uac-import - SKILL.md Agent Skill

name: "uac-import" description: "Inspects external prompt-like sources and maps them into Core-Prompts plan, judge, and apply flows. Use when importing, planning, judging, or applying a new or updated capability."

UAC Import — Capability Intake, Quality Review, and Uplift

Purpose

Take one or more external prompt-like sources and turn them into deterministic UAC assessments, layered manifests, canonical SSOT entries, machine-readable descriptors, quality-review artifacts, and advisory handoff contracts.

Supported inputs:

local file path
local folder path
raw public HTTPS URL
GitHub repo or folder URL
multiple --source values in one run
repomix-reduced repo input when available

Supported modes:

import
audit
explain
plan
judge
apply

Primary Objective

Classify the source safely, recommend the right surface area, and refuse landing until the candidate is structurally strong enough to become canonical SSOT plus descriptor state.

When structural quality is near the bar but behavioral confidence is still weak, escalate to a bounded behavioral proof workflow instead of overstating readiness.

Invocation Contract

Use this capability as the AI-facing intake contract.

Primary entrypoints:

bin/uac for intake, audit, planning, judging, and apply
bin/capability-fabric for surface build, validation, and deploy steps after landing

Canonical state:

ssot/<slug>.md is the human-readable prompt source of truth
.meta/capabilities/<slug>.json is the machine-readable descriptor source of truth
sources/ssot-baselines/<slug>/baseline.md is the baseline prompt-body fidelity source of truth used by UAC quality judging
generated surfaces under .codex/, .gemini/, .claude/, and .kiro/ are derived artifacts

Operational rule:

apply may mutate canonical repo state after confirmation
deploy is separate and must never be implied by apply
when shell entrypoints are required to complete the workflow, say so explicitly instead of assuming the caller knows them

Workflow

Ingest the source through the deterministic pipeline.
Produce a clean summary.
Run uplift to extract objective, scope, and constraints.
Run semantic routing.
If the source is a folder or repo subtree, inventory prompt-like files and classify them one by one.
Cluster broad repos into candidate families before recommending any landing.
Classify accepted sources as skill, agent, both, or manual_review.
Build layered manifests, cross-analysis, and advisory handoff data.
Select a quality profile and benchmark set.
Resolve the canonical baseline source from sources/ssot-baselines/ before judging fidelity.
On judge, run the built-in quality loop and return judge packets plus pass/fail reports without landing repo state.
If judge finds that structural quality is close to passing but behavioral confidence is insufficient, route to auto-research for bounded capability evaluation instead of guessing.
Search for benchmark sources only when the source is generic or fit confidence is weak.
On apply, refuse landing unless the quality loop reaches ship; if it does, materialize or preserve the canonical baseline source under sources/ssot-baselines/, write canonical repo state under ssot/ and .meta/capabilities/, persist quality reviews, then rebuild and validate generated surfaces.
Keep deployment separate from apply.

Tool Boundaries

allowed: inventory sources, run deterministic uplift and classification, produce advisory manifests, and land canonical SSOT plus descriptor state when the quality gate passes
forbidden: inventing capability strength that the source does not support, deploying to user homes during apply, or silently widening surface recommendations to satisfy packaging preferences
escalation: if the work shifts from intake to architecture, docs quality, testing, or release readiness, route to the companion capability with a concrete handoff instead of stretching the import workflow

Rules

Prefer existing pipeline code over ad-hoc parsing.
Keep results deterministic and roleplay-free.
Fail closed for unsuitable URL content.
For folders or repo trees, only group files that were actually inventoried.
If the source is config-only, require manual review instead of pretending it is a prompt.
If the source is already an agent definition, preserve its control-plane boundaries.
Never make orchestration or delegation decisions. Publish advisory metadata only.
Run cross-analysis against current SSOT before any apply is considered safe.
Treat commands, plugins, powers, and extensions as deployment wrappers, not capability types.
Quality review artifacts are advisory evidence; they must not encode runtime routing policy.
Do not make UAC the long-term owner of behavioral evaluation logic; route to auto-research when bounded behavioral proof is needed.

Invocation Hints

Use this capability when the user asks for any of the following, even without naming the skill:

import a prompt, prompt pack, or capability into this repo
classify whether this source should become a skill, agent, or manual review
explain how this external source would land into SSOT and descriptors
judge whether a candidate is ready to apply
tell me whether this import needs stronger behavioral proof before landing

Required Inputs

one or more explicit sources
desired mode such as plan, judge, or apply
target system or install target when that materially affects the recommendation
any benchmark or quality expectations when the caller wants a stricter gate

Required Output

Return a concise structured result with these sections:

Source
Summary
Uplift
Routing
UAC Classification
Collection Recommendation
Layered Manifest
Cross-Analysis
Quality Plan / Judge Reports
Install Target
Advisory Handoff Contract
Recommended Surface
Modernization Focus
Next Actions

When judge escalates to behavioral proof, also include:

Behavioral Confidence
Escalation Reason
Auto-Research Handoff

Companion Capability Matrix

If the import uncovers this need	Route to	Required handoff
The candidate needs deeper prompt hardening before it can pass the benchmark gate	`supercharge`	source excerpt, intended user outcome, weak sections, target capability style
The candidate is structurally close to passing but needs bounded behavioral proof against baseline or competing variants	`auto-research`	baseline artifact, candidate artifact or variants, claimed job, bounded task set or examples, pass/fail threshold
The candidate is structurally sound but needs final decision synthesis across several landing options	`converge`	candidate options, trade-offs, target install surfaces, decision criteria
The imported capability is fundamentally architectural or system-design oriented	`architecture`	source summary, design scope, affected boundaries, unresolved design questions
The imported capability needs documentation-quality review before landing	`docs-review-expert`	draft SSOT, descriptor summary, naming questions, drift or IA concerns
The imported capability requires stronger validation or test coverage in this repo	`testing`	changed scripts, validator paths, expected behaviors, missing coverage risks
The imported capability is ready to land but release, packaging, or CI readiness is the real question	`gitops-review`	applied diff, generated artifacts, validation output, release and deploy intent

Constraints

No hidden execution.
No packaging claims without evidence.
No deployment during apply.
For local folders or GitHub repos, inventory the files first and justify whether they belong under one roof.

Examples

Example Request

Import this prompt folder, tell me whether it belongs as one capability or several, and refuse apply if it misses the benchmark bar.

Example Output Shape

source inventory
classification and fit assessment
benchmark and quality status
canonical landing recommendation
next actions

Evaluation Rubric

Check	What Passing Looks Like
Source fidelity	The recommendation reflects the actual source set rather than guessed structure
Baseline fidelity	The candidate is judged against the repo-resident baseline source, not against metadata polish alone
Classification rigor	The result explains why the source is `skill`, `agent`, `both`, or `manual_review`
Landing safety	`apply` is blocked until benchmark and quality gates pass
Canonical output	The result names the SSOT, descriptor, and generated-surface consequences clearly
Boundary clarity	Deployment wrappers are not confused with capability types

Capability resource: .codex/skills/uac-import/resources/capability.json