Skip to content

kontourai/matt-pocock-engineering

v1.0.0MIT

Station-curated visual introductions to Matt Pocock engineering Skills, pinned to an attributed MIT source revision.

GitHub

Station-curated Matt Pocock engineering collection

This ordinary Agent Plugin supplies five visual entry points: Clarify an idea, Clarify project decisions, Write a specification, Break a specification into tickets, and Implement reviewed tickets. Each can be selected independently. Authored optional stage choices prepare the next draft in the same conversation; they never send or implement automatically. The package is Station-curated; Matt Pocock authored the upstream Skills, not these visual interfaces, and no endorsement is implied.

The source is mattpocock/skills at d81f3a183412e71a5b1e84ca21bc1a35eea03a60, under the retained MIT license. Source review records the conservative dependency graph and gaps; adaptations records the portable frontmatter changes. Every definition maps each bundled Skill to its original path and revision. Immediate-child skills/<name>/SKILL.md materialization follows ordinary Agent Plugins discovery; there is no core special case for this collection.

With a CLI built from this source revision on Node 24:

# In this directory:
station plugin build
# Against a running local Station using an isolated home:
station plugin preview /absolute/path/to/matt-pocock-engineering
station plugin install /absolute/path/to/matt-pocock-engineering

Author validation has been executed. It establishes schema, portable Skill parsing and pinned bundled bytes. One recorded browser/model journey at source1bd clarified a local Focus Board idea with Codex, produced a reviewed specification and published six local tickets after approval. Reload at source746 retained its three immutable stage snapshots. Questions used ordinary chat text; this receipt does not qualify canonical questionnaire answers, later composer changes, application implementation, physical devices or a released collection. Only that observed evaluation case passed; other planned cases remain not-run. Native, rich-pane, registry and release evidence retain their own scopes as explained in the author learning path.

Move an idea into engineering work

Start with a real idea. Clarify an idea delegates into grilling, which asks the independent decision frontier, waits for answers, recomputes the next frontier and ends only when the frontier is empty. It must wait for shared understanding before acting. Environment facts belong to agent exploration and subagents; a provider without that capability is unsuitable. The optional visible decision summary is Station-added; upstream promises shared understanding, not an exported summary.

Use Clarify project decisions when resolved terms should update the project glossary and consequential choices warrant ADRs. It additionally loads domain-modeling; glossary updates happen inline, while ADRs are offered only under its three criteria. Project access and write permissions still apply.

Continue the same conversation by explicitly asking for to-spec, naming the settled decisions. In a new conversation, supply the reviewed decision or spec reference and actual supporting artifacts. The skill synthesizes current context without another interview, explores the project glossary/ADRs, confirms testing seams, writes its extensive source template and publishes to the configured issue tracker. Missing tracker configuration stops for the user to select setup-matt-pocock-skills; the interface does not invent configuration or claim publication from draft prose.

Next explicitly select to-tickets and name the reviewed specification. It reads the full reference/comments, drafts tracer-bullet vertical slices and blocking edges, and quizzes the user about granularity/edges before publishing. For local tracking it writes one file per ticket in dependency order; for a real tracker it publishes real identifiers and relationships. It must not close or modify a parent issue. The configured tracker, local files and actual publication actions retain the agent's normal authority boundary.

Implementation is a separate user decision. After reviewing the work breakdown, explicitly choose Implement reviewed tickets, supply the selected ticket references and confirm Project, branch, tools and pre-agreed testing seams before sending. This authored stage calls the bundled implement entry; it is never an automatic transition from tickets. It uses tdd where possible, checks the work, invokes code-review and commits to the current branch. Installing or finishing a visual experience grants none of these actions. Starting implementation does not reset reviewed decisions. There is no automatic funnel or alternate session lifecycle.

Source adaptations and limits

Upstream's disable-model-invocation frontmatter is outside the portable Agent Skills SDK grammar. Adapted Skill files remove only that field; originals are retained under upstream/, and agents/openai.yaml policies remain. These are explicit-user entry points. Never automatically invoke setup, implementation or another user-invoked Skill. The visual adapter preserves selection and stop instructions; it does not establish enforcement of every provider's implicit invocation policy. Inspect that provider's actual policy support before using portable automatic discovery.

The conservative inspector includes setup/router references, so this package bundles the full selected 16-Skill graph and the linked wizard template rather than assuming only three wrapper files suffice. Templates may mention example project paths or hypothetical slash routes; these remain source review leads, not required tools. The source parser also reports the unselected upstream pr Skill's nested metadata as incompatible; it is not installed here.

Some source dependencies create scripts/HTML reports, open browsers, use CDN assets, run subagents, inspect Git history, or contact an issue tracker. Source presence does not execute or qualify those paths. No template is executed while authoring or validating. Setup asks before changing the project config.

The interview descriptor allows up to 12 canonical questions per rendered round. This is a Station presentation descriptor, not the source frontier limit. Larger or prose-only frontiers stay answerable in chat without dropping questions. Guided/alongside presentation must preserve the same requests, answers, decisions and artifacts. Definitions declare suitability, not grants.