marcelocruzrpa/uipath-ai-skills/uipath-core

UiPath Core Skill

⚠️ Safety Rules (I-1, I-2 in rules.md)

- Playwright/browser inspection is READ-ONLY. Login page → STOP, WAIT for user. See ui-inspection.md → Login Gate.

- Desktop inspection via PowerShell (inspect-ui-tree.ps1): read-only tree inspection only.

- NEVER generate credentials, tokens, or passwords — real or fake.

Generate production-quality UiPath automation artifacts using real Studio-exported templates and comprehensive reference documentation. Template baseline: Studio 24.10 Windows.

Hard Rule: Never Write .xaml by Hand

If you are about to use the Write tool with a .xaml extension, STOP. Every .xaml file in this skill MUST come from python -m uipath_core.scaffold_project or python scripts/generate_workflow.py. Hand-written XAML bypasses lint rules 1–122 and produces unrunnable workflows. If generator output is wrong, fix the JSON spec — do not edit .xaml manually except for narrow modify_framework.py operations.

This rule overrides any plan you would have produced from training memory.

Planning Workflow (read this BEFORE drafting any plan)

When the user hands you a PDD, requirements doc, or "build a UiPath project that does X" — STOP before drafting architecture. Read this section first; it overrides any plan you would have produced from training memory. The full text of every gate below lives in references/scaffolding.md § "Generating a Full Project" — this is the planner-side hoist.

PLAN SELF-CHECK — answer BEFORE presenting your plan to the user

• [ ] For EACH workflow in my plan: does it implement the PDD's specified approach, or did I substitute a "better" alternative? (e.g., InvokeCode SHA1 instead of browser automation of sha1-online.com)

→ If ANY substitution exists: REMOVE it and implement the PDD way. Propose the alternative AFTER.

• [ ] Does my plan include an explicit inspection step for EVERY target app?

→ Browser apps: "Inspect [app] with Playwright" (if Playwright MCP available) → Desktop apps: "Inspect [app] with PowerShell" (if PowerShell available on Windows) → If ANY app lacks an inspection step and the tool is available: ADD it now.

If you cannot answer YES to all checks, revise your plan before presenting it.

Echo-back requirement: the produced plan MUST contain a PLAN SELF-CHECK: block (in Phase 0 or as Phase 0.5) with two ✓ lines, one per box, naming the specifics — for example:

PLAN SELF-CHECK:
  ✓ No PDD substitutions — SHA1 computed via sha1-online.com browser flow as PDD specifies
  ✓ Phase 2 includes inspection for every target app: System1 + SHA1Online

For a single-app or non-UI process, the second line still appears — e.g. ✓ Phase 2 explicitly notes no UI inspection required (no browser/desktop apps in scope). A plan without this echo-back block is incomplete; revise before presenting.

PLAN OUTPUT FORMAT — present your plan using this structure

Phase 0:  Architecture decision — Sequence vs Dispatcher+Performer (P-2); A-1 wiring
Phase 1:  Workflows + arguments — one entry per .xaml with In/Out/InOut signatures
Phase 2:  EACH app to inspect — one bullet per target app, e.g. "Inspect acme-test.com with Playwright"
Phase 2b: Write selectors.json (apps → screens → elements with taxonomy_type)
Phase 2c: generate_object_repository.py --from-selectors selectors.json --project-dir <each project>
Phase 2d: Pass --project-dir to every generate_workflow.py call (auto-wires obj_repo refs)
Phase 3:  Generate sub-workflows ONE AT A TIME — validate_xaml --lint after EACH file
Phase 3b: modify_framework.py wire-uielement <project> <AppName> for EACH app with UI
Phase 3c: (Dispatcher only) wire SCAFFOLD.DISPATCHER_GET_ITEM marker — generate snippet first (G-3)
Phase 4:  Wire Main.xaml — orchestration only, log bookends (A-7)
Phase 5:  Project-level validation — validate_xaml <project_folder> --lint
Phase 6:  Output Config.xlsx required keys grouped by sheet — validate_xaml --config-keys <project>

If Phase 2 is empty or missing, you are violating the skill rules. Even when no UI inspection is needed (no browser/desktop apps), state that explicitly: Phase 2: no UI inspection required (no browser/desktop apps in scope).

Per-phase reference anchors

Config layout rule of thumb

When listing required Config.xlsx keys (Phase 6), apply this rule before falling back to config-sample.md's decision flowchart:

• URLs, API endpoints, file/network paths, shared values across projects, or anything whose value may change frequently → Assets sheet. They are environment-specific, fetched from Orchestrator at InitAllSettings, and admins update them without redeploying. If you expect the value to be tuned post-deployment (timeouts, thresholds, recipient lists, feature flags, batch sizes), it belongs here too — Settings/Constants require a redeploy to change.
• Credential asset NAMES → Settings sheet. The string in the cell is just a name passed to GetRobotCredential; storing it in Assets would trigger a needless GetRobotAsset round-trip. The credential value itself never appears in Config.xlsx — fetch it at point of use.
• Project identity (queue name, process name, processes-to-kill list) → Settings sheet. Local to the project, never changes per environment.
• Framework knobs (retry counts, log message templates) → Constants sheet. Identical across dev/staging/prod.

validate_xaml --config-keys <project> auto-classifies via this rule and emits a per-key reason — review the suggestion before manually editing Config.xlsx.

Compat-v2 (version-band) requirement — ENFORCED: every scaffold_project.py invocation MUST include --band <stable_year> (e.g., --band 25). To pick the band: run python scripts/resolve_nuget.py --all and read the major year of UiPath.System.Activities in the output — that's the latest stable. Available band profiles live under references/version-profiles/UiPath.System.Activities/<year>.<minor>.json. Every dependency MUST be resolved via python scripts/resolve_nuget.py <package> — no version literals from training memory (G-5). If project.json is generated without versionBand, lints 120–122 will REJECT it. The 23.10.x and 24.x bands are end-of-life and MUST NOT be used for new projects in 2026.

Final gate

If you are about to present an architectural plan and have NOT produced a Phase 2: inspection list with one entry per target app (or an explicit "no UI in scope" line), you are violating skill rules — go back and add Phase 2 first.

When To Read Which Reference

Start here. Match the user's task to the right file, then read only what's needed. For files > 200 lines, use grep or line-range reads — never read entire large files or XAML assets.

Common Tasks (check these first)

Activity-Specific Reference

Expression Reference (large files — use targeted reads)

Project & Configuration

Ground Rules

Full rule definitions with rationale: references/rules.md. Rules below are referenced by ID.