Tarsier
One-shot creative generator. Produces a complete SVG of a tarsier riding a bicycle, a rasterized 500×500 PNG with padding on white, and a markdown transcript recording what was made and how.
Invocation
/tarsierNo arguments. The prompt is static and never changes. Model name and reasoning level come from the harness when it exposes them.
What you produce
Three files in a new folder under the caller's current working directory:
{DIR}/
├── {BASE}.svg # the raw SVG you authored
├── {BASE}.png # 500x500 raster, 16px padding, white background
└── {BASE}.md # transcript with metadata and the full SVG source{BASE} follows the pattern tarsier-<model>-<reasoning>-<NZST-timestamp>.
Execution
Follow these steps in order. Each step has a verification cue — don't skip it.
Step 1 — Setup
- Model name: Read the current model ID from harness metadata. Strip any context-window suffix in brackets (e.g.
claude-opus-4-6[1m]→claude-opus-4-6). When the harness does not expose a model ID, useunknown-model. - Reasoning level: Read the active reasoning or effort level from harness metadata or the conversation. When unavailable, use
default. - Timestamp (NZST):
TZ='Pacific/Auckland' date '+%Y%m%dT%H%M%S' - Construct base name:
tarsier-{model}-{reasoning}-{timestamp}— all lowercase, hyphen-separated. - Create the output folder in the current working directory:
mkdir -p./{BASE}
Step 2 — Generate the SVG
Respond to the prompt below as if it were the only instruction given. No embellishment, no negotiation, no asking the user for style guidance:
Generate an SVG of a tarsier riding a bicycle
Write your SVG to {DIR}/{BASE}.svg.
Technical requirements (for rasterization compatibility only — these do not constrain the art):
- Complete
<svg>element withxmlns="http://www.w3.org/2000/svg" - Include a
viewBoxattribute - No text or prose outside the SVG markup itself
- Self-contained — no external image refs, no external font imports
A tarsier is a small nocturnal primate with huge round eyes, long fingers, a long tail, and a small furry body. Show it clearly on a bicycle. Beyond that, the art direction is up to you — detail level, style, and colour are intentionally unconstrained.
Step 3 — Rasterize to PNG
Run the rasterization script from this skill's scripts/ directory. The script wraps the SVG in a 500×500 white canvas with 16px padding so projectors and blog thumbnails render it consistently:
python3 <skill-dir>/scripts/rasterize.py "{DIR}/{BASE}.svg" "{DIR}/{BASE}.png"<skill-dir> is the path where this skill was installed. If the harness does not expose it, resolve scripts/rasterize.py relative to the location of this SKILL.md file.
Verify the PNG was created and is non-empty:
test -s "{DIR}/{BASE}.png" && echo "PNG OK: $(wc -c < "{DIR}/{BASE}.png") bytes"Step 4 — Write the transcript
Write {DIR}/{BASE}.md with this exact format (replace bracketed fields, keep the structure):
# Tarsier
| Field | Value |
|-------|-------|
| Date | {NZST datetime, e.g. 2026-04-10T14:30:45 NZST} |
| Model | {model} |
| Reasoning | {reasoning} |
| Prompt | Generate an SVG of a tarsier riding a bicycle |
## SVG Output
```svg
{the complete SVG code — verbatim from Step 2}
```Step 5 — Report
Tell the user the output folder and list its contents. Example:
./{BASE}/
├── {BASE}.svg
├── {BASE}.png (500x500)
└── {BASE}.mdDo not offer to upload, deploy, share, or host the output. Do not ask the user if they want another variant. The skill is one-shot by design.
The prompt
The prompt is always exactly:
Generate an SVG of a tarsier riding a bicycle
It is static. Do not paraphrase it in the transcript. Do not expand it with extra guidance inside your own reasoning. The value of a one-shot generator is that the input is fixed; the SVG is the only variable.
Naming convention
tarsier-<model>-<reasoning>-<NZST-timestamp>
- Model: harness-reported model ID, lowercase with hyphens.
unknown-modelwhen unavailable. Strip context-window suffixes like[1m]. - Reasoning: effort level for this session, lowercase (
default,low,standard,high,thinking, etc.). - Timestamp: NZST in
YYYYMMDDTHHmmSSformat (no punctuation between date and time parts).
Example: tarsier-claude-opus-4-6-default-20260410T143045
Gotchas
1. rsvg-convert must be installed. The rasterization script shells out to rsvg-convert (from librsvg). On macOS: brew install librsvg. On Debian/Ubuntu: apt install librsvg2-bin. If the binary is missing the script will fail at Step 3 with a non-zero exit code — surface that error to the user instead of silently skipping the PNG.
2. SVGs without a viewBox will be cropped. The rasterizer extracts viewBox and falls back to width/height attributes. If neither exists it assumes 0 0 100 100, which almost always cuts off the drawing. Always include a viewBox — this is the single most common cause of a blank or clipped PNG.
3. Do not put anything outside the <svg> element. The rasterizer only copies content between <svg...> and </svg>. HTML comments, prose, or a DOCTYPE before the SVG tag will be silently discarded. Put comments inside the SVG with <!--... --> if you need them.
4. Timestamp must be NZST, not UTC or local. The TZ='Pacific/Auckland' prefix is intentional so outputs sort consistently across machines regardless of where the harness runs. Omitting it will produce inconsistent folder names across runs.
5. The prompt is frozen, the art is variable. The point of a one-shot generator is that the input never changes. Do not add creative guidance ("make it detailed", "use flat colours") to your own interpretation — the variance between runs is supposed to come from the model, not from you editorialising the prompt.
6. Output folder lives in the caller's CWD. mkdir -p./{BASE} creates the folder where the user invoked the command, not inside the skill directory. This is deliberate — users expect their files next to whatever they were working on.
Files in this skill
scripts/rasterize.py— SVG → 500×500 PNG wrapper. Executed in Step 3. Do not modify per-run.evals/evals.json— test cases for the skill (smoke, edge, negative, disclosure).templates/transcript.md— the transcript format from Step 4 as a ready-to-fill template.