Skillstead
English · 한국어
Practical, portable skills for agentic work — create clearer artifacts, check public claims, guide safer GitHub
releases, turn rough or translated text into natural, precise writing, and interpret portraits.
[!TIP]
Skillstead = skill + homestead. A small, durable place for skills that coding agents can carry into
real repositories. Each public support claim is tied to examples and runtime evidence.
Start here
- Choose one skill for the task in front of you. Each skill works independently.
- Follow the installation guide to copy that skill's complete folder at a pinned version.
- Open the skill's README, copy a request that is close to your task, and adapt it to your material.
- Check the skill's stated limits before relying on the result in a release or other consequential workflow.
Looking for a particular guide? The documentation index separates user guides from
maintainer references.
Choose a skill
| Skill | Best for | Version | Runtime support | Maturity |
|---|
svg-infographic | Turning architecture notes, process flows, comparisons, and technical concepts into editable SVG + verified 2× PNG | 0.12.0 | Supported: Claude Code + Codex | Stable |
docs-claim-check | Checking whether public documentation claims are supported by supplied evidence | 0.11.0 | Claude Code | Stable |
github-release-guide | Guiding a private repository's first public transition and every later version release, with separate approval before each change | 0.9.1 | Supported: Claude Code + Codex | Stable |
writing-quality-editor | Composing and revising user-facing text, plus natural English↔Korean adaptation, without inventing or changing facts, intent, voice, or operational constraints | 0.19.0 | Supported: Claude Code + Codex | Stable |
street-portrait-artist | Creating a caricature, pen-and-watercolor portrait, or drawn paper-cut figure from supplied photos | 0.2.0 | Supported: ChatGPT + Codex | Experimental |
Each skill is self-contained and can be installed independently. You do not need to install the entire
catalog—copy only the complete folder for the skill you want to use. See
docs/INSTALL.md for global/project paths, pinned tags, clean updates, Windows commands,
and the per-skill runtime matrix.
The Version column above is per skill, not a catalog version. See
docs/VERSIONING.md for what it means and how it changes.
For more examples—including natural requests, the WQE shorthand, and requests that involve more than one
skill—see the repository-only
intent and invocation contract.
GitHub's Latest badge identifies the most recently published individual skill release. It does not represent
a catalog version.
Highlights
Explore the SVG gallery
svg-infographic covers more than conventional architecture boxes. The six above were chosen for how
differently they are shaped; they clear the lint, layout and typography gates, and they predate the TypePack
receipts, so no receipt is claimed for them. The gallery behind the image carries the nine TypePacks with their
receipts alongside these. The contact sheet intentionally keeps an English evidence frame around the locale-specific
artifacts; the live gallery localizes the full page. Open the live gallery, or
browse the legacy examples and the verified TypePack catalog.
Use one skill—or combine them across a release
Each skill installs and works independently. When a project needs a wider path, use
writing-quality-editor for clear prose, svg-infographic for visual explanation, docs-claim-check for
evidence-bounded public claims, and github-release-guide for approval-gated release decisions. This is not a
required pipeline: start with the skill you need, skip the others, and recheck an earlier artifact when it changes.
street-portrait-artist is a separate creative workflow for character portraits from supplied references; it does
not need to participate in that release path.
Skill details
svg-infographic
Technical diagrams often begin as prose and end as hard-to-edit screenshots. svg-infographic computes a
layout before drawing, authors structured SVG, checks the source, then exports a dimension-verified 2× PNG.
Use it for architecture and cloud topology, process or approval flows, before/after migrations, roadmaps,
layer models, qualitative matrices, and Korean/CJK-ready technical one-pagers.
docs-claim-check
Release-facing docs can sound certain even when their evidence is partial or stale. docs-claim-check splits
objective statements into atomic claims and labels each one verified, unsupported, stale-suspected, or
needs-human within an explicit reviewed-input scope.
Use it before publishing a README, install guide, release note, or announcement. It is advisory only: the
contract runs no commands during assessment and does not generate fixes, code review, or security verdicts.
- Friendly guide:
docs-claim-check README
- Validation material: synthetic AcmeTask fixture and worked assessment
- Name the skill:
Use docs-claim-check to assess these release-note claims against the supplied tag and CI evidence.
- Or ask naturally:
Check whether these README claims are supported by the evidence below. Report findings only; do not rewrite the document.
github-release-guide
GitHub releases combine documentation work with changes to visibility, branches, tags, settings, and GitHub
Releases that can be difficult to undo. github-release-guide first checks readiness without changing the
repository. It then shows each proposed change, checks the current state again, and asks for approval in two
distinct scopes — running a repository command, and changing the repository — before verifying the result and
moving on.
V1 can be used at two points: when an existing private github.com repository becomes public for the first time,
and whenever that public repository publishes a new version afterward. It does not bootstrap repositories,
publish packages, sign binaries, deploy cloud services, claim a security audit, force-push, or rewrite history.
First, pick the mode. Assess inspects without changing anything; Guided starts only after Assess is
complete, release-critical blockers are resolved, and you explicitly choose to switch.
Then, in Guided, one change at a time. Approval is asked in two distinct scopes: running a command is
approved separately from changing the repository, and neither implies the other.
- Friendly guide:
github-release-guide README
- Validation material and diagrams: synthetic scenarios, answer key, and worked outputs
- Natural readiness check:
Check whether this public GitHub repository is ready for its next version release. Inspect only and do not change it.
- Assess example:
Use github-release-guide in Assess mode for this public repository's upcoming version release.
- Guided example:
Use github-release-guide in Guided mode to prepare this private repository for first publication. Start with Assess, then show only the first proposed change. Do not change the repository until I approve that exact step.
- Safety boundary: immediately before a repository becomes public, the guide explains what cannot be undone and
asks the user to approve that visibility change separately. The release decision remains with the user.
writing-quality-editor
writing-quality-editor writes new text from notes and evidence or makes existing prose easier to understand.
Its editing rules require preserving facts, intent, author voice, commands, conditions, limitations, risks, and
next actions. Those rules do not make every result natural or accurate; important writing still needs a final review.
For wording and naturalness requests, it now starts with the smallest complete span that has a concrete reader
problem and leaves the surrounding text unchanged. Ambiguous discretion, approval, or notification wording stays unchanged
for human judgment while other safe edits continue.
Its Adapt mode rewrites between English and Korean for the target-language reader instead of copying the source
sentence structure. It may change information order, sentence rhythm, idioms, and explanation density, but it
does not invent claims or hide ambiguity. AI-detector gaming and provenance concealment are explicit non-goals.
- Friendly guide:
writing-quality-editor README
- Validation material: 36 scenarios and a separate answer key
- Name the skill:
Use writing-quality-editor to make the document below read naturally. Preserve its core facts, conditions, and requirements.
- Use the shorthand:
Use WQE to review this onboarding guide. Identify problems, but do not revise it yet.
- Or ask naturally:
Review this README. Do not revise the prose yet. · Write a new README using only information supported by the material below. · Rewrite this English release note so it reads naturally to Korean readers. Preserve its meaning and conditions.
- Specify a mode only when needed:
Use writing-quality-editor in Assess mode to review this release note. Do not draft revisions.
The versions listed above include a conditional WQE paragraph-flow check and portrait consistency, Editorial
Watercolor, and Paper-Cut Illustration. See the respective skill guides for development evidence and limitations.
street-portrait-artist
street-portrait-artist studies face shape, feature spacing, expression, and the main distinguishing feature before
drawing. It uses those observations for a kind, near-monochrome Street Caricature, a pen-and-watercolor
Romance Watercolor portrait, or a flat-color Paper-Cut Illustration in the supplied scene. All modes are
designed to retain the same person’s recognizable features.
Provide one or more clear portrait references. Supplied portraits, analysis, and outputs remain task-scoped unless the
user separately asks to publish or retain them.
- Friendly guide:
street-portrait-artist README
- Public-safe gallery: two synthetic Twin Portrait examples
- Name the skill:
Use street-portrait-artist to make a kind Street Caricature from my attached portrait.
- Or ask naturally:
Turn this portrait into a delicate pen-and-watercolor character portrait. Keep the jaw, hairline, and expression recognizable.
Playbooks (maintainer reference)
For changes to Skillstead itself, start with the maintainer workflow:
document alignment, root changelog, adopted personal-skill synchronization with commit and push, version
announcements, blog follow-up, and optional social drafts.
playbooks/public-release contains the canonical public-release playbook:
generic checklists and templates for taking a private repository public and verifying it afterward. These are
reference documents for maintainers — not installable skills, and installing any skill never requires them.
The github-release-guide skill mirrors the playbook's rules inside its own self-contained package. English is
canonical, and each document has a Korean mirror with the .ko.md suffix. Update both languages in the same pull
request whenever the meaning changes.
playbooks/skill-development defines the repository standard for
designing, validating, independently reviewing, releasing, and retiring a skill. Its package template is
deliberately non-installable until the reserved sample-skill identity is replaced. The tracked retirement and
major-transition evidence enforced by the release gates is specified in
docs/VALIDATION.md.
Quality and evidence bar
Every public skill must have:
- a clear description of what it does and does not do,
- synthetic, non-client validation material,
- runtime support and maturity labels limited to what was actually tested,
- public paths free of credentials, private provenance, and host-specific data,
- and a repeatable validation path appropriate to its output.
Runtime support is per skill, not catalog-wide.
svg-infographic has passed the same three frozen, fresh-context briefs on Claude Code and Codex. The checks
covered project-local discovery, source lint before rendering, exact 2× browser export, Korean/CJK text,
containment, connector semantics, and fail-closed recovery when browser launch crossed a sandbox boundary.
Codex evidence includes macOS Codex CLI and a fresh Codex App task in a Windows 11 ARM64 VM. It is
Supported for Claude Code and Codex within that recorded evidence scope.
github-release-guide has passed clean Claude Code/Codex material-parity checks, the disposable first-public
live E2E, pinned v0.5.0 project installation and discovery smoke, and the final strict claim audit. It is
Supported for Claude Code and Codex within that recorded evidence scope.
writing-quality-editor is Stable as of 0.18.0, following bounded validation and explicit owner acceptance.
Its existing Supported labels for Claude Code and Codex remain within the recorded evidence scope. The fixed
candidate was checked through eight regression outputs and forty final outputs on the two CLIs, including
actual package reads and same-conversation follow-ups. These checks do not establish universal discovery,
perfect preservation, general superiority, or new runtime support. Historical failures and remaining wording
limitations are retained in the WQE guide.
street-portrait-artist has two public-safe synthetic Twin Portrait cases that document its intended visual direction
and privacy boundary. Fresh installations of the published 0.1.0 package were then discovered and invoked in ChatGPT and
Codex with synthetic portrait input, reference-image generation, fail-visible size fallback, and output delivery. Both
runtimes are Supported within that recorded scope. Maturity remains Experimental because likeness and the refined
visual behavior have not passed a broad cross-subject regression set.
Current limitations
svg-infographic browser rendering is verified on macOS and in the recorded Windows 11 ARM64 VM.
This does not claim every Windows machine or filesystem; Linux rendering remains documented but unverified.docs-claim-check is advisory and evidence-bound; it does not execute verification commands.github-release-guide v1 is github.com-only. It covers the one-time private-to-public transition and each
version release after the repository is public.- For
github-release-guide, a clean assessment only means that no issue was found within the inspected scope;
it does not prove that the repository being assessed has no secrets, private information, or security risks.
writing-quality-editor is designed as locale-neutral, but its initial localization fixtures cover only
English↔Korean (ko-KR for Korean output). Other locale pairs are not yet claimed as behaviorally validated.street-portrait-artist likeness and image quality remain non-deterministic, exact pixel export depends on the
active image surface, and final artwork requires human review. The recorded ChatGPT and Codex support does not
promise exact dimensions or consistent visual quality across different image-generation runs.- Agent output is non-deterministic.
writing-quality-editor defaults to local-first edits and holds ambiguous
wording for review, but it cannot guarantee that every run avoids preference-driven changes outside the marked
spans. Review the final delta before publishing or relying on an important document.
License
Apache-2.0.
