Matt Pocock releases skills v1.3 with /retro transcript reviews

• Skills v1.3 ships /retro for reviewing old coding-agent transcripts, as Matt Pocock, the collection's author, announced in the release post.
• The migration renames CONTEXT.md to GLOSSARY.md and audits installed skills and recent usage, according to his upgrade prompt.
• The collection remains harness-agnostic; its Claude Code plugin updates automatically.

/retro reserves written coding standards for judgment calls, routing mechanical mistakes to executable checks. The accompanying v1.3.1 release notes also describe parallel implementation across worktrees and a YAML bug that hid six skills from installer discovery.

The shipped skill file marks /retro as user-invoked with disable-model-invocation: true. It reads the session sources the user specifies, potentially searching local logs; without a specified session, it defaults to the current one.

Pocock's earlier example targets repository navigation across ten sessions:

The skill loads writing-for-agents and examines seven categories:

• Navigation: difficulty finding relevant files or discovering dependencies between them.
• Automated checks: mistakes existing checks could catch, including checks that are unwired or silently broken.
• Coding standards: rules the reviewer missed, or existing rules needing clarification or removal.
• Global AGENTS.md: oversized steering instructions in repository or user-global files.
• Tool economy: expensive tool calls and token-inefficient CLIs or MCP tools.
• No-ops: steering instructions that do not change agent behavior.
• Information access: missing evidence, such as development-server logs or read-only access to third-party services.

Candidates come back ordered by severity. A repository with neither a pre-commit guardrail nor CI running its checks is itself a finding.

Mechanical rules versus judgment calls

• Mechanical violations: implement a deterministic check through the repository's linter, a pre-commit hook or CI.
• Judgment calls: keep them in CODING_STANDARDS.md, covering matters such as cross-file consistency that a deterministic check cannot replace.

A welcome bias toward checks that actually run.

The specification assigns coding standards to the reviewer, which receives a diff, rather than adding them to the implementer's exploration and debugging context. It treats CLAUDE.md and AGENTS.md primarily as sparse navigation pointers; CODING_STANDARDS.md is read during review.

Updated skills only look for GLOSSARY.md and GLOSSARY-MAP.md, according to the versioned migration notes. The filename change affects both readers and writers across the collection.

Pocock bundled repository migration and usage review in his upgrade prompt:

1. Rename domain docs: CONTEXT.md becomes GLOSSARY.md. The release notes additionally cover CONTEXT-MAP.md becoming GLOSSARY-MAP.md.
2. Check setup changes: inspect /setup-matt-pocock-skills for changes required by the release.
3. Compare installed skills: diff repository-local copies against upstream.
4. Audit usage: scan the last 25 sessions for skill invocations and assess how the release fits the existing workflow.

The original CONTEXT naming drew on domain-driven design's bounded context, Pocock explained in a reply.

The tagged README documents two distribution paths:

• Claude Code plugin: claude plugins install mattpocock-skills installs a managed, read-only bundle from the official marketplace. Pocock confirmed automatic updates.
• Codex and other agents: npx skills@latest add mattpocock/skills copies editable skill files into the project. Existing installations update with npx skills update, as he clarified in an update reply.

Here, skills names the installer package, rather than Pocock's particular skill collection. The README says installing both paths produces duplicate skills.

Jira is supported through /setup-matt-pocock-skills use JIRA as a ticket tracker, Pocock confirmed. The broader requirement is an issue tracker exposed through a CLI or MCP:

The v1.3.1 release graduates two other capabilities into the Engineering bucket and Claude Code plugin:

• implement-spec, user-invoked: reads tickets as a task graph and dispatches implementers in separate worktrees across the ready frontier. Work lands on one integration branch and ends with code-review; implementers use TDD and synchronize with the integration tip before reporting completion, enabling fast-forward landing.
• pr, model-invoked: builds a PR body around a compact visual summary, before/after evidence and a merge-risk statement covering reversibility and blast radius.

A draft PR for implement-spec opens only when the tracker closes work through PRs or the user requests one, and only after the first merge. Missing tracker configuration prompts a request to run setup rather than silently defaulting to gh.

Subagent names distinguish roles; Pocock said general-use agents work fine. The release also removes resolving-merge-conflicts, leaving merge and rebase conflicts to the agent without a dedicated skill.

Malformed YAML caused skills.sh to skip six skills during discovery. The patch notes trace the failure to an unquoted colon-space left after a prose edit and fix the description front matter in:

• to-spec
• code-review
• setup-matt-pocock-skills
• writing-fragments
• writing-shape
• wait-what

Cross-skill handoffs also changed:

• Model-invoked skills: instructions explicitly say Call the Skill tool with "name". Naming a slash command in prose had not reliably loaded it; multiple dependencies require separate calls.
• User-invoked skills: other skills cannot call them. Setup prerequisites now tell the human to run /setup-matt-pocock-skills, and an unreliable autonomous handoff from diagnosing-bugs to improve-codebase-architecture was removed.

Pocock explained why he does not maintain evals for the skills:

• He lacks a model provider's token budget to spend on them.
• He estimated that, if he wrote evals, maintaining them would consume 95% of his time instead of writing and shipping skills.
• He prefers listening closely to users over trying to emulate their usage ahead of release.

In the discussion under the /retro prompt, he also dismissed a reply for offering no concrete alternative.