<- all tokdocs
2026.10.13@mattpocockuk

Matt Pocock says /retro will add rules to your steering files; the skill's own SKILL.md says it only proposes and waits

AI
Watch on TikTok

View on TikTok ->

The artifact is a 62.90-second vertical clip at 1080x1920 in an MP4 container, video encoded as HEVC (yt-dlp format bytevc1_1080p_448368-1) at 60 fps and 375,712 bps, audio as AAC at 44,100 Hz stereo and 64,121 bps, overall bitrate 448,368 bps, file size 3,525,297 bytes. The post timestamp is 1791901458, which is 2026-10-13 14:24:18 UTC. The handle is mattpocockuk, the channel nickname is "Matt Pocock", and the audio is credited as "original sound - Matt Pocock", so there is no licensed music. The description is 29 characters and identical to the title: "/retro is a banger of a skill". At capture on 2026-10-19 05:16:31 UTC the metadata recorded 6,243 views, 341 likes, 6 comments, 28 reposts, and 172 saves. I read 13 of the 31 extracted frames (001, 003, 006, 008, 009, 012, 015, 018, 021, 024, 027, 029, 031) and all 222 transcript words across 19 Whisper segments. Frames 001 to 006 (0 to 10 s) are a full-frame talking head: a bald man in round tortoiseshell glasses with a short reddish beard, in a blue-and-grey plaid flannel over a white tee, seated in a dark high-back chair in front of wooden bookcases, with a black dynamic microphone on a boom arm at chest height. Yellow monospace captions read "What do you do when your" (001), "Or do you think, how do" (003), and "I have to achieve / that? Well, you" (006). Frame 001 also carries a lower-third card reading "AIhero.dev" and "Go from Zero To AI Hero". From frame 008 (14 s) to frame 031 (60 s) the top half is a screen recording of a dark aihero.dev page: the "AIHero" wordmark, a metadata line "08 / 27 | AI Skills for Real Engineers · 7 min read · Updated Oct 5", a purple app icon beside the heading "The /retro Skill", the subtitle "Look back on a session and suggest repo improvements.", an author row "Matt Pocock" with a "Subscribe" button, the label "Install this skill", the code block "$ npx skills@latest add mattpocock/skills --skill=retro", and the line "Then type /retro in your coding agent." The bottom half is the same speaker under warmer light with a bookshelf behind him. Captions in that stretch include "for v1.3. And what it does is" (009), "and turn them into either." (015), "to make sure that code review" (021), and "in." (031). No terminal, editor, agent output, or run of /retro appears in any frame I read. The transcript is one spoken argument: when an agent errs, ask what mechanism makes that error harder next time, and the answer is "the retro skill", which "conducts a retrospective on your previous sessions" and turns mistakes into "automated checks" or "steering files like codingstandards.md" so that "code review agents catch those mistakes before they go to production". The last segment ends at 62.78 s against 62.90 s of audio, so no transcript text sits past the recording.

The v1.3 date holds, and the skill is older than "fresh" suggests

The transcript at 13.84 s says "This is a skill that's fresh in my skills repo for V1.3." The GitHub API shows release v1.3.0 published 2026-10-04T12:47:22Z and v1.3.1 at 12:48:18Z the same day, both carrying the changeset "Graduate retro into the Engineering bucket, so it ships in the Claude Code plugin, gets a docs page, and is routed by ask-matt as the last step of the main flow, after code-review." The video went up 2026-10-13, nine days after the release. The on-screen header "Updated Oct 5" in frames 008 through 031 matches the updated stamp of 2026-10-05T10:57:45Z on the aihero.dev v1.3 changelog post. The version claim is correct.

"Fresh" needs one qualification. The retro skill was first committed on 2026-08-24 as skills/in-progress/retro/SKILL.md (commit 8fa18865, "feat: add initial retro skill and OpenAI agent configuration"). PR #1120, which moved it to skills/engineering/retro/, was opened 2026-09-24 and merged 2026-09-29. What v1.3 changed was the skill's status: it left the in-progress bucket, entered the Claude Code plugin manifest, got a docs page, and was wired into the ask-matt router as the final step of the chain grill-with-docs → to-spec → to-tickets → implement → code-review → retro. PR #1120's evidence block records the plugin going from 25 skills to 27, which is consistent with the "08 / 27" counter on the page in the frames. The same PR removed resolving-merge-conflicts and renamed CONTEXT.md to GLOSSARY.md across eleven skills, neither of which the video mentions.

I tried to fetch Matt Pocock's X post announcing v1.3 (status 2107062474914550098) and received HTTP 402, so I am not citing it. The GitHub release and the aihero.dev changelog cover the same ground.

"Previous sessions" reverses the default; retro reads the current session unless you point it elsewhere

At 18.08 s the video says retro "conducts a retrospective on your previous sessions", and at 22.16 s, "if a mistake happens in your previous sessions, you run retro and it will do a retrospective." SKILL.md step 2 reads: "Read the primary sources for the session the user specifies. This may mean searching through session logs on this machine. If the user doesn't specify a session, default to the current one." The repo's docs page (docs/engineering/retro.md, published at aihero.dev/skills-retro) says the same: "It reads the session's own record (the current one by default, or one you point it at in the session logs)."

Two things follow. First, the plural misdescribes a single run. The docs' FAQ says "It sees one session," and gives that as the reason it "cannot tell you that a rule now fires too often or that the bug behind it is gone." Second, the current session is the recommended input. The docs say "By default it reviews the current session. That is the best case, because the struggles are still in the context window." Only when the session "has already moved out of the smart zone" do they tell you to clear it and "point a fresh /retro at the previous session in the session logs."

The video's framing works as one use case, since you can aim the skill at a prior session's log, but it inverts the default. A viewer who installs the skill and types /retro at the end of a working session gets a retrospective on that session, which is the designed path.

"It will add them" overstates what the skill does; it proposes and waits

The core claim runs from 27.04 s to 46.00 s: retro "will try to find the mistakes, try to find the inefficiencies in your process and turn them into either automated checks that deterministically check that your agent is doing the right thing. Or it will add them into steering files like codingstandards.md". The verb "add" does not match the skill. SKILL.md step 4 is the last step, and it reads "Present these candidates to the user, in order of severity." The docs page says: "It also only proposes. Nothing changes until you pick a candidate." The FAQ is more explicit: "It waits. retro only proposes. Nothing changes until you pick a candidate, so it edits nothing by hand and installs no hook automatically. This is deliberate. One user asked for exactly this after being 'burned by auto-hooks that blocked good changes.'" The docs add that "There is no dry-run mode. A proposed check is code like any other, so try it against the repo before you let it block merges."

The frontmatter enforces the same boundary on the input side: disable-model-invocation: true. Per the Claude Code skills reference, that field means Claude cannot invoke the skill on its own and the user must run /retro. The release notes label it "(user-invoked)" for this reason, and the docs say "the agent won't reach for it on its own."

The distinction matters for the video's pitch. The opening question at 4.20 s is "how do I make that mistake harder for the agent to make next time?", and the answer the clip gives is a tool that does it for you. The tool as shipped produces a ranked list. The human picks a candidate, and only then does the agent write the lint rule, hook, CI job, or standards line.

The two destinations are real, and the skill's own rule is to prefer the check over the sentence

The video's split between "automated checks that deterministically check" and "steering files like codingstandards.md" tracks the skill closely. SKILL.md's coding-standards bullet reads: "Classify the violation first: a mechanical one (a fixed syntactic pattern, a banned API, an import shape, a file-location rule) gets a deterministic check, full stop: a custom rule in the repo's own linter, a new pre-commit hook, or a new CI job, whichever the repo's language and existing guardrail make cheapest. Default to building the check over writing the rule. Reserve CODING_STANDARDS.md for genuine judgement calls". The docs give the reason: "a check can fail and a sentence in a standards file can't." The automated-checks bullet adds a case the video skips: "A repo with no guardrail (no pre-commit hook and no CI job running its lint/typecheck/test command) is itself a finding."

The code-review handoff at 38.38 s to 46.00 s also checks out. The code-review SKILL.md's step 3 says "When CODING_STANDARDS.md or CONTRIBUTING.md exists, it must be on the list" of standards sources, and the retro SKILL.md's reference section says "CODING_STANDARDS.md: this file is read during review, not implementation." Per the docs FAQ, no skill ships the file; retro proposes creating it the first time a judgement-call rule appears, and code-review reads it after you accept.

One wording issue remains. The video calls CODING_STANDARDS.md a steering file. In the skill's own taxonomy, steering files are CLAUDE.md and AGENTS.md, which "are pushed to the context window of any agent working in this repo" and "should be used incredibly sparingly, usually only for navigation pointers to other files." CODING_STANDARDS.md is a review input. The docs say retro "pushes back on" adding lines to CLAUDE.md "more than anything else," and one of its seven categories exists to move steering out of AGENTS.md into standards or checks. The transcript's "codingstandards.md" is Whisper flattening an uppercase, underscored filename; the page on screen shows no filename at all.

What "skill" means here, checked against Anthropic's live docs and the installer on screen

The clip never explains the mechanism, so here it is from primary sources, checked 2026-10-19. Claude Code's skills reference at code.claude.com defines a skill as a SKILL.md file with YAML frontmatter (name, description, plus optional fields such as disable-model-invocation, allowed-tools, and context: fork) followed by Markdown instructions. The directory name becomes the slash command, so skills/engineering/retro/SKILL.md yields /retro. Skill content loads only when used, which the docs give as the reason to prefer skills over CLAUDE.md for long reference material. The page also states that custom commands "have been merged into skills": a .claude/commands/deploy.md and a .claude/skills/deploy/SKILL.md both produce /deploy. Skills live in ~/.claude/skills/, .claude/skills/, or <plugin>/skills/. The retro SKILL.md's own reference section applies the same rule: "use skills for docs (since their description goes into the agent's context window), or for user-invoked commands."

The install line on screen, npx skills@latest add mattpocock/skills --skill=retro, calls the Vercel Labs skills CLI ("The open agent skills tool - npx skills", MIT), which supports a --skill flag and installs into Claude Code, Cursor, Codex, GitHub Copilot, Windsurf, and others. The repo README lists two more paths: claude plugin install mattpocock-skills@claude-plugins-official and copilot plugin install mattpocock-skills@mattpocock. The retro folder also carries agents/openai.yaml for Codex.

The format itself is Anthropic's. The "Introducing Agent Skills" post on claude.com is dated October 16, 2025, and agentskills.io says the format "was originally developed by Anthropic, released as an open standard," listing Claude Code, Claude, Cursor, Gemini CLI, GitHub Copilot, ChatGPT and Codex, and dozens of other clients.

The repo documents three weaknesses the clip leaves out

"Retro has been absolutely invaluable for me" (46.32 s) and "it's really forced me to consider that maybe my code base is not as good as I thought it was" (48.68 s) are testimony, and no run of the skill appears on screen. The repo's docs are more candid than the video. Under "Won't it just invent generic advice to fill its categories?" the docs call this "the strongest criticism of it" and quote a user: "once the job is finished, the AI tends to forget the struggles from the middle of the session and invents generic advice to satisfy the retro categories." The docs' mitigation is a reading rule for the user: "Discard any candidate you can't trace to a specific moment."

Second, under "Won't this pile up lint rules forever?", the docs answer "Partly, and this is its weakest spot," because retro "does not audit the lint rules, hooks, or CI jobs it proposed last month." Third, the docs say retro "is not a memory system. It does not store what happened," which bounds the video's "make that mistake harder next time" to the checks and rules you accept.

The "put up with without moaning" line at 53.14 s has a literal counterpart in the SKILL.md's "Implementation vs Review" section: the implementation agent "has the most context pressure," so standards are enforced by the reviewer, which "receives a diff, so no exploration needed." That design reasoning is the most substantive idea in the skill, and the clip skips it.

On the creator: the ai.engineer speaker page lists Matt Pocock as creator of Total TypeScript and AI Hero, formerly on the XState core team at Stately and a contractor on Vercel's Turborepo and Turbopack documentation. The Pragmatic Engineer's September 17, 2026 interview reports Total TypeScript "generated more than $2.5 million in total sales." The skills repo (created 2026-02-03, MIT, homepage aihero.dev/skills) showed 297,200 stars and 24,952 forks when I queried the GitHub API on 2026-10-19.

Key Takeaways

  • Verified: The artifact facts are confirmed by ffprobe and the post metadata. 62.90 seconds, 1080x1920, HEVC at 60 fps, AAC audio, 3,525,297 bytes, posted 2026-10-13 14:24:18 UTC, 6,243 views and 341 likes at capture on 2026-10-19.
  • Verified: mattpocock/skills v1.3.0 was published 2026-10-04T12:47:22Z with retro graduated into the Engineering bucket and the Claude Code plugin, nine days before the video.
  • Verified: Retro's two output destinations are a deterministic check (linter rule, pre-commit hook, CI job) for mechanical violations and CODING_STANDARDS.md for judgement calls, and the code-review skill reads CODING_STANDARDS.md as a standards source.
  • Correction: The video says retro "will add them into steering files like codingstandards.md". SKILL.md's final step is "Present these candidates to the user, in order of severity," and the docs say "It waits. retro only proposes. Nothing changes until you pick a candidate."
  • Partial correction: The video says retro runs "on your previous sessions". SKILL.md defaults to the current session ("If the user doesn't specify a session, default to the current one"), reads one session per run, and the docs recommend the current session as "the best case."
  • Clarification: "Fresh" refers to graduation, not creation. The skill was first committed as in-progress on 2026-08-24, and PR #1120 moving it to Engineering merged 2026-09-29.
  • Clarification: CODING_STANDARDS.md is a review input in the skill's taxonomy. Steering files are CLAUDE.md and AGENTS.md, and the docs say retro pushes back on adding lines there "more than anything else."
  • Unverified: "Retro has been absolutely invaluable for me" and the claim that it exposed problems in his codebase are personal testimony. No run of the skill, no finding, and no before-and-after appears in the clip.
  • Context: The SKILL.md format was announced by Anthropic on October 16, 2025 and is now an open standard at agentskills.io. Claude Code's docs state that custom slash commands have been merged into skills, and disable-model-invocation: true in retro's frontmatter is what makes it user-only.

Resources

Published October 13, 2026. Writeup generated from a favorited TikTok.