Matt Pocock's Skills (Skills for Real Engineers)

An MIT-licensed library of small, composable, model-agnostic agent skills - "Skills for Real Engineers... not vibe coding." It is the source of Kevin's 7-phase development pipeline, his forked grill-with-docs, the local engineering-flow skills kevin-engineering-flow, setup-kevin-engineering-flow, Agent Engineering Skills, Agent Engineering Skills, Agent Engineering Skills, Agent Engineering Skills, Agent Engineering Skills, Agent Engineering Skills, Agent Engineering Skills, wayfinder, Agent Engineering Skills, Agent Engineering Skills, Agent Engineering Skills, and Agent Engineering Skills, plus Agent Operations Skills and the long-form writing trio Writing and Content Skills, Writing and Content Skills, and Writing and Content Skills. Source: mattpocock/skills commit 84fdeffd12f2ee307994d1eb6feb48173b6e0502, 2026-08-11

Current Source Snapshot

The frozen 2026-08-11 source snapshot is main at 84fdeffd12f2ee307994d1eb6feb48173b6e0502, package 1.2.3, MIT, with 35 SKILL.md files. GitHub reported 213,557 stars, 18,429 forks, and an upstream push on 2026-08-07 at capture. The released source now includes /wayfinder, /research, /code-review, /to-spec, and /to-tickets; they are no longer speculative names. Kevin still uses curated local adaptations rather than blindly installing the complete repo. The correlated Meng To source also points to davidondrej/skills; its 49 procedures were reviewed beside all 35 Matt procedures because the post's claim is about the combination. Exactly one missing job was promoted: wayfinder. Source: frozen GitHub API/source captures, 2026-08-11; X/@MengTo, 2026-07-09

A Japanese OSS-trend roundup reported mattpocock/skills as its weekly number one AI repository, with a claimed gain of 10,651 stars. This is adoption evidence only: the delta and roundup links were not independently verified, and it does not change the curated-import rule. Source: X/@so_ainsight 2080144636170117582, reviewed 2026-08-05

External skill-source intake

The wider saved-skill cohort is a candidate feed, not a bulk-install manifest. Karpathy-derived guidance, error-discovery, GBrain, animations.dev, Browserbase, Swift/Xcode collections, problem-first, experiment-design, PR walkthroughs, measurement skills, agentic-engineering articles, GSkills, and registry results are all retained. For each exact source, resolve authorship and license, freeze the revision, inspect triggers/procedure/scripts/dependencies/network and write authority, compare overlap with the local owner, then run positive, negative, and adversarial trigger cases plus representative behavior and cost tests. Missing license or inaccessible source blocks copying/installing, not retention. Extend an existing skill when the job is already owned; create a new executable owner only for a distinct repeated procedure with a verifier. Source: 23 saved skill signals including X Articles 2042696610484781056, 2061440101411102721, 2063182047024500736, 2070393925760958464, 2076327982847647744, and 2076987913599094784; current Matt Pocock, Karpathy-derived, error-discovery, GBrain, and repository receipts, reviewed 2026-08-12

Philosophy

The skills are deliberately small, easy to adapt, and composable — the opposite of process-owning frameworks (GSD, BMAD, Spec-Kit) that "take away your control and make bugs in the process hard to resolve." Hack them, make them your own. They target four recurring agent failure modes: Source: repo README, 2026-06-15

Skill inventory

The current source contains 18 engineering, seven productivity, six in-progress, and four miscellaneous skills. All 35 are recorded in the correlated review artifact, alongside all 49 davidondrej/skills procedures. Existing local owners absorb almost all of them. wayfinder is the sole new executable owner because a multi-session decision map is materially different from a PRD, implementation tickets, a goal loop, or research. /research stays on the source-compilation stack, /code-review stays on Security and Review Skills, /to-spec aliases Agent Engineering Skills, and /to-tickets aliases Agent Engineering Skills. Source: exact-revision source audit, 2026-08-11

/setup-kevin-engineering-flow is the one-time per-repo scaffold: it configures the issue tracker (GitHub / Linear / local files), the triage label vocabulary, and where docs are saved - config the other engineering skills consume.

/loop-me is the newest high-signal addition: it interviews the user about recurring work loops and writes workflow specs for AI delegation. Kevin now has a local executable adaptation as Agent Operations Skills, routed through the no-one-off/workflow stack rather than treated as a generic planning skill. The reviewed bookmark had no visual artifact; its value was the resolved GitHub SKILL.md and self-thread that named /loop-me as the new in-progress skill. Source: X/@mattpocockuk, 2026-06-24; Source: GitHub SKILL.md

Current Command Surface

The 1.2.3 source resolves July's open questions. Kevin exposes the useful vocabulary while keeping one executable owner per job. wayfinder is manual-only because an ordinary planning request should not silently expand into a persistent multi-session map. Source: mattpocock/skills 84fdeffd12f2ee307994d1eb6feb48173b6e0502, 2026-08-11

This command surface is generated into Slash Command Index so root agent docs can stay small. The index is the listable surface; the skill files remain the executable source.

Local command policy: expose upstream vocabulary as /COMMAND handles, but add a local skill only when it owns a different repeatable workflow with a clear source, output, authority, stop, and verification contract. /to-spec and /to-tickets are aliases; /research, /writing-great-skills, /code-review, and /triage route to existing owners; /wayfinder is distinct and promoted. Source: User request, 2026-07-07; exact-revision audit, 2026-08-11

The local media for this cluster is mostly tweet/repo screenshot evidence rather than an executable artifact; the value is the sequence relationship across the rename posts, code-review vocabulary, and v1.1 planning notes. Source: X bookmark artifact audit, 2026-07-06

Fowler Review Baseline

PR #394 adds a curated Fowler smell baseline to the Standards axis of the in-progress review skill. The important design choice is that this is not a third review axis and not a hard lint rule. It is a Standards-axis heuristic set, with two binding constraints: documented repo standards override the baseline, and every smell is reported as a judgement call.

Kevin imported the vocabulary into Security and Review Skills rather than adding a duplicate /review command. The local review smell list is: Mysterious Name, Duplicated Code, Feature Envy, Data Clumps, Primitive Obsession, Repeated Switches, Shotgun Surgery, Divergent Change, Speculative Generality, Message Chains, Middle Man, and Refused Bequest. Source: mattpocock/skills PR #394 head 7a4c7561d4d8d866a9ae6e0073648ac94785ac8a, checked 2026-07-06

Writing-Great-Skills v1.1 Notes

The current writing-great-skills source frames skill quality around predictability, not size for its own sake. The vocabulary Kevin should keep using in Agent Operations Skills is: invocation, context load, cognitive load, router skill, information hierarchy, steps vs reference, progressive disclosure, context pointer, co-location, branch, leading word, completion criterion, legwork, post-completion steps, single source of truth, no-op, sediment, sprawl, duplication, relevance, and negation.

Local decision: /writing-great-skills continues to route to Agent Operations Skills and Agent Operations Skills. Do not install a second skill-authoring command unless upstream gains a materially new executable workflow. Source: mattpocock/skills skills/productivity/writing-great-skills, commit 16a2a5cd00b4416f673f4ff38c7971a04dd708e7, checked 2026-07-06

Bundle Pattern

The repo's setup pattern is the part Kevin should copy into his own skill distribution:

1. Choose one source mode. Claude's managed plugin is appropriate when Claude owns updates; skills.sh/vendored source is appropriate when Kevin needs editable canonical adaptations. Never install both because duplicate names make routing ambiguous.
2. Force a setup skill before first use: /setup-kevin-engineering-flow records issue tracker, triage labels, and domain-doc layout.
3. Split user-invoked from model-invoked skills. User-invoked skills orchestrate flows; model-invoked skills hold reusable discipline and can be reached automatically.
4. Use reference files for branch depth. SKILL.md stays small; files such as DEEPENING.md, CONTEXT-FORMAT.md, tests.md, and HTML-REPORT.md load only when that branch needs them.
5. Keep local adaptations explicit. Kevin's copies preserve upstream intent but route to local owners such as Agent Operations Skills, Security and Review Skills, Security and Review Skills, Security and Review Skills, and git safety rules.

This is the minimum viable shape for Kevin's distributed skills: source owner, setup skill, router skill, small executable skills, branch reference files, generated registry rows, family synthesis pages, resolver rows, and trigger evals.

The v1 split between user-invoked and model-invoked skills is the reason Kevin's local import keeps two boundaries: user-facing routers such as kevin-engineering-flow and setup flows carry Kevin-owned names, while upstream source names remain provenance. Model-invoked discipline skills can stay small and composable because they are not trying to be the command surface. Source: X/@mattpocockuk, 2026-06-17

Patterns worth adopting (the "great patterns")

1. Ubiquitous language as a living CONTEXT.md. A glossary (not a spec): opinionated, tight one-or-two-sentence definitions, project-specific terms only, with an _Avoid_ list of rejected synonyms. Multi-context repos use a CONTEXT-MAP.md. Pocock calls it "maybe the single coolest technique in this repo" — it makes the agent more concise, names variables/files consistently, and spends fewer thinking tokens. (Mirrors Kevin's The Brain-Agent Loop compiled-knowledge move.)
2. ADRs, offered sparingly. A 1-3 sentence record of a decision, created only when all three hold: hard to reverse, surprising without context, the result of a real trade-off. Otherwise skip it.
3. Tracer-bullet vertical slices. to-issues and tdd both insist on thin slices that cut through every layer end-to-end (schema→API→UI→tests), never horizontal "all tests then all code." Each slice is demoable; many thin > few thick. Slices tagged HITL (needs a human) vs AFK (agent can merge solo).
4. Deep modules + the deletion test. improve-codebase-architecture uses a strict vocabulary (module, interface, implementation, depth, seam, adapter, locality). The deletion test: imagine deleting a module — if complexity vanishes it was a pass-through; if it reappears across N callers it earned its keep. "One adapter = hypothetical seam; two adapters = real seam." (Ousterhout's A Philosophy of Software Design.)
5. Prototype = throwaway code that answers one question. Two branches: LOGIC (a runnable terminal app to push a state machine through hard cases) or UI (several radically different variations on one route, toggled by a URL param + floating bar). Throwaway from day one; keep only the answer (into an ADR/issue/NOTES.md).
6. PRD by synthesis, not interview. to-prd does not re-interview — it synthesizes the conversation into Problem / Solution / exhaustive User Stories / Implementation Decisions / Testing Decisions / Out of Scope, picking the highest test seam first. No file paths or code (they go stale) — except a prototype-derived snippet that encodes a decision precisely.
7. Review along two axes in parallel sub-agents. review checks Standards (does it follow this repo's documented standards?) and Spec (does it match the originating issue/PRD?) concurrently.
8. Skill authoring = progressive disclosure. write-a-skill: SKILL.md stays small; overflow goes to REFERENCE.md/EXAMPLES.md/scripts/. The description is the only thing the agent sees when choosing a skill, so it must encode what + when to trigger. (Same doctrine as Kevin's Agent Operations Skills.)
9. No-op pruning is contextual deletion testing. A skill line is only valuable if deleting it changes routing, first action, tool/file choice, output contract, stop condition, or verification. Vague virtue instructions and repeated advice are skill sediment unless rewritten into checkable behavior. Source: X/@mattpocockuk thread, 2026-06-24
10. Long-form writing splits explore from exploit. writing-fragments mines a raw pile without structure; writing-beats turns the pile into a reader journey one grounded beat at a time; writing-shape produces a coherent article paragraph by paragraph. This maps cleanly onto Kevin's content stack: gather raw material first, shape the article second, adapt for X/LinkedIn last. Source: X/@mattpocockuk, 2026-06-25; Source: GitHub in-progress skills, commit 5d78bd0903420f97c791f834201e550c765699f8

Local Adoption Matrix

Relevance to Kevin

Kevin uses this framework as his 7 phases. The local stack now covers most of the repo through natural owners instead of a pile of overlapping commands: grill-with-docs for alignment, domain-modeling for vocabulary/ADRs, prototype for evidence, to-prd for synthesis, to-issues for tracer-bullet execution, tdd and diagnosing-bugs for feedback loops, codebase-design for deep modules, triage for intake, and loop-me for recurring work delegation.

Late-July operating corrections

Deep replay of the later saved posts changed four local contracts that an inventory-only review had left stale:

1. Grilling uses dependency-frontier rounds. The 125-second demo and exact v1.2 skill remove one-question-at-a-time latency without mixing dependent decisions. Kevin's shape-first pass remains the map; the frontier is how the map is resolved. [Sources: X 2078077849785815465; grilling/SKILL.md at 84fdeffd]
2. Session transitions are a routing decision. Continue, discard/fresh, handoff, delegate, and compact preserve different things. The local agent-iteration-loop now selects between them at phase boundaries, while handoff explicitly owns recipient/context transfer rather than generic summarization. Source: saved visual from X 2079879414297330146
3. Comprehension debt needs a why digest. The weekly cross-project learning run now relates repository revisions to intent, decisions, behavior, failures, and proof before any optional audio presentation. A podcast file is not evidence that the changes were understood. Source: X/@mattpocockuk, 2026-07-14
4. Distribution is capability compilation. Matt's proposed installer and released plugin/copy split imply agent-specific adapters, dependency and invocation checks, one update owner, duplicate-command detection, and rollback proof. The environment bootstrap workflow owns this rather than a second global skill installation. [Sources: X 2075495703028142364; mattpocock/skills@84fdeffd]

The reviewed seven-phase visual still defines Grill, optional Research, optional Prototype, PRD, Issues, Implement, and Review. Kevin's local Grill phase now combines the prior breadth-first shape pass with the later upstream dependency-frontier rounds before high-fidelity branch work. [Sources: X/@mattpocockuk visual artifacts 2066451514672009337 and 2078077849785815465]

---

Timeline

• 2026-08-12 | Replayed the late-July Matt Pocock cohort beyond the repo inventory: inspected the full round-based grilling video, exact v1.2 grilling body, session-transition visual, comprehension-debt proposal, and installer/plugin surfaces. Corrected the executable planning skill, development workflow, session transition, cross-project digest, and environment bootstrap contracts. [Sources: X 2078077849785815465, 2079879414297330146, 2076257280501129336, 2075495703028142364; mattpocock/skills@84fdeffd]

• 2026-08-12 | Replayed the repository at exact revision 84fdeffd12f2ee307994d1eb6feb48173b6e0502 with a structured exploration receipt: all 35 skill manifests, 25 selected supporting documents, four executable/config support surfaces, release workflow, license, and local routing were inspected. The durable refinements are an explicit user-only/agent-or-user workflow invocation boundary, complete skill-manifest capture before repository evidence caps, and per-script authority review; the curated local routes remain preferable to a duplicate bulk install. Source: reviews/source-explorations/mattpocock-skills-2026-08-12.json; mattpocock/skills@84fdeffd

• 2026-08-12 | Added one external-skill intake gate across the recovered Swift, agentic engineering, design, measurement, experiment, error discovery, and registry sources. All signals remain available; only distinct, licensed, non-overlapping, behavior-tested jobs become executable owners. Source: quality/skills deep-adoption cohort

• 2026-08-11 | Replayed Meng To's paired engineering/orchestration source at exact revisions: all 84 procedures are accounted for; wayfinder is the only new executable owner; /to-tickets gained DAG/frontier and expand–migrate–contract rules; David Ondrej's guardrail and database-role examples became security/harness evidence rather than global installs. Source: X/@MengTo, 2026-07-09; mattpocock/skills 84fdeffd12f2ee307994d1eb6feb48173b6e0502; davidondrej/skills 081d4d52cddb1f24e9358add21d47a45d82dd3fd

• 2026-07-12 | Reabsorbed the July 7–11 skill cluster: /wayfinder clarification, installation-CLI/portable-source discussion, v1.1 expand-migrate-contract, and Meng To's companion library. The local decision is unchanged: retain one canonical skill library plus harness projections, and keep watch-only commands from becoming duplicate local owners. Source: X Bookmark Ingest 2026-07-12; X/@mattpocockuk, 2026-07-07–11

• 2026-07-07 | Tightened the local command policy for the v1.1 surface: /COMMAND vocabulary is listable through Slash Command Index, but aliases and watch-only routes do not become duplicate skills without a distinct owner contract. Source: User request, 2026-07-07

• 2026-07-06 | Absorbed the July v1.1 command-surface sequence and exposed it through Slash Command Index. Local decision: /to-spec aliases Agent Engineering Skills, /to-tickets aliases Agent Engineering Skills, /triage stays Agent Engineering Skills, /writing-great-skills stays Agent Operations Skills, /code-review routes through existing reviewers, and /wayfinder / /research remain watch-only until source semantics are clear. Current upstream HEAD: 16a2a5cd00b4416f673f4ff38c7971a04dd708e7; latest release remains v1.0.1. Source: X/@mattpocockuk, 2026-07-02/06; GitHub mattpocock/skills, 2026-07-06

• 2026-07-06 | Source-reviewed PR #394 head 7a4c7561d4d8d866a9ae6e0073648ac94785ac8a and folded the Fowler smell baseline into Security and Review Skills as Standards-axis heuristics. Reread writing-great-skills and kept Agent Operations Skills as the local route for predictability, progressive disclosure, no-op pruning, and leading-word vocabulary. Source: GitHub mattpocock/skills, 2026-07-06

• 2026-07-04 | Source-reviewed the .agents/skills symlink and proposed /research skill fragment. Local decision: the symlink note reinforces Cross-Agent Harness Portability, while /research stays watch-only until the owner boundary is clear because the described behavior is still close to ordinary "spawn research agent and save markdown" prompting. Source: X/@mattpocockuk, 2026-06-28

• 2026-06-30 | Deep-reviewed the /loop-me root bookmark and current upstream source. Current source snapshot: main HEAD 0877403d1e867fd9d574117e9b34ade404f36d2a, latest tag v1.0.1, 151,542 GitHub stars, MIT license. Local decision: keep Kevin's existing Agent Operations Skills adaptation, update provenance, and do not import a duplicate upstream router. Source: X/@mattpocockuk, 2026-06-24; Source: GitHub mattpocock/skills, 2026-06-30

• 2026-07-01 | Deep-reviewed the no-op pruning thread's resolved-link/reply evidence and refreshed source state to main HEAD 7a83a3a, 36 skills, 152,934 stars, and latest release v1.0.1. Local decision unchanged: no-op pruning belongs in Agent Operations Skills, not a new duplicate skill. Source: X/@mattpocockuk, 2026-06-24; Source: GitHub/skills CLI, 2026-07-01

• 2026-06-30 | Deep-reviewed the writing-trio bookmark and current upstream source. Local decision: keep the three installed Kevin-owned skills, expand their bodies with the missing upstream mechanics, and avoid a duplicate bundle import. Refreshed source state: main HEAD 801dca688564c529fa84f247f64472520d9ebe28, v1.0.1 latest release, Brin conservative caution. Source: X/@mattpocockuk, 2026-06-25; Source: GitHub mattpocock/skills, 2026-06-30

• 2026-06-30 | Deep-reviewed the /loop-me bookmark as part of the one-by-one artifact audit. Local decision: Agent Operations Skills is already installed and routed through no-one-off/workflow codification, so no duplicate skill is needed; refreshed upstream main to 801dca688564c529fa84f247f64472520d9ebe28. Source: X/@mattpocockuk, 2026-06-24; Source: GitHub mattpocock/skills, 2026-06-30

• 2026-06-30 | Deep-reviewed the seven-phase AI-powered development bookmark and local visual artifact; confirmed it updates The 7 Phases of AI-Powered Development and Agent Engineering Skills but does not require a duplicate skill. Refreshed upstream main to 999745ead1b37119380ad1f4de2bcdda5aa5bc84. Source: X/@mattpocockuk, 2026-06-15; GitHub mattpocock/skills, 2026-06-30

• 2026-06-30 | Deep-reviewed the v1 release bookmark: repaired the local screenshot artifact, verified latest upstream release v1.0.1 / main HEAD 43ea0884b07a3e67a5a07f025ce92aefa983177b, and recorded the user-invoked vs model-invoked split as the source-boundary rule behind Kevin-owned route names. Source: X/@mattpocockuk, 2026-06-17; Source: GitHub mattpocock/skills, 2026-06-30

• 2026-06-29 | Renamed upstream ask-matt and setup-matt-pocock-skills at the local route boundary to kevin-engineering-flow and setup-kevin-engineering-flow so imported source vocabulary does not leak into Kevin's live command surface. Source: User correction, 2026-06-29

• 2026-06-27 | Completed the local engineering import from current upstream: added kevin-engineering-flow, setup-kevin-engineering-flow, Agent Engineering Skills, Agent Engineering Skills, and Agent Engineering Skills; added branch reference files to existing local imports; documented the bundle/setup/user-vs-model-invoked pattern as the minimum shape for Kevin's distributed skills. Source: github.com/mattpocock/skills commit 5d78bd0903420f97c791f834201e550c765699f8; raw/registries/agent-capabilities/2026-06-27/mattpocock-skills-tree.json

• 2026-06-25 | Added Pocock's no-op pruning thread to the local skill-authoring route: no generic no-op phrase list, only contextual deletion testing for behavior change. Joe Fernandez's bookmarked reply reported 75-80% character reduction after applying the principle to workflow skills. Source: X/@mattpocockuk, 2026-06-24; Source: X/@joenandez, 2026-06-25

• 2026-06-25 | Promoted Pocock's updated in-progress writing trio into local executable skills: Writing and Content Skills for exploration, Writing and Content Skills for grounded article beats, and Writing and Content Skills for paragraph/block article shaping. Source: X/@mattpocockuk, 2026-06-25; Source: GitHub commit 5d78bd0903420f97c791f834201e550c765699f8

• 2026-06-25 | Re-read the upstream skill tree at commit 8370e760d0251a3738e006aeacec6d1cb31dd208, captured the 35-skill raw manifest, and promoted eight additional local skills: Agent Engineering Skills, Agent Engineering Skills, Agent Engineering Skills, Agent Engineering Skills, Agent Engineering Skills, Agent Engineering Skills, Agent Engineering Skills, and Agent Engineering Skills. Updated Agent Operations Skills with upstream loop vocabulary. Source: github.com/mattpocock/skills, 2026-06-25; raw/registries/agent-capabilities/2026-06-25/mattpocock-skills-tree.json

• 2026-06-25 | Rechecked the upstream repo through npx skills add mattpocock/skills --list --full-depth: 35 skills visible, ~145K GitHub stars, and mattpocock/skills@loop-me visible on skills.sh with ~1.9K installs. Added Kevin's local Agent Operations Skills adaptation. Source: CLI output, 2026-06-25; X/@mattpocockuk, 2026-06-24

• 2026-06-15 | Researched the repo (README, philosophy, full skill tree, and the grill-with-docs / prototype / to-prd / to-issues / tdd / improve-codebase-architecture / write-a-skill skills) after Kevin flagged it for extensive study and selective adoption. ~130K stars, MIT, updated same day. Source: github.com/mattpocock/skills, 2026-06-15