Production-tested system prompts for building agents on Solar. Each agent in this repository is a single agent.md file you can copy into any OpenAI-compatible client, together with real outputs it produced, so you can see exactly what you will get before writing a line of code.
Every published artifact here (spreadsheets, reports, slide decks, games, simulations, dashboards) was generated by solar-pro4 through the included runner (one simulation, weather-island, is retained from solar-open2 after repeated pro4 attempts failed its browser gate; each output's run-meta.json records the exact model), with pass criteria in a checklist. Nothing is hand-edited. When a generation falls short, we regenerate and promote the best run.
About the model. Solar Pro 4 is Upstage's flagship model, specialized for agentic use. It is particularly strong at document-based work, coding, and business-related tasks, supports a 512K context length with full English, Korean, and Japanese coverage, and offers a reasoning mode. Exactly the profile this cookbook exercises: office documents, games and dashboards, and tool-calling agent loops.
You need a Solar API key (create one at console.upstage.ai/api-keys) and Python 3.
pip install requests
export UPSTAGE_API_KEY=your_key
cd solar-agent-cookbook
python3 runner/run.py --agent simulation-builder \
--task simulation-builder/golden/01-aquarium/task.md \
--out /tmp/aquarium-runThis sends the agent's system prompt and the task to solar-pro4, saves the generated file to /tmp/aquarium-run/, and writes a run-meta.json describing the run. Open the resulting aquarium.html in your browser and you should see a living aquarium.
Prefer notebooks? quickstart.ipynb walks through the same flow with plain API calls, including the tool-calling loop, with executed outputs you can read before running anything. And any OpenAI-compatible client works: put the contents of agent.md in the system message, your task in the user message, and call model solar-pro4.
Short, single-purpose notebooks in capabilities/, each isolating one prompting technique with a baseline-versus-improved measurement on real solar-pro4 output. Read them before building the equivalent capability into an agent or skill.
| Guide | What you learn | Technique |
|---|---|---|
| parameters | Pick the right reasoning_effort, temperature, top_p, max_tokens, stream settings | defaults + ranges + live effect demos |
| classification | Route Korean inquiries into 6 support labels | label definitions + few-shot + output contract + abstain rule |
| summarization | Keep a press-release summary faithful and on-format | format contract + faithfulness rules + extract-then-summarize |
| structured-extraction | Turn Korean emails into validated JSON | schema-in-prompt + null rule + validate-and-retry loop |
| translation | Lock business terminology across a translation | glossary injection + structure/number rules + tone control |
Every preview below is a real, unedited solar-pro4 output committed in this repository.
 |
 |
 |
| excel-agent: site-evaluation workbook | doc-agent: feasibility report | ppt-agent: board-approval deck |
 |
 |
 |
| simulation-builder: deep-sea aquarium | game-builder: neon breakout roguelite | frontend-artifact: finance dashboard |
| Agent | What it builds | Run mode |
|---|---|---|
| excel-agent | Excel workbooks computed from input data, with formatting and a verification pass | loop |
| doc-agent | Word reports whose numbers are pulled from upstream data, never re-derived | loop |
| ppt-agent | PowerPoint decks with charts and speaker notes on every slide | loop |
| game-builder | Complete, playable browser games in one HTML file | oneshot |
| simulation-builder | Self-running animated simulations (physics, nature scenes) in one HTML file | oneshot |
| frontend-artifact | Practical single-file web UIs such as dashboards and internal tools | oneshot |
| research-agent | Research briefs where every claim carries a citation, like the committed model brief | loop |
The three office agents are designed to chain. The workbook's checkpoint files feed the report, and both feed the deck, so numbers stay consistent across all three documents by contract rather than by luck.
Skills live in solar-skills/: reusable capabilities any agent can load, each shipped with a golden run as proof.
| Skill | What it does | Golden proof |
|---|---|---|
| document-qa | Reads documents through Upstage Document Parse and answers with page-level citations like (document, p.30) |
5/5 questions on a 173-page quarterly report answered with verified page citations |
agent.md anatomy. Every prompt follows the same five-part structure, documented in TEMPLATE.md: a persona with a quality bar, hard rules written from observed failures, a process with checkpoints, an output contract with self-verification, and communication rules. For the office agents you can watch the model actually following these rules in the committed transcript.json files (where available; one run's transcript was lost and is disclosed in its checklist).
Two run modes. oneshot sends one request and saves the response as a file. It fits anything that is a single artifact, like a game or a dashboard. loop gives the model tools (bash, write_file, read_file, fetch_url, finish) and lets it work step by step, which is what office documents and research need. The runner is about 100 lines per mode and exists so results are reproducible, not to be a framework. Swap in your own harness freely.
Golden sets. Each agent ships golden/NN-slug/ containing the exact task, its inputs, the real outputs, and a checklist.md a human can judge in five minutes. Goldens are the regression baseline: prompt changes must keep them passing.
New agents, new golden tasks, and prompt improvements are all welcome. See CONTRIBUTING.md for the workflow. The two house rules: outputs must be genuine Solar generations with run-meta.json provenance, and prompt changes must keep the golden checklists passing.
solar-agent-cookbook/ # agents: <name>/agent.md + golden/ + README with preview
runner/ # minimal reproducible runner (oneshot + tool loop)
TEMPLATE.md # how to write an agent.md (the 5-part anatomy)
solar-skills/ # skills: <name>/SKILL.md + scripts/ + golden/