The agent skill pack with a workflow spine.
English · 简体中文 · vendored under MIT & Apache-2.0 — see ATTRIBUTION
Advisory skills don't ship design systems — contracts do. oh-my-trellis vendors
mattpocock/skills,
Everything Claude Code and
pbakaus/impeccable into one repo, then
binds their loose nouns — "the issue tracker", "PRODUCT.md", "DESIGN.md" — to
Trellis's .trellis/ spine so the workflow is enforced, not suggested.
3 vendored upstreams · 1,741 catalog assets · 43 default skills · 22 platforms · 1 contract layer — zero patches to vendored files.
| You want… | Run |
|---|---|
| The Trellis contract specs only | trellis init --registry gh:ScoFan-official/oh-my-trellis/marketplace --template agent-workflow --append |
The recommended skill set (all 37 mp skills + impeccable + the pack's own mp-trellis-bridge / oh-my-update / trellis-domains / trellis-run / dispatch-roster / hygiene-sweep / rules-audit) |
npx skills add ScoFan-official/oh-my-trellis --agent <platform> --copy |
| Anything else from the 1,741-asset catalog | python scripts/install.py --platform <platform> --target <repo> |
Repo is public — all three channels work without auth. Mix and match freely.
Hand this to Devin / Codex / ZCode / Claude / Cursor and it will set everything up itself:
Set up Trellis and oh-my-trellis in this repo: ① install our Trellis CLI and init —
npm i -g https://github.com/ScoFan-official/oh-my-trellis/releases/download/cli-v0.6.18-ohmy.5/oh-my-trellis-0.6.18-ohmy.5.tgz(if@mindfoldhq/trellisis already installed globally,npm rm -g @mindfoldhq/trellisfirst — our package owns thetrellisbin), thentrellis init --devin— registry and workflow defaults are baked into the fork; ② runtrellis init --template agent-workflow --appendto install the spec contracts; ③ install the skills for YOUR platform — prefernpx skills add ScoFan-official/oh-my-trellis --agent <your-platform> --copy; if your platform isn't covered, rungit clone --depth 1 https://github.com/ScoFan-official/oh-my-trellis /tmp/mtp && python /tmp/mtp/scripts/install.py --platform <your-platform> --target . --only-coreinstead; ④ invoke themp-trellis-bridgeskill so it finishes seedingAGENTS.md, then read.trellis/spec/agents/issue-tracker.mdand report back which tracker verbs you'll use for specs, tickets, triage and implementation. To update later, invoke theoh-my-updateskill.
For non-Trellis repos, skip to skills only:
Install the mattpocock skill set here:
npx skills add ScoFan-official/oh-my-trellis --agent <your-platform> --copy, then browsecatalog/INDEX.mdin https://github.com/ScoFan-official/oh-my-trellis and install extra components for this platform withscripts/install.pyif needed.
Skills that say "publish to the issue tracker" get pointed at .trellis/tasks/.
Skills that say "update PRODUCT.md / DESIGN.md" get pointed at .trellis/spec/.
Delivery is index-driven: Trellis's session-start hook injects the spec
packages' index.md files, and the agent reads each contract the index lists —
agents/index.md first. The frontend gate is mechanical: the CLI refuses to
archive a frontend task whose implement.md carries no ## Design review
section.
trellis init --registry gh:ScoFan-official/oh-my-trellis/marketplace --template agent-workflow --appendInstalls into .trellis/spec/:
.trellis/spec/
├── agents/
│ ├── index.md
│ ├── issue-tracker.md # THE contract — read before publishing to the tracker
│ ├── triage-labels.md # five roles → meta.triage / Status: lines
│ ├── frontend-craft.md # impeccable binding + the mechanical design-review gate
│ ├── domain.md # GLOSSARY.md + docs/adr/ conventions
│ └── voice.md # bilingual technical voice contract (STC + STE rules)
└── guides/
└── mp-integration.md # phase map, lane rules, skill precedence (pull-mode doc; guides/index.md stays yours)
Each package ships an index.md; the session-start hook injects those indexes,
and the agent reads the contracts they list. The contracts themselves are
pull-mode docs — nothing parses per-file path globs. voice.md anchors the
bilingual tone standard (Simplified Technical Chinese / STC across C1–C10/D1–D8/I1–I8
and Simplified Technical English / STE per ASD-STE100); all pack-owned skills and
templates are retrofitted to this standard.
--append adds missing files only — safe on existing spec trees.
After install, the files are yours to edit (Trellis's project-ownership model).
npx skills add ScoFan-official/oh-my-trellis --agent <platform> --copyInstalls 43 skills — mattpocock's full collection
(engineering, productivity, misc, in-progress) plus impeccable, the frontend
design workflow, plus the pack's seven first-party skills:
mp-trellis-bridge, oh-my-update, trellis-domains, trellis-run,
dispatch-roster, hygiene-sweep, and rules-audit. --agent accepts devin, codex,
claude, cursor and the other platforms the skills CLI knows.
mp-trellis-bridge is a bootstrapper + routing contract: on first load it
verifies .trellis/, installs the channel-1 spec contracts (registry first,
bundled templates as fallback), appends an ## Agent skills block to
AGENTS.md, checks the mp skills are present, and gates the close-out question
toward dispatch-roster when planning finishes with tickets. No-op afterwards.
scripts/install.py materializes any subset of the catalog into a target repo,
converted to that platform's native format.
# Inventory
python scripts/install.py --list
python scripts/install.py --list --platform codex # + where things land
# Everything, one platform
python scripts/install.py --platform codex --target /path/to/repo
python scripts/install.py --platform devin --target /path/to/repo
python scripts/install.py --platform zcode --target /path/to/repo
# Recommended core only (bridge + mp main skills, no in-progress/misc)
python scripts/install.py --platform zcode --target . --only-core
# Surgical picks — by id, by name, or by glob
python scripts/install.py --platform claude --target . \
--component "ecc:agents/security/*" "ecc:commands/git-workflow/commit" "mp:engineering/tdd"
# Filter by type or source
python scripts/install.py --platform cursor --target . --type skill agent
python scripts/install.py --platform copilot --target . --source mp
# Preview
python scripts/install.py --platform devin --target . --dry-runFlags: --platform (required), --target (repo root, default .),
--type skill|agent|command|mcp, --source mp|ecc,
--component <id|name|glob>, --only-core, --dry-run, --list.
Asset IDs follow source:type-path: mp:engineering/tdd,
ecc:skills/development/api-design-principles,
ecc:agents/security/security-auditor, ecc:commands/git-workflow/commit,
ecc:mcps/database/dbhub. Full IDs live in catalog/manifest.json.
- skills →
<platform skills dir>/<name>/— plain folder copy (universal format) - agents → platform's sub-agent file + converted frontmatter
- commands → platform's command/workflow/prompt file, or a user-invocable
skill (
ecc-<name>/SKILL.mdwithdisable-model-invocation) on platforms with no command primitive - mcps → merged into the platform's MCP JSON config; Codex gets
.codex/mcp-<name>.tomlsnippets to merge intoconfig.toml
| Type | Trigger model | You say… |
|---|---|---|
| skills (889) | Semantic auto-match: platform injects each SKILL.md's name+description into the agent's available-skill list; it self-loads when your task matches ("something is broken" → diagnosing-bugs). Explicit mention also works. |
"debug this" or "use api-design-principles" |
| commands (288) | Explicit: with a native command primitive → platform command syntax; without (codex/devin/kiro) → wrapped as disable-model-invocation skills, invoked by name. |
"/ecc-commit" or "run ecc-commit" |
| agents (422) | Delegation: platform's sub-agent mechanism (Claude Task, Codex subagents); auto-delegated on description match or invoked by name. Skipped on devin/kilo/antigravity (no primitive — read the body as an inline persona if needed). |
"have security-auditor review this" |
| mcps (104) | Config activation: merged into platform MCP config, handshake on restart, tools join the toolbox. Most need credentials — fill <your-...> placeholders first. |
tools just appear |
Don't bulk-install all 889 skills — every name+description costs system-prompt space and dilutes matching precision. Use --only-core or --component picks; treat catalog/INDEX.md as the shelf and pull from it as needed.
| Platform | skills dir | sub-agents | commands | mcp | status |
|---|---|---|---|---|---|
| claude | .claude/skills |
.claude/agents/*.md |
.claude/commands/ecc/* |
.mcp.json |
✅ verified |
| cursor | .cursor/skills |
.cursor/agents/*.md |
.cursor/commands/ecc-* |
.cursor/mcp.json |
✅ |
| opencode | .opencode/skills |
.opencode/agents/*.md (+permission:) |
.opencode/commands/ecc/* |
opencode.json |
✅ |
| codex | .agents/skills |
.codex/agents/*.toml (developer_instructions) |
→ skills | .codex/mcp-*.toml |
✅ |
| kiro | .kiro/skills |
.kiro/agents/*.json |
→ skills | .kiro/settings/mcp.json |
✅ |
| gemini | .agents/skills |
.gemini/agents/*.md |
.gemini/commands/ecc/*.toml |
.gemini/settings.json |
✅ |
| qoder | .qoder/skills |
.qoder/agents/*.md |
.qoder/commands/ecc-* |
.qoder/mcp.json |
✅ |
| codebuddy | .codebuddy/skills |
.codebuddy/agents/*.md |
.codebuddy/commands/ecc/* |
mcp.json |
✅ |
| copilot | .github/skills |
.github/agents/*.agent.md |
.github/prompts/*.prompt.md |
.vscode/mcp.json |
✅ |
| droid | .factory/skills |
.factory/droids/*.md |
.factory/commands/ecc/* |
.factory/mcp.json |
✅ |
| pi | .agents/skills |
.pi/agents/*.md |
.pi/prompts/ecc-* |
.pi/mcp.json |
✅ |
| devin | .devin/skills |
— (no primitive; inline) | .devin/workflows/ecc-*.md |
.devin/mcp.json |
✅ |
| kilo | .kilocode/skills |
— (inline) | .kilocode/workflows/ecc-* |
.kilocode/mcp.json |
✅ |
| zcode | .zcode/skills |
.zcode/agents/*.md (no tools:) |
.zcode/commands/ecc-* |
.zcode/mcp.json |
|
| omp / reasonix / trae / grok / kimi / snow / dsh / antigravity | .X/skills (resp.) |
.X/agents/*.md (conv.) |
`.X/commands | workflows/ecc-*` | .X/mcp.json |
✅ = file locations follow Trellis's documented per-platform tables.
verified: false in scripts/platforms.py /
manifest.json); treat generated agent/command files as drafts to review.
zcode agents are verified against the real Trellis zcode template; its
commands/MCP paths are conventional.
Per-component "when to use": see catalog/INDEX.md —
an auto-generated index where every one of the 1,741 components carries its
own description ("use when…") extracted from its source frontmatter.
Regenerate with python scripts/build_index.py.
The engineering spine of the pack. Main flow: grill → spec → tickets → implement → retro, with two intake ramps and standalone utilities.
| Skill | When to use |
|---|---|
ask-matt |
Router — "which skill fits my situation?" Ask first when unsure. |
grill-with-docs |
Align on an idea: agent interviews you through a design tree, lands terms in GLOSSARY.md + decisions in docs/adr/. Always use in-repo (over grill-me). |
grill-me / grilling |
Same interview, no doc side-effects (no working dir). grilling is the engine; grill-me the user-invoked entry. |
to-spec |
Freeze a grilled conversation into a spec (Problem/Solution/User Stories/Decisions) → Trellis task + prd.md. |
to-tickets |
Slice a spec into tracer-bullet tickets → child tasks + blocked_by meta. Confirm the split before it writes. |
implement |
Build one ticket: TDD red-green inside pre-agreed seams → typecheck/tests → code-review → commit. One ticket per session. |
implement-spec |
Batch-orchestrate a whole spec's ticket graph (integration branch + parallel worktrees). On Devin, run sequentially or via run_subagent. |
tdd |
Red-green-refactor loop, embedded inside implement. |
code-review |
Two-axis review (Standards + Spec) of changes since a ref. Complements trellis-check. |
codebase-design |
Deep-module vocabulary (interface/depth/seam/deletion test) for design discussions. |
improve-codebase-architecture |
Scan for shallow modules → visual HTML report → grill through a pick. Run every few days on growing code. |
diagnosing-bugs |
Hard bugs/regressions: failing reproduction first, then hypothesize-instrument-fix-regress. |
domain-modeling |
Build/maintain GLOSSARY.md + ADRs; fires lazily inside grill/wayfinder. |
research |
Delegate a question to high-trust sources → markdown findings file. |
prototype |
Throwaway prototype to answer a design question (the grill↔prototype loop). |
triage |
Triage external reports: needs-triage→needs-info/ready-for-agent/ready-for-human/wontfix. Inbox in .scratch/, promote to tasks. |
wayfinder |
Fog-level projects: parent task + map.md + decision tickets (research/prototype/grilling/task), max one resolved per session. |
retro |
Post-build env tune-up: navigation pointers, deterministic checks, review checklists. |
handoff |
Compress session → handoff doc for the next agent/session; auto-redacts secrets. |
teach |
Turn a directory into a multi-session learning workspace (mission→resources→lessons→records). |
to-questionnaire |
Turn a stuck decision into a markdown questionnaire for the right human. |
wait-what |
"Re-explain that" — simplified technical English + glossary terms. |
writing-for-agents |
Authoring/editing skills, AGENTS.md, CLAUDE.md. |
pr / wizard |
PR body; interactive bash wizard for human-only steps (infra, dashboards). |
setup-matt-pocock-skills |
Legacy bootstrapper for plain (non-Trellis) repos — superseded here by mp-trellis-bridge. |
claude-handoff, loop-me, setup-ts-deep-modules, writing-beats, writing-fragments, writing-shape |
in-progress collection — usable but less polished. |
git-guardrails-claude-code, migrate-to-shoehorn, scaffold-exercises, setup-pre-commit |
misc — platform-specific or niche. |
mp-trellis-bridge |
Ours. The contract + router binding all of the above to Trellis. Verifies .trellis/, writes AGENTS.md block, checks mp skills, and hands off to dispatch-roster when planning finishes. |
oh-my-update |
Ours. Gate-checked update of the whole stack — fork CLI, .trellis/ templates, installed skills, .trellis/spec/ — to the latest ScoFan-official releases (filters cli-v* tags). |
trellis-domains |
Ours. Thin shell over the .trellis/domains/ layer — 归口 routing, board flag order criterion, takeover read-chain, worklog/收工 and 对账 checklists that point at the repo's DISCIPLINE.md / WORKLOG-PROTOCOL.md / REGISTRY.md (the protocol text never lives here). |
trellis-run |
Ours. Thin shell over trellis run, the unattended frontier loop — when the loop is the right tool, what the tier + delivery.auto_push_refs must already allow, where the run ledger lives (and why it is never a reconciliation source), five stop reasons, always_stop guard, and the orchestrator boundary with dispatch-roster. |
dispatch-roster |
Ours. Staffs the dispatch roster (派工表) for unattended runs: probes ACP targets (grok-build / codex / claude) via lightweight handshake + credential checks (3 states), settles four decisions (orchestrator, role defaults, ticket overrides, max_parallel), writes to disk, and hands off to trellis-run. |
hygiene-sweep |
Ours. Read-only inventory audit inside a declared scope prefix — dead code, duplicated logic, oversized files — writing a durable findings list before any edit. Bundles scope_check.py, the mechanical gate (scope prefix + 800-line split-plan rule) that cleanup tickets run against spec/agents/hygiene.md. |
rules-audit |
Ours. Reconciles always-loaded rule files (AGENTS.md / CLAUDE.md) against reality: per-rule evidence check, prune the evidence-free, annotate the kept with why they exist. |
pbakaus/impeccable, Apache-2.0 — the
frontend craft workflow. Vendored markdown-only: the /impeccable <cmd>
skill (shape, craft, critique, audit, polish, bolder, typeset…)
works on every platform; the 61 deterministic detector rules, live, generate,
hooks, doctor and the impeccable context boot loader need upstream's
engine — npx impeccable install in the target project. Without it, commands
fall back to the skill's built-in degraded mode.
In Trellis repos, agents/frontend-craft.md binds impeccable's artifacts:
PRODUCT.md → .trellis/spec/product.md, DESIGN.md → .trellis/spec/design-system.md;
wires shape → craft → audit+critique → polish into the phases; and gates
frontend tasks — no ## Design review section in implement.md, no READY.
Largest community catalog of agent components (aitmpl.com). Counts are real,
from install.py --list:
| Type | Count | Top categories | Notes |
|---|---|---|---|
| skills | 889 | development 230, scientific 135, ai-research 131, productivity 50, business-marketing 48, security 47, creative-design 33, enterprise-communication 33, web-development 30, career 21… | SKILL.md dirs w/ references/scripts; copy to any platform's skills dir |
| agents | 422 | expert-advisors 52, programming-languages 50, devops-infrastructure 40, data-ai 40, development-tools 35, security 25, business-marketing 21, development-team 18, deep-research-team 16, web-tools 16… | Claude md → converted per platform |
| commands | 288 | google-workspace 47, utilities 21, project-management 20, testing 17, svelte 16, orchestration 15, git-workflow 14, sync 14, team 14, deployment 11… | Claude commands → converted/wrapped |
| mcps | 104 | devtools 49, web-data 10, database 8, integration 8, browser_automation 6, web 6, productivity 5… | mcpServers JSON snippets |
Not ported (kept in catalog/ as reference, Claude-Code-specific):
hooks 81, settings 75, mods ~40, sandbox 23 (cloudflare/docker/e2b), loops 18.
Every component's when-to-use: catalog/INDEX.md.
marketplace/ # channel 1: Trellis spec registry (index.json + agent-workflow w/ voice.md)
skills/ # channel 2: npx-skills surface — 5 pack-owned skills + 37 mp skills + impeccable
catalog/
├── INDEX.md # generated: every component + when-to-use
├── manifest.json # generated: every asset → per-platform target paths
├── upstream.json # pinned upstream SHAs
├── ATTRIBUTION.md # MIT provenance
├── mattpocock/ # skills(4 collections) + docs + .claude-plugin + .agents + LICENSE
├── impeccable/ # skills/impeccable + docs + LICENSE + NOTICE.md (Apache-2.0)
├── ecc/ # skills agents commands mcps hooks loops mods settings sandbox + LICENSE
└── voice/ # bilingual voice lint toolchain (voice_lint.py, stc/, ste/)
scripts/
├── platforms.py # 22-platform location table (single source of truth)
├── install.py # materialize catalog → platform-native files
├── build_manifest.py # regen manifest.json
├── build_index.py # regen INDEX.md
└── sync_upstream.py # re-vendor upstreams (--check = drift report)
Consumers: single entry point — the /oh-my-update workflow (Devin) or
the oh-my-update skill (skills/oh-my-update/). It checks the fork CLI
against this repo's cli-v* releases (the CLI tarball channel lives here
alongside pack v* releases), refreshes installed skills and
appends spec contracts, with a confirm gate. Layer-specific fallback:
npx skills add ScoFan-official/oh-my-trellis --agent <platform> --copy.
Pack maintainers (re-vendoring upstreams):
python scripts/sync_upstream.py --check # did upstreams move?
python scripts/sync_upstream.py # re-vendor catalog/ (overwrites catalog only)
python scripts/build_manifest.py && python scripts/build_index.pyInstalled spec files are project-owned (Trellis model) — merge upstream spec
changes intentionally. catalog/ is vendored; never patch it in place.
- Python 3.8+, git
npx skills(Node.js) for channel 2- A Trellis-managed repo only for channel 1 (the spec registry)
- Claude-format extras (hooks/settings/etc.) require Claude Code to use as-is
Private/forked copies? Repo is public — all three channels work without
auth. Fork to pin your own snapshot; swap --registry gh:you/fork.
Which platform am I on? install.py --list --platform X prints where each
asset type lands for that platform before you install anything.
Too many skills? Channel 3 defaults to your explicit selection; --only-core
gives the ~28-skill recommended set. npx skills add --skill <name> installs
single skills from skills/.
Found a wrong platform path? Fix scripts/platforms.py — it's the single
source of truth — then rerun install.py and build_manifest.py.