SDLC Agent-Skill Wiki

Matt Pocock — Skills for Real Engineers

13 min read

Workflow — the primary skill per SDLC stage this framework runs, top to bottom (folded and off-stage steps omitted).

flowchart TB
    A1["mp-wayfinder<br/><i>Align</i>"] --> S["mp-to-spec<br/><i>Specify</i>"]
    A2["mp-grill-with-docs<br/><i>Align</i>"] --> S
    S --> P["mp-to-tickets<br/><i>Plan</i>"]
    P --> I1["mp-implement<br/><i>Implement · per ticket</i>"] --> R["mp-code-review<br/><i>Review</i>"]
    P --> I2["mp-implement-spec<br/><i>Implement · parallel</i>"] --> R
    R --> PR["mp-pr<br/><i>Release · PR body</i>"] --> L["mp-retro<br/><i>Learn</i>"]

A curated set of small, composable agent skills by Matt Pocock “for real engineering work — not ‘vibe coding’”, sourced from his personal .claude directory and designed to be adapted (“Hack around with them. Make them your own.”). Works with any model and any coding agent. Explicitly positioned against owning-the-process frameworks: “Approaches like GSD, BMAD, and Spec-Kit try to help by owning the process. But while doing so, they take away your control and make bugs in the process hard to resolve.” These skills stay small and composable instead.

The current release is v1.3.1 (2026-10-04): 37 skills in engineering/, productivity/, misc/, and a beta-channel in-progress/ folder.

  • Install: two paths, “two philosophies”. The Claude Code plugin (claude plugins install mattpocock-skills, in the official marketplace) is a managed, read-only bundle that self-updates: “you subscribe rather than fork”. npx skills@latest add mattpocock/skills (skills.sh) copies editable files for Codex and other agents. “Pick one: installing both leaves you with every skill twice.”
  • Configure: /setup-matt-pocock-skills (issue tracker — GitHub/Linear/local files, triage labels, doc paths).

Dual-harness. Every skill ships an agents/openai.yaml beside its SKILL.md with Codex UI metadata, and user-invoked skills set policy.allow_implicit_invocation: false (the Codex analog of disable-model-invocation: true). AGENTS.md is a symlink to CLAUDE.md. A native Codex plugin is deferred (ADR 0002). runs_on: lists only [[claude-code]], since the wiki has no Codex harness page (support-scope rule).

Unlike gsd — a single end-to-end workflow engine — this is a loose toolkit of independent skills, each user-invoked (typed as a slash command; orchestrates) or model-invoked (typed or auto-reached-for by the agent; holds reusable discipline). A user-invoked skill may invoke model-invoked ones, never another user-invoked one. Cross-skill calls are written as an explicit “call the Skill tool with …” instruction, which only works for model-invoked targets.

The main flow (via mp-ask-matt)

The router mp-ask-matt frames the skills as flows. The main flow, idea → ship:

mp-grill-with-docs → (prototype detour bridged by mp-handoff when a question needs a runnable answer: mp-prototype) → mp-to-spec → mp-to-tickets → mp-implement (per ticket, clearing context between each) or mp-implement-spec (whole spec, parallel) → mp-code-review → mp-pr (when it goes up as a PR) → mp-retro (closes the loop)

Three on-ramps merge onto it: mp-triage (raw incoming issues), mp-diagnosing-bugs (the “something’s broken” entry, followed by /retro once fixed), and mp-wayfinder (an effort too big for one session, merging back at /to-spec). A codebase-health lane (mp-improve-codebase-architecture) generates ideas for the front of the flow. A vocabulary layer runs underneath (mp-domain-modeling, mp-codebase-design). At every phase boundary, the router’s tree orders the options continue → /clear → /handoff → subagent → /compact.

The flow is a loop, not a line. mp-retro feeds findings back into the environment (checks, coding standards, navigation pointers, tooling) that the next build starts from. That puts the toolkit in stage-learn, in the environment-compounding flavor (see topic-compounding-loop).

Four failure modes addressed

Capabilities

Skills — align / elicitation & intake

  • mp-grill-me — comprehensive interview on plans/designs until decisions resolve.
  • mp-grilling — the reusable interview loop underneath the grill skills: asks the whole question frontier round by round, sends facts to background sub-agents, puts decisions to the human, and waits for confirmation before acting.
  • mp-grill-with-docs — grilling that also builds a domain model + ADRs.
  • mp-triage — move issues/PRs through a triage state machine → agent-ready briefs (intake).
  • mp-wayfinder — map an oversized effort as a shared map of decision tickets on the tracker, resolving decisions until the destination is clear; research tickets burned down by parallel subagents (situational on-ramp; upstream of the spec).
  • mp-to-questionnaire — turn a decision you can’t answer into a questionnaire for the one person who can; grills the send, not the subject.

Skills — specify

  • mp-to-spec — synthesize the conversation into a spec on the issue tracker; sketches testing seams; no interview.

Skills — plan (decompose, design, research)

  • mp-to-tickets — break a plan/spec into tracer-bullet tickets with blocking edges; handles wide refactors via expand–contract.
  • mp-research — background agent investigating primary sources → cited Markdown (artifact-research-md).
  • mp-domain-modeling — build/sharpen the project domain model + ubiquitous language.
  • mp-codebase-design — shared vocabulary for designing deep modules at clean seams.

Skills — implement (build, fix)

  • mp-implement — build the work from a spec/tickets, one ticket per context; drives mp-tdd at seams, closes with mp-code-review, commits.
  • mp-implement-spec — implement a whole spec in one run: tickets as a task graph, implementer subagents in worktrees across the ready frontier, one integration branch, one closing code-review.
  • mp-tdd — test-first red → green loop at pre-agreed seams.
  • mp-prototype — a throwaway prototype for one design question: a single shareable HTML file (logic) or toggleable UI variations, kept on a prototype/<name> branch as a primary source.
  • mp-diagnosing-bugs — reproduce → minimize → hypothesize → instrument → fix → regression-test, with secrets redacted throughout.

Skills — review (quality gate)

  • mp-code-review — two-axis diff review, Standards (+ Fowler smell baseline) ∥ Spec, parallel sub-agents.
  • mp-improve-codebase-architecture — scan recently changed paths for deepening opportunities; visual HTML report, then grill the one you pick.

Skills — release (PR write-up)

  • mp-pr — model-invoked shape for a PR body: smallest-visual Summary, before/after Evidence, Merge Danger (one-/two-way door + blast radius). Writes the body only; it does not push or merge.

Skills — learn (loop close)

  • mp-retro — retrospective on a coding session that suggests changes to the agent’s environment: navigation pointers, automated checks, coding standards, steering files, tool economy, information access; mechanical findings become deterministic checks.

Skills — cross-cutting

  • mp-handoff — compact a conversation into a handoff document when context must travel (new harness, new directory, a colleague, a mid-phase fork).
  • mp-ask-matt — router over the skills; describes the flows above and the phase-boundary tree.
  • mp-wizard — model-invoked generator of an interactive bash wizard for steps only a human can perform (provisioning, credentials/CI secrets, dashboards, cutovers).
  • mp-wait-what — three-line corrective: re-pitch the last message in Simplified Technical English using the GLOSSARY.md vocabulary.
  • mp-writing-for-agents — model-invoked reference for writing any agent-facing document (skills, AGENTS.md/CLAUDE.md): context pointers, the two loads, information hierarchy, leading words, pruning.

Catalogued (non-lifecycle infra / tooling — not paged)

  • setup-matt-pocock-skills — configure issue tracker, triage labels (asked only if triage is installed), and doc layout (single-context by default) once per repo; local tickets live one file per ticket under .scratch/<feature>/issues/.
  • git-guardrails-claude-code — block dangerous git commands via pre-execution hooks (an pattern-edit-guardrails relative).
  • setup-pre-commit — Husky hooks: lint-staged, Prettier, type-check, tests.
  • migrate-to-shoehorn — migrate as test assertions to @total-typescript/shoehorn.
  • scaffold-exercises — generate exercise directory structures.
  • teach — multi-session teaching workspace (productivity).

In-progress (the beta channel, not paged)

skills/in-progress/ is “a beta channel, published on purpose, installable one skill at a time through skills.sh”: claude-handoff, loop-me, setup-ts-deep-modules (wire dependency-cruiser so each TS package is a deep module reachable only via its entry points), writing-beats, writing-fragments, writing-shape.

Artifacts produced

Patterns applied

See Also

Version history

Newest first. Each entry records what changed in that release relative to the one before; the sections above describe only the current release.

v1.3.1 (2026-10-04; ingested 2026-10-11)

v1.3.0 (2026-10-04)

  • Main flow gained a tail: mp-pr (PR body) and mp-retro (environment retrospective) graduated from in-progress/. Before this the flow ended at mp-code-review, and the toolkit had no stage-release or stage-learn capability.
  • mp-implement-spec graduated as the parallel, whole-spec alternative to per-ticket mp-implement.
  • Removed: resolving-merge-conflicts (mp-resolving-merge-conflicts, kept as a historical page), “no longer needed, and nothing replaces it”. It had been an implement-stage skill applying pattern-trunk-based-development and producing atomic commits.
  • Renamed: the domain doc CONTEXT.md / CONTEXT-MAP.md → GLOSSARY.md / GLOSSARY-MAP.md across all skills (existing files must be git mv’d). This ended the filename clash with GSD’s artifact-phase-context.
  • Cross-skill invocation standardized on “call the Skill tool with …”, and preconditions that tried to call the user-invoked setup-matt-pocock-skills now tell the human to run it. diagnosing-bugs’ Phase 6 hand-off was removed for the same reason.
  • Em-dashes removed from all prose.

v1.2.0 – v1.2.3 (2026-08-05 – 2026-08-06)

  • Distribution: shipped as a native Claude Code plugin in the official marketplace; added Codex metadata (agents/openai.yaml) beside every skill, with AGENTS.md symlinked to CLAUDE.md.
  • Promoted: mp-to-questionnaire, mp-wizard (made model-invoked; v1.2.3 dropped its time estimate), and mp-wait-what (new).
  • Renamed: writing-great-skills → mp-writing-for-agents (breaking, no alias), widened from skills to all agent-facing documents; made model-invoked again for Codex in v1.2.2.
  • mp-grilling moved from one-question-at-a-time to round-by-round over a question frontier, and was reworded for any plan, decision, or idea.
  • mp-prototype reshaped: the logic branch became a single shareable HTML file instead of a terminal app, and prototypes are kept on a prototype/<name> branch rather than deleted.
  • mp-wayfinder’s unit was renamed from “investigation ticket” to decision ticket, and research tickets are burned down by parallel subagents.
  • mp-ask-matt gained the five-option phase-boundary tree, narrowed mp-handoff to when context must travel, and raised the smart-zone figure from ~120k to ~150k tokens.
  • mp-improve-codebase-architecture gained a YAGNI scoping filter (recent commit paths).
  • mp-diagnosing-bugs gained secret redaction (v1.2.3); sub-agent dispatch in mp-code-review, mp-codebase-design, and improve-codebase-architecture became harness-neutral.
  • setup-matt-pocock-skills got friendlier defaults, and local tickets became one file per ticket.
  • The to-prd → to-spec rename was finished in all shipped text.
  • Deleted: the four v1.1 deprecations (ubiquitous-language → mp-domain-modeling; design-an-interface → mp-codebase-design, which ships “design it twice” as DESIGN-IT-TWICE.md; qa → mp-triage + mp-to-tickets; request-refactor-plan → mp-to-spec + mp-improve-codebase-architecture) and the personal/ bucket (edit-article, obsidian-vault).

v1.1.0 (2026-07-08; ingested 2026-07-09)

Initial ingest (2026-06-28)

  • First documented: 12 core SDLC skills paged (including to-prd and to-issues), 9 meta/niche skills catalogued. Building was covered only by the test-first mp-tdd (then red-green-refactor) and the throwaway mp-prototype; there was no execute skill that walked a plan, and no diff review beyond the architecture scan in mp-improve-codebase-architecture.

Graph View

Backlinks