prompt-workflow-architect

A Claude Code skill that interviews you before you write the prompt — so you end up with an outcome, guardrails, and a way for Claude to check its own work, instead of a numbered procedure that drifts.

The problem it solves

Two habits quietly waste most of the time people spend prompting.

You write a procedure. "First read the files, then group them, then write the output." That's what older, weaker models needed. Current models do better with the end state, the constraints, and a way to verify — then left to run. A procedure locks the model into your guess at the right path, and breaks the moment reality differs from your assumptions.

You skip verification. This is the expensive one. If Claude has no way to confirm its own work, you have to inspect the output yourself — which defeats the point of automating it. Unattended runs stall and drift for exactly this reason.

This skill forces the better shape through a short interview, then tells you whether the task should run as a single agent, a dynamic workflow, a local loop, or a cloud routine. It drafts only — it never creates or schedules a cron job, routine, or file on your behalf.

Who it's for

Operators who run real automations and are tired of prompts that work once and then quietly degrade: RevOps and CS leads, DTC and marketing-ops folks, and anyone standing up recurring agent jobs against tools they already pay for. You don't need to write code to use it.

Install

Claude Code — plugin marketplace (recommended)

/plugin marketplace add promptmetrics/prompt-workflow-architecture
/plugin install promptmetrics-prompt-workflow@prompt-workflow-architecture
/reload-plugins

The slash command is /promptmetrics-prompt-workflow:prompt-workflow-architect. It also triggers on its own when you ask a prompt- or workflow-shaped question — no slash needed.

Claude Code — local clone

git clone https://github.com/promptmetrics/prompt-workflow-architecture
cd prompt-workflow-architecture
claude --plugin-dir .

Or copy the skill into your personal skills directory so it loads in every project:

cp -r skills/prompt-workflow-architect ~/.claude/skills/

Claude.ai

Prerequisite: enable Code execution and file creation in Settings ▸ Capabilities.

Download prompt-workflow-architect.zip from the v0.1.0 release, or build it locally with ./scripts/build-zip.sh.
Open Customize ▸ Skills → + → + Create skill → Upload a skill → upload the zip.
Enable it in Settings ▸ Capabilities.

Known limitation. This skill's description is 878 characters — well under Anthropic's documented 1024-character limit, and deliberately long so its trigger phrases stay in the matcher. Some upload paths enforce a tighter ceiling. The Claude Code plugin install above is the verified route; if Claude.ai rejects the zip on description length, that's why. Shorten the description: line in SKILL.md and rebuild if you hit it.

Claude Cowork

Prerequisite: enable Code execution and file creation in Settings ▸ Capabilities.

Standalone skill: upload the same zip via Customize ▸ Skills.
Bundled in a plugin: add the skill to a Cowork plugin via Customize ▸ Plugins.
Org-wide: an owner enables sharing in Organization settings ▸ Skills, then the skill's Share action shares it with specific people or the whole org.

What a session looks like

You bring a rough idea — "I want our changelog to stop going stale." The skill asks one question at a time:

Outcome, not steps. What does the end state look like? If you answer with a numbered procedure, it pushes back and makes you restate it as a result. It won't proceed until you do.
Scope. One-time or recurring? One-time branches on size (small → single agent, large or multi-stage → dynamic workflow). Recurring branches on locality and cadence (needs your machine → loop, can run in the cloud → routine).
Verification. How will Claude know it's done without you checking? This is a hard gate. "I'll see if it looks right" gets rejected, and it keeps asking until there's something runnable — a test suite, a diff against a reference, a concrete command whose output you can check.
Guardrails. What must not be touched or broken? "Nothing" is fine, but it gets recorded explicitly rather than dropped.
Existing prompt. Is something already covering this? If yes, the output tells you to remove the old instruction first and only add pieces back if the same failure actually repeats — instead of stacking two instructions that conflict in ways you can't attribute.

You get back one markdown block: the prompt as prose, the execution mode with a one-line reason, the invocation snippet, and — if relevant — the ablation note.

Full detail, including the mode decision table and a complete worked example: docs/how-it-works.md.

Repository contents

prompt-workflow-architecture/
├── README.md                       # this file
├── LICENSE                         # MIT
├── CONTRIBUTING.md                 # how to change the skill and cut a release
├── CHANGELOG.md                    # version history
├── .claude-plugin/
│   ├── plugin.json                 # promptmetrics-prompt-workflow plugin manifest
│   └── marketplace.json            # prompt-workflow-architecture marketplace manifest
├── docs/
│   └── how-it-works.md             # the five steps, mode decision table, worked example
├── scripts/
│   └── build-zip.sh                # builds dist/prompt-workflow-architect.zip for upload
└── skills/
    └── prompt-workflow-architect/
        ├── SKILL.md                # the interview, output contract, guardrails — model-facing
        ├── references/
        │   └── output-template.md  # exact output structure, loaded only when needed
        └── evals/
            └── evals.json          # trigger / non-trigger test cases

Per Anthropic's skills guidance there is deliberately no README.md inside the skill folder — human-facing documentation lives here and in docs/, so only SKILL.md and references/ ship as model context.

Compatibility

No scripts, no MCP servers, no credentials, no network calls — prompt-workflow-architect is a pure conversational skill, and it writes no files. That's why the same folder ships unchanged to Claude Code, Claude.ai, and Claude Cowork, following the open Agent Skills spec. The .claude-plugin/ manifests are read only by Claude Code's plugin loader; Claude.ai and Cowork use the uploaded zip directly.

Contributing

See CONTRIBUTING.md for the description-length rules, how to run the eval cases, and the release checklist.

Prompt Workflow Architecture

At a glance

README