Skills v1.3: implement-spec, PR body and retro skills
- YT :: https://www.youtube.com/watch?v=BsJGo1wFTvQ
- Original title :: New Skills! v1.3 brings /pr, /implement-spec, and /retro
Matt Pocock walks through version 1.3 of his skills repo, which adds three skills. implement-spec has an orchestrator agent land a whole spec by delegating its tickets to sub-agents and producing a single PR. The PR skill is a PR-body template built around a summary, before/after evidence and merge danger. retro reviews real agent sessions and proposes improvements to the repo and environment. The release also renames context.md to glossary.md in the domain-modeling skill.
implement-spec: an agent babysits the ticket loop
Big work is split into a spec (the destination) and tickets (sessions small enough to stay out of the "dumb zone" and avoid auto-compact). Three ways to run the tickets: the manual loop (the human is the for loop, clearing context between tickets), a deterministic script loop (reliable and cheap, but hard to set up, so out of range for beginners), and the middle ground in this skill, where an agent orchestrates sub-agents. That became possible because sub-agents can now spawn sub-agents. It is worse than a deterministic loop but a good on-ramp to AFK workflows.
The skill treats tickets as a task graph with blocking relationships, so there is always a frontier of ready tickets that can run in parallel. Its nine steps: read spec and tickets, explore in a sub-agent, create an integration branch, implement each ticket in a sub-agent (TDD, worktrees), merge into the integration branch, repeat until done, run code review on the integration branch, clean up, and leave one PR ready for review. It is a basic version meant to be remixed (stacked PRs, etc.).
PR skill: make human review easy
PRs are still the bottleneck for getting work to main. Inspired by Dex Horthy's "Show Me" skill (concise diagrams and pseudo-code), the skill is just a PR-body template with four parts. A summary in show-me style. Evidence, before and after: asking for it makes the agent run an extra test or take a screenshot instead of claiming "that probably works". Merge danger: a one-way door (deletes data, costly to roll back) versus a two-way door (easy revert). Blast radius: the potential ramifications. A small blast radius on a two-way door needs little review, so this goes near the top. Matt's emphasis on verification follows his recent interview with Lauren (poteto).
It is one of the most consistently model-invoked skills he has seen (on Opus 5.5), so users just get better PR bodies; people with their own PR skill can borrow from it.
context.md becomes glossary.md
The domain-modeling skill (grill with docs) now writes glossary.md instead of context.md. The old name came from DDD's bounded context, but it was vague, did not make the agent pull the file in at the right moment, and confused users. The content had shrunk to just a glossary anyway. Only the file name changes, but many skills depend on it, so users must update.
retro: a retrospective on real agent sessions
Matt used to ask users to maintain their repos for agents (agents.md, skills, token efficiency, lint rules, navigation) and realised he did this himself, so he bundled it into retro. Run it on the current session, the previous one or several; because it reads actual sessions it sees inefficiencies the agent hides, since agents persist and rarely complain. Example findings from his repos: an irreversible release run without asking (hence no 1.3.0, only 1.3.1), a check script that nothing runs (add CI), context loss across compactions (move repeated instructions into an animatic skill), and a custom CLI wasting tokens.
Retro is meant to be human-in-the-loop: do not automate it, because the agent will chase false positives and drift the repo somewhere bad. Run it on a sample of sessions, especially ones that went wrong. Categories it checks: codebase navigability, automated checks, coding standards for the automated reviewer, health of the global agents.md, tool economy, no-op steering instructions, and missing information. Matt reports a large gain in token efficiency and work quality from running it regularly. He closes by plugging his AI Coding Crash Course on aihero.dev, updated with an implement-spec lesson.