phira-ai Use this skill when generating research directions for the Phira team.
Purpose
- Convert vague goals (or paper-derived mechanisms) into multiple actionable directions.
- Ensure each direction is internally consistent and can be handed to
phira-prototyper.
Required output (option cards)
Produce 3-5 option cards. Each card MUST include:
Title + one-line thesis
Core mechanism (math-first)
- Define symbols before first use.
- Include the objective and/or update rules when applicable.
- State key assumptions explicitly.
- Training pipeline (end-to-end)
- Data: inputs, preprocessing, split protocol, leakage risks.
- Model: modules/interfaces; key shapes if relevant.
- Objective: losses/regularisers/constraints; weighting.
- Optimisation: optimiser/schedule/stability tricks; any distributed/precision assumptions.
- Outputs: checkpoints, logs, metrics, artefacts.
- Inference pipeline (end-to-end)
- Inputs/outputs; runtime steps (retrieval/decoding/postprocessing if any).
- Calibration/uncertainty or failure handling assumptions if relevant.
Failure modes and tradeoffs (ranked)
Validation hooks (light)
- 1-3 quick checks that would confirm the direction is implemented/plumbed correctly.
- Avoid long experiment plans.
Recommendation + handoff
End with:
- A recommended option and the criteria used.
- A handoff packet for prototyping: the minimal interfaces and unknowns to resolve.
Notation rules
- $\mathcal{C}$ denotes a set.
- Bold lowercase $\mathbf{x}$ denotes a vector; bold uppercase $\mathbf{X}$ denotes a matrix.
- Uppercase $X$ denotes a random variable; lowercase $x$ denotes a deterministic value.
- Use $...$ for inline maths and $$...$$ for display maths.