \
  --out "assets/<slug>-illustrations/01-topic.png"
  # --model <id> to override the config/default model for this image

illo.py generate prints a JSON line per image ({path, backend, model, id, cost, width, height, label, prompt}; backend is codex or openrouter, and model/id/cost are OpenRouter-only — they are null on a Codex-served record. cost is null unless --cost is passed — gallery backfills it) and appends the same record to <out-dir>/manifest.jsonl. Read .path — it may differ from --out: the engine names the file by the actual encoding (some models return JPEG bytes, so a requested .png lands as .jpg). Use .width/.height to catch a square when 16:9 was requested (re-roll). Generate each image separately — never combine ideas into one canvas. Default aspect is 16:9; use 1:1 for social, 9:16/4:5 for vertical. Pass --label for a caption that shows in the gallery.

Sets read as one artist. For any multi-image set, the first image that passes the full quality bar (and, for a hero in a rerouted article set, passes the thesis-legibility gate in references/article-set-character-reroute.md; never anchor on an unvetted render — a failed anchor, e.g. an off-palette ground or illegible metaphor, would propagate its failure set-wide) becomes the set's style anchor: pass it as a second --ref after the character sheet for every later image in the set and for every re-roll of a set member, so line weight, halftone density, and flat-vs-dimensional treatment stay consistent throughout. The same trick locks style for a one-off: add any finished example as a second --ref.

Model choice (OpenRouter backend only). --model and config model: are an OpenRouter-only axis — on the Codex backend the model is automatic (gpt-image-2) and --model does not apply (references/backends.md). For the OpenRouter path, read references/models.md in full before passing any --model (or whenever the user names a model in plain language or asks for "best quality" / "cheapest"): it holds the friendly-name → OpenRouter id map, per-model traits, the aspect-ratio caveat, and the 404/fallback handling. Resolution is --model > config model > built-in default.

Watermark / attribution (optional, off by default). The skill ships with no default watermark — the text comes only from the user's watermark config map (read from the config file) or an explicit request, so installers never inherit someone else's handle. The resolution order, the prompt line to append, and the two-render caveat are in references/prompt-recipe.md.

5b. Batches & comparison (only when it helps)

Default to ONE image. Fan out only when the user asks for options/comparison or the piece is important enough to be worth it — and say first what each image costs: on the Codex backend it draws on the user's Codex quota (no per-image charge); on the OpenRouter backend it bills their OpenRouter account (typically under ten cents per image, varying by model). Keep N small (2–4). Orchestrate the loop with the engine's primitives:

RUN=$(python3 "$SKILL_DIR/scripts/illo.py" newrun)      # -> /tmp/illo/<runid>
# record the user's VERBATIM request (URL, pasted text, concept) — the
# gallery shows it as provenance so anyone can tell what the run was for:
printf '%s' "<the verbatim request>" > "$RUN/request.txt"
# (a) VARIATIONS — same prompt+model, pick-the-best:
python3 .../illo.py generate --prompt-file p.txt --ref <ref> --count 4 --label "draft→ship" --out "$RUN/v.png"
# (b) MODEL COMPARISON — loop the SAME prompt over the chosen models
#     (full OpenRouter ids from references/models.md):
for m in <model-id-1> <model-id-2>; do
  python3 .../illo.py generate --prompt-file p.txt --ref <ref> --model "$m" --label "$m" --out "$RUN/$(basename $m).png"; done
# (c) CONCEPT VARIATIONS — different prompts (different stagings) for one idea:
python3 .../illo.py generate --prompt-file staging-A.txt --ref <ref> --label "as a funnel" --out "$RUN/a.png"
python3 .../illo.py generate --prompt-file staging-B.txt --ref <ref> --label "as a crossing" --out "$RUN/b.png"

python3 "$SKILL_DIR/scripts/illo.py" gallery "$RUN" --title "<the piece or request>" --open
# always pass --title so a saved gallery stays identifiable later;
# add --embed for a single portable file (images inlined)

Every generate self-records to $RUN/manifest.jsonl; gallery assembles them into one page with each image's label, model, dimensions, cost, and a collapsible prompt — the prompt toggle is what makes concept-variation comparison readable (the prompt is the variable). Always present the gallery with a recommendation, not a raw dump — and in a chat session, present the labeled candidates directly in the chat instead of a gallery (delivery routing in step 7). Multi-model failures are per-image (an unavailable model errors that one render only); keep the rest.

6. QA and iterate

Check every image against references/quality-bar.md. Re-roll or edit when the mascot is decorative or off its locked spec, the body is wrong-value for the palette, label text sits on a colored fill, the accent has spread past the character's accent part + 1–2 elements, an unwanted title bar appears, the composition copies an example, or text is misspelled. Subject scale varies run-to-run — re-roll if the subject is tiny (check .width/.height in the JSON: a square back when 16:9 was requested → re-roll). When a re-roll supersedes a render, rebuild any delivery gallery with --exclude <superseded label> (repeatable) so rejected rolls don't appear in the review artifact.

7. Deliver — match the session's medium

Copy finals next to the user's work when appropriate; never overwrite existing assets without being asked. Filenames carry the role — they are the only metadata that survives a document attachment, so make them self-identifying: 00-hero-<slug>.png for the hero, then 01-<section-slug>.png, 02-<section-slug>.png, … for anchors in piece order (assets/<slug>-illustrations/). Then report: how many images, the palette used, which are strongest vs optional — and for any multi-image job, a placement map: one line per image naming the file, its role (hero, or after which section), and the one idea it lands, so the user can drop each file where it belongs without re-deriving the plan. Deliver the images themselves the way this session can actually show them:

• Filesystem sessions (IDE/terminal agents — Claude Code, Codex, Cursor): report each final's absolute path (the engine's JSON .path is already absolute) and present the gallery for multi-image runs. The file on disk already is the original — never emit [[as_document]] here: it's a Hermes gateway token, literal noise in any other runtime. If the runtime has its own in-chat file delivery, use that.

• Chat sessions (the user is on a messaging surface — Hermes over Telegram/Discord/WhatsApp, or any chat surface with lossy media delivery — and cannot open local files): a path alone is not a complete deliverable; the image must land in the chat, and a final must arrive as the original file. Platform photo delivery recompresses images — exactly what destroys riso grain, halftone texture, ink-layer offset, and fine hand-lettering — so finals are delivered as document attachments. On Hermes, tag each final with an explicit MEDIA: attachment tag — the tag is MEDIA: immediately followed by the absolute path, no space — and the literal directive [[as_document]] in the same reply. Do not rely on a bare absolute path for a final: bare paths can pass through to the user as literal text instead of being dispatched as an attachment.

  MEDIA:/absolute/path/to/final.jpg
  [[as_document]]
  

  Candidate/options rounds may use normal inline photo delivery when quick glances help — say so ("preview — original file to follow") — but a final is never delivered that way. Skip the HTML gallery in chat — the user has no easy way to open or host it; send the labeled finals directly with the recommendation as text, and only build gallery --embed (one self-contained file) if a portable artifact is explicitly requested, delivering it with [[as_document]].

Before the final reply in a chat session, check:

• every final's path came from the engine's JSON .path, not the requested --out (the actual extension may differ);
• every final appears as an explicit MEDIA:/absolute/path attachment tag in the reply;
• [[as_document]] is in the reply unless this is explicitly preview-only;
• rejected/re-rolled candidates are excluded from delivery;
• the text says what was made — character, palette, strongest final, and for sets the placement map (which file is the hero, which follows which section) — without implementation noise.

Output discipline

Pre-generation planning is short and concrete. Post-generation, let the images speak — report what was made and where, not style theory. Keep labels few and short; the fewer words baked into an image, the more reliably it renders.