Power Platform

power-platform · an Agent Skill by SkillEra for Claude Code, Codex, Cursor and any agent that reads SKILL.md

Power Platform development the way software is built, and proved in the product. Canvas apps, Dataverse solutions and Power Automate flows with the definition in git, a portable artifact built from it, a deliberate deployment, and every change proved by performing the task in the published app, driven by Playwright.

Free
Open source · MIT
Public beta
Latest tagged version: v0.7.1 (opens in a new tab)

View on GitHub Install

What you seeWhat actually happened

pac solution pack exits 0: A component was dropped from the zip
The canvas compile passes: It validated against no data sources at all
The import reports success: It imported whatever zip it was handed, including the wrong one
A flow saves and turns on: It listens to the wrong event, or re-triggers itself on its own write
A button does nothing: One failing step abandoned the rest of the formula without a message
It works in Studio Preview: The published app resolves against a cached copy of your table metadata that no longer matches Dataverse

Almost every Power Platform failure is silent. A change can compile clean, work in Preview, pass every audit, and do nothing for every user.

A clean compile is not enough

This skill is a working method built from those failures. Every trap it describes comes with its signature, so a symptom is recognised in minutes instead of days, and every step of shipping states what it proves and what it does not.

It loads on its own for Power Platform tasks — a Save button that works in Studio but does nothing in the published app, a picker over a large Dataverse table, solution flow JSON, or setting up a repository for canvas app and Dataverse work.

From a change to proof in the published app

Described from the README at v0.6.0 (opens in a new tab) — a newer tagged version is available on GitHub, and may change what follows.

Every non-trivial change moves through the same cycle. Each step proves one thing, and skipping one is how a change reaches users unproven.

1Pre-flight: branch, clean tree, pac org who, read the state fileProves: You are changing the right thing in the right environment
2Specify: who uses it, what proves it, what it touchesProves: There is a test before there is a change
3Build in source: .pa.yaml, solution XML, flow JSON; hooks check each writeProves: The known compile-killers are absent
4Audit: a stale input is unverified, not a passProves: Project rules hold on current inputs
5Compile against a live Studio session, and read the first line of the resultProves: Formulas bind to real data sources
6Build on the live manifest and assert on the artifactProves: The zip contains every component and marker you changed
7Import, then confirm what landed: LoadFromYaml, markers in the half that runs, data-source countProves: The environment runs what you built
8Perform the task in the published app: fresh build, right role, effect checked in the database, data restoredProves: The feature works for the person who uses it
9Document in the same changeProves: The next session starts from the truth
10Refresh the audit inputsProves: The next audit describes the app that now exists

What you get

The method and references need nothing installed. The tools need Node 20 or later, Python 3 and the Power Platform CLI; Playwright is optional, for browser verification.

The method: SKILL.md: ten non-negotiables, a ten-step working loop, and a routing table into the references.
18 references: Self-contained guides loaded only when a task needs them: canvas shipping, manifest caches, Power Fx and .pa.yaml, controls, layout, browser verification, Dataverse, the Web API, security, data migration, flows, CI/CD pipelines, audits and more.
Five tools: A ship pipeline, a metadata drift checker, an artifact inspector, a cloud-flow linter, and a Playwright driver for Studio and the published player. Each proves it can fail with --selftest.
Four hooks: Claude Code hooks that stop known compile-killers at the moment a file is written, plus a session pre-flight and an end-of-turn audit.
Templates: Hook settings, config examples, an example browser scenario, and state, decisions and dependency templates.

Measured, with and without it

Measured on version 0.5.1. On ten realistic Power Platform tasks, run twice each, the same model passed:

With the skill: 133/148; graded checks (90%)
Without it: 98/148; graded checks (66%)

The same model and prompt ran twice with the skill and twice without, in fresh sessions, graded against checks fixed before the runs.

The cost is real: about 2.9 times the tokens, 1.2 times the output, and about 26 seconds more per task. Nothing ran against a live tenant. Four checks still fail with the skill.

The evaluation report, every check and the limits of the measurement

The published app is the test harness

A canvas app has no test framework. Opening Studio, adding a data source, capturing a network trace and running the app are all browser work, and a browser can be driven. Verifying in the published app is the standard here, not a fallback.

The skill never installs anything silently. Without a browser, the change is reported as unverified, not passed.

Prefer the published player to Studio Preview: Preview runs live metadata; users run the published app and its cached copy.
Clear the stale player first: The player serves a cached build from IndexedDB behind a late-arriving banner; the scenario asserts the build stamp.
Perform the task as the role that uses it: An admin session proves nothing about a restriction. If only an admin account is available, the result is reported as unverified.
Confirm the effect where it lands: Read the row back through the Web API, not the screen that wrote it.
Restore what it wrote: A scenario that writes must declare a restore, and the driver refuses to run it without --allow-writes.

Install

Then open your Power Platform repository in your agent and ask for real work.

Or copy skills/power-platform/ to ~/.claude/skills/power-platform/, or to the equivalent skills folder for your agent. The hooks are optional and recommended.

Any agent that reads SKILL.md

Claude Code, Codex, Cursor and others, through the skills installer.

npx skills add KDavidP1987/power-platform-skill

Claude Code plugin

Add the marketplace once, then install the plugin.

/plugin marketplace add KDavidP1987/power-platform-skill
/plugin install power-platform@power-platform-skill

Rules it will not bend

The repo is the source of truth. Studio is never the last place a change was made.
Assert on the finished artifact, never on an exit code.
Verify by performing the task in the published app, as the role that uses it.
Verify the effect where it lands: in the table, not on the screen that wrote it.
Schema before screens, and never in one step.
Security roles stay out of the solution.
A flow that writes its own trigger table guards on a value that write changes.
Before any bulk write, count the messages and park the sender.
An audit that can pass vacuously will: floors, self-tests, stale-input detection.

Some limits of this version

The ship pipeline has run real ships on one app. More apps will find more cases; its refusals print the remedy they know.
The cloud-flow linter is still a floor: anything it does not understand is treated as unknown, which never counts as a guard. Activation is still the only compile.
Browser selectors track today's player and Studio, and Microsoft changes both. The driver fails loudly rather than passing when a selector stops matching.
Model-driven apps, Power Pages, Copilot Studio and Power BI are outside the method, apart from model-driven forms edited by script.

See every check

The evaluation report shows every check, what went wrong without the skill, and the limits of the measurement. The skill is MIT licensed; issues and pull requests are welcome on GitHub.

The evaluation View on GitHub

Loading…

Power Platform

power-platform · an Agent Skill by SkillEra for Claude Code, Codex, Cursor and any agent that reads SKILL.md

A clean compile is not enough

From a change to proof in the published app

Described from the README at v0.6.0 — a newer tagged version is available on GitHub, and may change what follows.

Every non-trivial change moves through the same cycle. Each step proves one thing, and skipping one is how a change reaches users unproven.

1Pre-flight: branch, clean tree, pac org who, read the state fileYou are changing the right thing in the right environment

2Specify: who uses it, what proves it, what it touchesThere is a test before there is a change

3Build in source: .pa.yaml, solution XML, flow JSON; hooks check each writeThe known compile-killers are absent

4Audit: a stale input is unverified, not a passProject rules hold on current inputs

5Compile against a live Studio session, and read the first line of the resultFormulas bind to real data sources

6Build on the live manifest and assert on the artifactThe zip contains every component and marker you changed

7Import, then confirm what landed: LoadFromYaml, markers in the half that runs, data-source countThe environment runs what you built

8Perform the task in the published app: fresh build, right role, effect checked in the database, data restoredThe feature works for the person who uses it

9Document in the same changeThe next session starts from the truth

10Refresh the audit inputsThe next audit describes the app that now exists

What you get

The method and references need nothing installed. The tools need Node 20 or later, Python 3 and the Power Platform CLI; Playwright is optional, for browser verification.

The method: SKILL.md: ten non-negotiables, a ten-step working loop, and a routing table into the references.
18 references: Self-contained guides loaded only when a task needs them: canvas shipping, manifest caches, Power Fx and .pa.yaml, controls, layout, browser verification, Dataverse, the Web API, security, data migration, flows, CI/CD pipelines, audits and more.
Five tools: A ship pipeline, a metadata drift checker, an artifact inspector, a cloud-flow linter, and a Playwright driver for Studio and the published player. Each proves it can fail with --selftest.
Four hooks: Claude Code hooks that stop known compile-killers at the moment a file is written, plus a session pre-flight and an end-of-turn audit.
Templates: Hook settings, config examples, an example browser scenario, and state, decisions and dependency templates.

Measured, with and without it

Measured on version 0.5.1. On ten realistic Power Platform tasks, run twice each, the same model passed:

With the skill: 133/148; graded checks (90%)
Without it: 98/148; graded checks (66%)

The same model and prompt ran twice with the skill and twice without, in fresh sessions, graded against checks fixed before the runs.

The cost is real: about 2.9 times the tokens, 1.2 times the output, and about 26 seconds more per task. Nothing ran against a live tenant. Four checks still fail with the skill.

The evaluation report, every check and the limits of the measurement

The published app is the test harness

The skill never installs anything silently. Without a browser, the change is reported as unverified, not passed.

Rules it will not bend

The repo is the source of truth. Studio is never the last place a change was made.

Assert on the finished artifact, never on an exit code.

Verify by performing the task in the published app, as the role that uses it.

Verify the effect where it lands: in the table, not on the screen that wrote it.

Schema before screens, and never in one step.

Security roles stay out of the solution.

A flow that writes its own trigger table guards on a value that write changes.

Before any bulk write, count the messages and park the sender.

An audit that can pass vacuously will: floors, self-tests, stale-input detection.

Some limits of this version

The ship pipeline has run real ships on one app. More apps will find more cases; its refusals print the remedy they know.

The cloud-flow linter is still a floor: anything it does not understand is treated as unknown, which never counts as a guard. Activation is still the only compile.

Browser selectors track today's player and Studio, and Microsoft changes both. The driver fails loudly rather than passing when a selector stops matching.

Model-driven apps, Power Pages, Copilot Studio and Power BI are outside the method, apart from model-driven forms edited by script.