Terminal Poster

Generate dense, on-brand infographic posters in a terminal-aesthetic visual style — dark + orange + pixel-bitmap + ASCII + brutalist minimalism. The look popularized by @shannholmberg for his viral Hermes Agent posts on X, generalized into a reusable system.

Pattern reverse-engineered from 27 image-summary posts (12 of them deep-audited). 5 reusable templates — A, B, C, D, E — share one design DNA. Cluster C and D each have 2 sub-modes (C1 painterly / C2 pixel, D1 step pipeline / D2 terminal-window). Tested live with Nano Banana Pro — Cluster A audited hit 99%, Cluster C1 painterly hit 95.6%.

Fresh agent? Start here: 1) Read this SKILL.md top-to-bottom. 2) Pick a cluster using the decision tree below. 3) Fast path — for Cluster A or C, write a YAML spec and run scripts/make-poster.sh spec.yaml out.png. See example-specs/. 4) Manual path — for B/D/E or custom edits, open the matching references/templates/cluster-X-*.md, fill placeholders, run scripts/generate.sh. 5) Vision-audit the output. The full design rules live in references/design-dna.md — read it once for context.

Decision tree — which template to use?

What is the topic?
├── Pure system architecture / framework / diagram (org charts, pipelines)
│   └── 🔴 USE CLUSTER A (ASCII Terminal) — highest fidelity (94-99% audited), cheapest to nail
│
├── Leveled comparison / maturity model / "5 ways to X"
│   └── 🟡 USE CLUSTER B (Color-Coded Levels)
│
├── Feature launch / viral hook / "X changed how I work"
│   └── 🔴 USE CLUSTER C (Cyborg Hero) — pick a sub-mode:
│       ├── 🔴 Mode C1 — PAINTERLY HERO (viral, 4/6 of Shann's Cluster C)
│       │     Use for serious launches, agent reveals, big-deal announcements
│       └── 🟡 Mode C2 — FLAT PIXEL HERO (whimsical, 2/6)
│             Use for lighter / playful / less-serious topics
│
├── Step-by-step tutorial / blueprint / process flow
│   └── 🟡 USE CLUSTER D (Blueprint) — pick a sub-mode:
│       ├── 🔴 Mode D1 — STEP PIPELINE (brand-book / playbook style)
│       │     Green-dominant, 4-7 step cards, macOS chrome dots
│       │     Use for sequential processes ("do X, then Y, then Z")
│       └── 🟡 Mode D2 — TERMINAL-WINDOW MOCKUP (premium thought-piece)
│             Fake terminal floating over painterly canvas
│             Use for single concepts framed as "the rule"
│
└── Brand identity / design system / voice doc
    └── 🟢 USE CLUSTER E (Editorial) — rare, only when topic is literally brand

Default: When in doubt, use Cluster A. It's the cheapest to nail and matches the highest share of the source corpus.

Topic → Cluster lookup (common cases)

🟢 Quick rule: Architecture → A. Launch → C1. Playbook → D1. Levels → B. Brand → E.

The common design DNA (applies to all 5 clusters)

⚠️ Audited May 24, 2026 against 12 real images. Many rules are CLUSTER-SPECIFIC, not universal. See references/design-dna.md for the full breakdown. Below is the abridged "what really holds" version.

🚀 Fast path — use the CLI

If you know which cluster you want, the fastest way to generate a poster is the make-poster.sh CLI:

bash scripts/make-poster.sh \
  scripts/example-specs/your-spec.yaml \
  /tmp/output.png

The CLI takes a YAML spec, picks the right cluster template, fills placeholders, calls generate.sh, and saves the prompt next to the output for reproducibility.

Supported clusters via CLI: a (ASCII Terminal) and c (Cyborg Hero — both c1 painterly and c2 flat pixel). Manual fill: b, d, e — use the template files directly (see Step 2 below).

Example specs: scripts/example-specs/

• cluster-a-stack.yaml — Cluster A engineer-poet zine, 3-panel agent stack
• cluster-b-maturity.yaml — Cluster B color-coded 4-level maturity ladder
• cluster-d-playbook.yaml — Cluster D blueprint 5-step playbook
• c2-smoketest.yaml — Cluster C2 whimsical flat-pixel hero

YAML spec format (Cluster C1):

cluster: c
mode: c1                     # c1 painterly OR c2 flat pixel
topic: yourapp               # top-left terminal prompt subject
status: shipping             # top-left terminal status
status_2: 6 agents online    # top-right MIRRORED status (NOT a version stamp)
handle: "@yourhandle"
eyebrow: AGENT-NATIVE
brand: YOURAPP
subtitle: 6 AGENTS · ONE BRAIN
hero_subject: >
  (multi-line painterly hero description — model fills in details from the 5 composition rules)
cards:
  - {header: AGENT ONE,   icon: clipboard,        body: "..."}
  - {header: AGENT TWO,   icon: magnifying-glass, body: "..."}
  # ... exactly 6 cards
tagline_separator: star      # star | pipe | period
tagline_phrases:
  - SHIP A BRAIN NOT A PROMPT
  - PROOF OVER PROMISES
  - ONE BRAIN. SIX AGENTS.

YAML spec format (Cluster A):

cluster: a
title: a typical agent stack
bottom_tagline: "spec in. action out. ship it."
handle: "@yourhandle"
panels:
  - label: PERCEPTION
    subject: input layer
    tagline: structured intent in
    flow: "user prompt ──▶ parser ──▶ structured spec"
    prose: "turns ambient natural-language input into a machine-readable plan."
    items: ["yaml or natural-language spec", "schema validation", "intent classification", "skill routing"]
  # ... 3-5 panels total

The CLI takes ~5 seconds + the generate.sh API call (~30 seconds). Cost ≈ $0.002.

Manual workflow (when CLI doesn't fit)

Step 1 — Gather inputs from the user

Ask for:

1. Topic / brand name — what the poster is about (e.g. "MarketIntell", "Hermes Agent", "Bookmark Brain")
2. Eyebrow/kicker — short framing label above the brand (e.g. "AGENT-NATIVE", "OPEN SOURCE")
3. Subtitle — what it is in 4-6 words (e.g. "the AI analyst desk")
4. 3-6 content blocks — for the body. Each block needs: a short ALL-CAPS header, an icon concept, and 1-3 lines of body copy.
5. Manifesto tagline — 1-3 phrases for the bottom bar
6. Handle / status string — for top-right corner. Cluster C uses a mirrored > sys$ [STATUS_2] @handle terminal prompt (e.g. "@ravidsrk uptime 47h" or "@ravidsrk ready"). NOT a version stamp — that was wave-1 thinking, audit replaced it with a second terminal prompt.
7. Cluster choice — A/B/C/D/E. If unsure, suggest Cluster A.

If user already gave most of this, infer the rest. Don't re-ask.

Step 2 — Gather accurate facts (if applicable)

If the topic is a real product/project, look up any existing notes or docs first so the poster reflects real numbers and taglines. Don't invent data.

Step 3 — Fill the prompt template

Use one of the 5 templates in references/templates/. Replace [BRACKETED PLACEHOLDERS] with the user's content.

Templates available:

• references/templates/cluster-a-ascii-terminal.md — default
• references/templates/cluster-b-color-coded.md
• references/templates/cluster-c-cyborg-hero.md — viral version
• references/templates/cluster-d-blueprint.md
• references/templates/cluster-e-editorial.md

Step 4 — Generate via Nano Banana Pro

Run the helper script:

bash scripts/generate.sh \
  /path/to/prompt.txt \
  /path/to/output.png

The script handles: JSON escaping → OpenRouter POST → base64 decode → save PNG.

Always use Nano Banana Pro (google/gemini-3-pro-image-preview) as default. Don't downgrade silently — Nano Banana 2 hallucinates duplicate cards on Cluster C.

Step 5 — Verify with vision

Use read(path, prompt="...") to grade the output. Vision-check for:

1. Tagline typos (the word "PERIOD" appearing literally — known model failure)
2. Duplicate cards
3. Painterly hero (when prompt asks for flat pixel-art)
4. Wrong palette (drift to pure black or pure white)

If score < 80%, regenerate with tighter constraints.

Step 6 — Preview

previewFile("/path/to/output.png")

⚠️ Critical pitfalls (learned the hard way)

🔴 Forbid the word "PERIOD" explicitly. Models spell it out as filler text in tagline bars, even when given literal . examples. Always include this block in any prompt with a tagline:

⚠️ CRITICAL TEXT RULES — read carefully before rendering letters:
• Spell every word completely. Do NOT abbreviate. Do NOT drop letters.
• The word "PERIOD" must NEVER appear in the tagline. Use a literal "." punctuation mark, not the spelled-out word.
• Each sentence ends with a single period character "." — not the word PERIOD.
• Render the dot as a small square pixel, not as letters.
• Avoid any duplicate or stuttered words.

🔴 Always specify EXACT grid dimensions for Cluster C. Say "EXACTLY a 3 columns × 2 rows = 6 cards" — not "6 cards in a grid". Models otherwise produce 4+3 asymmetric layouts or duplicate cards.

🔴 Scope "no painterly" to the Cluster C BODY CARD GRID only. The hero zone in Mode C1 SHOULD be painterly — that's the viral pattern (see line below). But inside the 3×2 card grid, explicitly say: "Cards are FLAT — no painterly, no shadows, no gradients inside cards." Without this scoping, the model bleeds painterly texture into the body cards and they stop reading as engineered chrome.

🟡 Brand names with short letters (e.g. "Polymarket") sometimes get a single glyph corrupted. If brand name is critical, either spell it out in the prompt OR run a 2nd pass.

🟡 Use OpenRouter, not direct Google API. The default Google GOOGLE_GENERATIVE_AI_API_KEY is project-blocked → 403. OpenRouter route is the reliable path.

🟢 Cluster A is the cheapest to nail — pure ASCII, two-tone, easy for the model. 94-99% on first try (audited). Default to it.

🔴 For viral Cluster C posters (gBrain, control-room, skill-bundles-ship style), ALLOW painterly hero illustration. An earlier version of this skill forbade "painterly" globally — that rule was wrong. The viral Cluster C images USE painterly cyborg portraits with radial orange glow. Forbid painterly only inside the body card grid; hero illustration may be painterly + radial glow. See cluster-c-cyborg-hero.md for the two hero modes (C1 painterly / C2 pixel-flat).

🔴 Pick the right orange for the cluster. #F26B1F is the Cluster C canonical (gBrain). Cluster B uses muted rust #B8541F. Cluster E uses red-orange #FC4A2B. Cluster D1 uses phosphor green #A8E060 as primary, not orange. Don't reach for #F26B1F reflexively.

🟡 Cluster A signature motifs are non-negotiable — ◆ diamond separator in panel headers, › right-angle quote bullets, "what runs here:" ritual line, inline LEVEL N: labels. The skill template includes these; if a Cluster A generation lacks them, regenerate.

Cost

• Nano Banana Pro: ~$0.002 / image
• Nano Banana 2: ~$0.0005 / image (use only for drafts)
• Typical session (3 iterations): ~$0.006

Files in this skill

terminal-poster/
├── SKILL.md                         ← this file
├── scripts/
│   ├── generate.sh                  ← Low-level bash helper, takes prompt text + output path
│   ├── make-poster.sh               ← 🚀 High-level CLI, takes YAML spec + output path (recommended)
│   └── example-specs/
│       ├── cluster-a-stack.yaml     ← Cluster A engineer-poet (3-panel agent stack)
│       ├── cluster-b-maturity.yaml  ← Cluster B color-coded (4-level ladder)
│       ├── cluster-d-playbook.yaml  ← Cluster D blueprint (5-step playbook)
│       └── c2-smoketest.yaml        ← Cluster C2 whimsical flat-pixel hero
└── references/
    ├── design-dna.md                ← full pattern doc (palette, fonts, layout rules)
    ├── templates/
    │   ├── cluster-a-ascii-terminal.md  ← DEFAULT
    │   ├── cluster-b-color-coded.md
    │   ├── cluster-c-cyborg-hero.md     ← VIRAL
    │   ├── cluster-d-blueprint.md
    │   └── cluster-e-editorial.md
    └── worked-examples.md           ← real test results, lessons learned

Worked examples

See references/worked-examples.md for an iteration log — scores from 70% (Nano Banana 2 first try) → 99% (audited Cluster A), with the lessons that produced each jump.

Credits

Visual pattern reverse-engineered from @shannholmberg on X. He didn't design this skill — he designed the look. This skill just makes the look reproducible across topics.