uwf-model-adaptation - SKILL.md Agent Skill

name: uwf-model-adaptation description: "Resolve the model profile and emit a structured steering policy for any UWF stage."

UWF Model Adaptation Skill

Overview

The model-adaptation skill resolves a model profile once at orchestrator startup and injects a structured steering_policy into every stage invocation. This lets stage agents adapt their instruction density, example verbosity, and schema-reminder frequency to the steering needs of the running model — without requiring separate prompt files per model tier.

Profiles

Three built-in profiles are defined in profiles.yaml:

Profile	Use when	Instruction density	Examples	Schema reminders
`compact`	Smaller / instruction-following models	`expanded`	yes	`exhaustive`
`balanced`	Mid-tier models (default)	`standard`	no	`standard`
`reasoning`	Large reasoning models (o1, opus, …)	`compact`	no	`concise`

The default profile is balanced. Unknown or unresolvable inputs always fall back to balanced.

Resolution Script

.github/skills/uwf-model-adaptation/resolve.mjs

Usage

node .github/skills/uwf-model-adaptation/resolve.mjs detect [--profile <name>] [--model <model_name>]

Resolution priority

--profile flag — explicit; exits 1 if the value is unknown
--model flag — matched against model_map in profiles.yaml
UWF_MODEL_PROFILE environment variable
UWF_MODEL_NAME environment variable (matched)
Default: balanced

Output shape

{
  "profile": "balanced",
  "model_name": "claude-sonnet",
  "steering_policy": {
    "instruction_density": "standard",
    "include_examples": false,
    "step_expansion": true,
    "schema_reminders": "standard",
    "self_check": true
  }
}

Exit codes

Code	Meaning
`0`	Success
`1`	Invalid explicit profile (`--profile` was given but unknown)
`2`	Usage error

Steering Policy Fields

Field	Values	Meaning
`instruction_density`	`expanded` \| `standard` \| `compact`	How much explanatory text to include in stage prompts
`include_examples`	`true` \| `false`	Whether to inline worked examples
`step_expansion`	`true` \| `false`	Whether to expand procedural steps into sub-bullets
`schema_reminders`	`exhaustive` \| `standard` \| `concise`	How verbosely to repeat output schema requirements
`self_check`	`true` \| `false`	Whether the agent should self-check outputs before emitting

Stage Overrides

profiles.yaml may define per-stage policy overrides under stage_overrides. These are applied on top of the base profile steering policy by the stage-resolution layer in stage-tracker.mjs when constructing the resolved stage object.

Non-negotiable Rules

This skill is the single source of truth for profile definitions. Do not copy profile logic into individual stage agents or workflow skills.
Model profile must be resolved once at orchestrator startup, not re-resolved per stage.
Unknown models must silently fall back to balanced — never fail on an unrecognized model name.
Invalid explicit --profile values must exit 1 so the orchestrator can surface the error.
Do not infer the model from response quality, tool use, or conversation content.