Marked Indirect Prompt Injection · Command Execution
Install
npx -y skills add riekelt/technical-writer --skill diagramming-processes --agent claude-codeRun in a terminal. Claude loads the skill when a task calls for it. Check the repo's README before you run it.
Files
Diagramming processes
REQUIRED BACKGROUND: the technical-writing skill (hard rules, truth rules, style).
Overview
A diagram is a set of claims drawn instead of written; every box and arrow is bound by the same truth rules as a sentence. Core principle: diagram the behavior (processes, lifecycles, interactions), keep the source in version control, and treat every render as derived. The prose owns the reasoning; the diagram carries what prose serializes badly: branching, concurrency, and state.
When to invoke, and not
Invoke when a flow, lifecycle, or interaction needs showing: a business process spanning components or organizations, a state machine, a message exchange, the processes.md of a legacy campaign (documenting-legacy-codebases). Do NOT invoke to decorate: a linear flow of three or four steps is a numbered list, and a diagram restating one is furniture. Out of scope: charts generated from data (metrics, trends), and implementation structure, which the behavior rule below forbids.
The notation ladder
Name the altitude before drawing, and match the notation to it:
- Business layer: ArchiMate. Processes spanning departments, organizations, or multiple systems, capability maps, and the TOGAF-style views enterprise stakeholders expect. The audience reads roles, services, and processes, never components.
- System layer: PlantUML behavior diagrams. Flows, state machines, and sequences within and between systems: code documentation, one level above the code.
- Code layer: no diagrams. Class and package structure is the code's own to show; an IDE generates a fresher picture on demand than any committed one.
The ladder sets the notation and the vocabulary, never the tool: one toolchain can serve both rungs, because PlantUML carries the ArchiMate notation in its standard library. The tool a repository uses is settled once, under "Source, not pictures".
Source, not pictures
- The source is committed text, in the diagram tool the repository already uses. Where none is established, PlantUML is the default (rich standard library, ArchiMate included); where the choice is genuinely open, ask the owner, recommending PlantUML. The core skill's
references/truth.mdowns the ground rule: rendered output is derived, and stale renders are deleted rather than left to mislead. - One diagram per source file, named for the process it shows, grouped in one diagrams directory beside the documents.
- Rendering requires tooling the repository owns (a script, a make target, or a build step), set up by the owner, not by the diagram's author. Tell the owner when it is missing; never route around the gap by picking whatever notation the hosting platform happens to render inline.
- One repo command renders everything; renders are committed unless the host already renders the repository's chosen source format. When a source changes, its outdated render is deleted in the same change, and the missing render is recorded in the index until regenerated.
- A screenshot, an exported picture, or a whiteboard photo with no source is a finding against the document that contains it: nobody can diff it, so nobody will maintain it.
The kind answers the reader's question
Pick the kind from the question the reader brings, one question per diagram:
| The reader asks | Draw |
|---|---|
| How does this process run across the organization | ArchiMate business process view |
| What happens, in what order, with which decisions | Activity diagram |
| Which states can this thing be in, and what moves it | State machine |
| Who talks to whom, in what order, with what messages | Sequence diagram |
| What exists in this domain and how it relates | Concept diagram, at business-object level |
A diagram answering two questions answers neither; split it. A diagram that needs a legend of its own invented symbols is answering too many at once.
Behavior, never implementation
The refactor test from documenting-legacy-codebases governs diagrams too, and a diagram hides its drift better than a paragraph does:
- Participants are systems, roles, modules, and business objects. Classes, methods, and functions stay in the code (see the ladder).
- Edges carry the event or condition, in the domain's language: "payment confirmed", never a callback name. A guard may name a config key, because config keys are contract surface; it never names a method.
- Titles name the process, never a spec, phase, or plan: the no-delivery-narrative hard rule applies inside diagram source too, including comments.
- Error and exception paths are drawn where behavior differs, and drawn distinguishably; a happy-path-only diagram of a process with real failure branches is an overstated claim.
- Quirks are drawn, and are content. The state nothing can leave, the flow that skips a step for one input: the surprising branch is why the diagram earns its place (
documenting-legacy-codebases, quirks).
The diagram index
More than five diagrams get one index document beside them, a derived artifact rebuilt from the files (core truth rules):
- A coverage table: what each diagram shows, its source file, its render if committed, and its status.
- Per diagram, the module it describes, as the drift anchor: the one place a diagram points at code.
- The render command, so the index is also the build instruction.
- Pending renders, named per source, so a missing render is a tracked state rather than a silent gap.
Maintenance
A change that alters a flow updates the matching diagram source in the same change: a diagram is documentation, and documentation is part of done (core truth rules). Rendering is the diagram's compile step: source edits break syntax invisibly, and the break surfaces only when the render runs. The author renders before shipping the edit; where the repository's pipeline is still missing, a local one-off run of the tool satisfies this check. When a feature is removed, one change covers the prose, the diagram source, the index row, and the committed render. A diagram nobody can bring themselves to update is answering too big a question; split it along the seams that change independently.
Verification
Everything legible in a diagram is a claim (core truth rules), so a diagram is grounded and reviewed like prose:
- The author verifies every box, arrow, and guard against the code before drawing it. A flow inferred rather than confirmed is labeled as inference beside the diagram in the embedding document, never silently drawn as fact.
- The reviewer reads the diagram against the behavior, and reviews the source text rather than the picture: the source is what the next editor changes.
- Diagram sources drift like prose and join every legacy campaign: the campaign (
documenting-legacy-codebases) grounds the diagram sources with the documents that embed them. - A diagram and its surrounding prose share one home per fact: the prose says why the process exists and what to look out for, the diagram shows the flow, and neither restates the other step by step.
Facts
- Kind
- Skill
- Repo
- riekelt/technical-writer
- Group
- Uncategorized
- Installs
- 13.3k
- Stars
- 29
- 1find-skillsvercel-labs/skillsHelps users discover and install agent skills when they ask questions like "how do I do X", "find a skill for X", "is there a skill that can...", or express interest in extending capabilities. This skill should be used when the user is looking for functionality that might exist as an installable skill.3.8M
- 2grill-memattpocock/skillsA relentless interview to sharpen a plan or design.1.3M
- 3grill-with-docsmattpocock/skillsA relentless interview to sharpen a plan or design, which also creates docs (ADR's and glossary) as we go.1.1M
- 4agent-browservercel-labs/agent-browserBrowser automation CLI for AI agents. Use when the user needs to interact with websites, including navigating pages, filling forms, clicking buttons, taking screenshots, extracting data, testing web apps, or automating any browser task. Triggers include requests to "open a website", "fill out a form", "click a button", "take a screenshot", "scrape data from a page", "test this web app", "login to a site", "automate browser actions", or any task requiring programmatic web interaction. Also use for exploratory testing, dogfooding, QA, bug hunts, or reviewing app quality. Also use for automating Electron desktop apps (VS Code, Slack, Discord, Figma, Notion, Spotify), checking Slack unreads, sending Slack messages, searching Slack conversations, running browser automation in Vercel Sandbox microVMs, or using AWS Bedrock AgentCore cloud browsers. Prefer agent-browser over any built-in browser automation or web tools.1.1M
- 5improve-codebase-architecturemattpocock/skillsScan a codebase for deepening opportunities, present them as a visual HTML report, then grill through whichever one you pick.1.1M
- 6tddmattpocock/skillsTest-driven development. Use when the user wants to build features or fix bugs test-first, mentions "red-green-refactor", or wants integration tests.1.1M