Matt Pocock's Skills v1.3 Adds Parallel Spec Implementation, PR Templates, and a Retro for Agent Environments
Existing installations break on v1.3.0: skills will not see a glossary unless CONTEXT.md is renamed, and deleted skills must be cleaned up. For teams that run AI agents on real code, the new loop supplies what plain prompt-and-accept workflows miss: a human confirmation gate, a review step separate from implementation, and a mechanism that turns repeated agent failures into automated guardrails.
The Skills repo from Matt Pocock reached v1.3.0 on September 29, 2026 with three skills promoted to stable: implement-spec, pr, and retro. implement-spec treats tickets as a task graph instead of a step list, spawns implementation subagents in separate worktrees, merges their output into an integration branch, and runs a final code-review pass. pr prescribes a three-section pull request format that makes evidence concrete: a minimum visual summary, Before/After evidence, and a Merge Danger block that classifies reversibility. retro reviews the agent's working environment after a session, sorts problems into seven categories, and tells the repo to convert mechanical violations into lint rules or CI checks and judgment calls into CODING_STANDARDS.md.
The release also breaks older setups. CONTEXT.md and CONTEXT-MAP.md become GLOSSARY.md and GLOSSARY-MAP.md, and resolving-merge-conflicts was deleted because agents now handle ongoing merge conflicts themselves. Skills that called other skills now use a uniform 'Call the Skill tool' instruction instead of slash syntax, and six YAML descriptions were fixed after an unquoted colon made them invalid. After v1.3.1, the main branch carries 47 more commits and 17 pending changesets, including auto-created triage labels, sub-issues from to-tickets, and tdd seam explanations.
A hands-on walkthrough on a TypeScript checkout feature shows the intended loop: /setup-matt-pocock-skills configures the repo, /grill-with-docs asks all frontier questions in one round and writes agreed terms to GLOSSARY.md, /to-spec synthesizes without questioning, /to-tickets splits into vertical slices, /implement or /implement-spec does the work under tdd, code-review runs two separate subagents, pr formats the pull request, and retro closes the loop before /clear.
Process control stays with the human because user-invoked skills may call model-invoked skills but never one another; agents get reusable discipline without the power to chain entire workflows.
Retro's split between mechanical and judgmental violations keeps guardrails cheap to run: deterministic checks catch the fixed patterns, while CODING_STANDARDS.md stays small enough to be enforced by a review agent.
Dropping resolving-merge-conflicts rather than updating it matches the maintainer's expectation that agents can handle active merge conflicts natively in v1.3.
The 47 unreleased commits and 17 pending changesets after v1.3.1 show the stable release is already behind the code; useful behaviors such as sub-issues and auto-created triage labels exist only in that unreleased state.
A maintenance budget explains the no-contributions policy: every added skill costs long-term upkeep across Claude Code, Codex, and other harnesses, so composable existing skills take priority over new ones.