The Agent Skills Directory

[PROMPT_INJECTION]: The skill exhibits an attack surface for indirect prompt injection by incorporating untrusted experimental data into prompts used by LLM-based optimizers.
Ingestion points: User-provided target colors, volume ratios, and labware configurations enter the workflow in references/colour-mixing-opt.md and references/viscosity-optimization.md.
Boundary markers: Prompt templates within scripts/optimizers.py utilize markdown headers and clear role definitions to delineate data history from model instructions.
Capability inventory: The agent has the capability to execute shell commands (e.g., via uv run), interact with local network hardware (Opentrons IP), and perform network requests to the OpenRouter API.
Sanitization: The scripts/optimizers.py implementation performs JSON schema validation and validates numeric constraints, such as ensuring well volumes sum correctly.
[COMMAND_EXECUTION]: The skill instructions require running shell commands to interface with local laboratory equipment.
Evidence: The file references/viscosity-optimization.md provides instructions to launch a mass balance service using the command uv run --package balance-edge python edge/balance.py.
[EXTERNAL_DOWNLOADS]: The skill documentation and scripts specify several external scientific and machine learning dependencies.
Evidence: The workflow depends on the installation of botorch, gpytorch, torch, openai, numpy, and Pillow.
Vendor Context: The balance-edge package is described as part of the vendor's (PUDAP) specialized laboratory setup for gravimetric feedback.

puda-workflows