| name | groundwork |
| description | Scaffold and maintain a reusable research → design → plan → orchestrate → act
folder for any non-trivial work — software features, marketing campaigns,
org changes. Drops a domain-agnostic spine (00-README · 01-plan · 02/03
research · 04-discussion newest-first · 05-tracking · 09-orchestration ·
artifact/board.html plan-board) plus stateless action-skills that augment
the docs in place without clobbering hand-written prose. Composes
ikenga-artifact-builder, huashu-design, frontend-design, ikenga-pkg-builder
when present; degrades gracefully when not. Profile-driven: `software`
(rich default, code work), `general` (lean, non-code — campaigns, org
changes), `content` (editorial/marketing with key art), `design-system`
(component/token systems — adds parts + quality-gate docs to the spine),
and `film` (filmmaking pre-production — logline, treatment, character /
location bible, lookbook, shot ledger + render budget; hands off to Studio
for execution). Pick `film` over `content` for anything shot-based — a
short film, music video, trailer, or any AI-generated film work.
TRIGGER when the user asks to start a real plan for non-trivial work
("plan a feature," "scaffold a plan folder," "set up groundwork for…"),
references an existing plans/ folder by groundwork structure, or runs
any of these actions: groundwork init / research / design / subplan /
review / clarify / orchestrate / refresh-board / refresh-living-spec /
explorer / plans-index / status.
DO NOT TRIGGER for one-off code changes, single-document writeups, ADRs,
or content that fits in a single markdown file — those don't need a
multi-doc plan folder. If the user just wants a single artifact
(dashboard, mockup), route to ikenga-artifact-builder instead.
|
| allowed-tools | Read, Write, Edit, Bash, Glob, Grep, AskUserQuestion, Agent, Skill, Workflow, TaskCreate, TaskUpdate, TaskList, TaskGet |
groundwork — research → plan → orchestrate → act, packaged
This file is a router. Each action and each agent role lives in its own file under actions/ and agents/; load on demand. The state mechanism (.groundwork.json, fences, hash-diff, IDs) is fully specified in lib/state.md — every action obeys it. Read lib/state.md before doing anything that writes to disk.
The full design lives in plans/groundwork/01-plan.md; locked decisions in plans/groundwork/04-discussion.md (Rounds 1–4). This skill is the executable form of that plan.
What groundwork is for
Stripped to one sentence: a stateless action-set that scaffolds and maintains a folder of living planning documents, with traceable IDs threading plan → tracking → orchestration → board, so re-runs augment instead of clobber and a multi-agent kickoff falls out the back end. Originated from the com.ikenga.studio planning workflow; generalized so it works for non-code, non-Ikenga work too.
Surface model:
| Surface | What it is | Trigger |
|---|
Scaffolder (init) | First-run interview + folder skeleton drop | User has a goal but no plan folder |
Action: research | Researcher agent fills 02/03 | After scaffolder, or to refresh |
Action: design | Produces ≥2 comparable designs/*.html for the current phase, locks one | Visual/UX work, profile-gated |
Action: subplan | Scaffolds a focused NN-*.md from one of three archetypes (diff-plan / decision-doc / bug-doc) | A hard piece needs its own focused doc (critical-path PR plan, between-round deliberation, postmortem) |
Action: review | Reviewer agent → new Round in 04 → re-sync via IDs | Recurring, highest-value |
Action: clarify | Readiness gate before orchestrate | Before kickoff |
Action: orchestrate | Generates 09-orchestration.md from 05 | When ready to kick off |
Action: refresh-board | (Re)generates artifact/board.html & artifact/explorer.html and opens/returns artifact/explorer.html | Whenever board falls behind |
Action: refresh-living-spec | Regenerates the spec-state fence inside artifact/index.html (Phasing/Decisions/Risks tabs) from .groundwork.json + 04 + 05 + 01 §Risks | When the living-spec's auto tabs fall behind |
Action: explorer | (Re)generates artifact/explorer.html — a file-tree + tabbed viewer (with search) of the whole plan folder; sibling to the board (opens it as a tab) | Browse the plan's files as one navigable view |
| Action: |
init is the one entry point that doesn't require an existing .groundwork.json. Every other action refuses to run without one (and tells the user to run init first).
Routing
Match the user's request against the table; load the matching action file and follow it. Do not inline action behavior in this file — the actions are the source of truth.
| If the user says… | Load this | Then |
|---|
| "scaffold groundwork", "init", "new plan folder", "start groundwork in X/" | actions/init.md | Interview goal + profile, scaffold, write .groundwork.json |
| "groundwork research", "do a research pass", "fill 02/03" | actions/research.md | Spawn researcher; write inside fences (--sweep → multi-modal Workflow fan-out) |
| "groundwork design", "design pass", "mock up the UI options" | actions/design.md | Profile-gated; compose design skill or plain HTML; lock one |
| "groundwork subplan", "scaffold an NN-*.md", "new diff-plan / decision-doc / bug-doc" | actions/subplan.md | Scaffold a focused sub-plan from one of three archetypes; register in .groundwork.json.subplans; cross-link from 00-README |
| "groundwork review", "review pass", "gap analysis" | actions/review.md | Spawn reviewer; append Round; re-sync via IDs (--panel → verified Workflow panel) |
| "groundwork clarify", "ready to orchestrate?" | actions/clarify.md | Scan for open questions + unspecified IDs + missing locked designs |
| "groundwork orchestrate", "generate 09", "wave plan" | actions/orchestrate.md | Runs clarify first; emits 09-orchestration.md (--emit-workflow → runnable .workflow.js + offer to run) |
| "refresh the board", "update artifact/board.html" | |
Discover vs. fast path
- Discover path — invoked cold ("scaffold a plan for X"). Run the
init interview to capture goal + profile, then suggest research as the next step.
- Fast path — invoked in an existing groundwork folder ("groundwork review"). Read
.groundwork.json, dispatch to the named action immediately. No interview unless the action itself needs one.
If invoked from inside a folder with no .groundwork.json, treat as discover even if the action name was given — refuse the action with a one-line "this folder isn't a groundwork plan yet; run init first."
Click-to-fire prompts (per-action)
Each action file under actions/ carries a ## Click-to-fire prompt section at the bottom with two flavors: a standalone slash form (/groundwork <action> [args]) and a seeded-session form (a self-contained brief that works even in a chat where the skill isn't loaded). The Phase-4 board surfaces (Kickoff card / ⌘K palette / argument pickers) read these as their canonical strings — they are the single source of truth shared between the FE and the orchestrator agent. Substitution variables are documented inline per action (typically {plan_folder}, {plan_title}, {plan_slug}, {profile}).
Profiles
A profile swaps vocabulary and optional blocks, not the spine. Templates use {{vocab.*}} placeholders (same substitution convention as ikenga-pkg-builder's {{id}}/{{slug}}).
| Profile | Default for | 05 work-unit | produces_designs | Optional 01 blocks |
|---|
software | Ikenga features, any code work | "work package" / PR | true | schema, manifest, adapter contracts, critical files |
general | Non-code work — campaigns, org changes, research | "workstream" / "deliverable" | false | stakeholders, deliverables, success metrics |
content | Editorial/marketing — content series, campaigns with key art | "piece" / "asset" | true | editorial standards, distribution plan |
design-system | Component/token systems — design-language work | "part" | true | mode, taxonomy, tokens, fixtures, gallery, quality_gate |
film | Filmmaking pre-production — short films, music videos, trailers, AI-generated film | "sequence" | true | treatment, lookbook, shotlist, shot_tracker, budget, schedule |
Choosing between content and film: both are creative and both produce designs, so the split is shot-based vs not. If the work has shots — a short film, music video, trailer, any AI-generated film — use film; its vocabulary (sequence / scene + reel / picture lock) and its shot ledger + render budget are what that work actually needs. Use content for editorial: articles, campaigns, key art. "It's creative, not code" is not a reason to reach for content — that reasoning lands film work in the wrong profile.
Maintenance model: a single profiles/_shared/ base + thin per-profile overlays. profile.json declares extends: "_shared"; only diffs need to live in the overlay. Format spec in 01-plan.md §"Domain profiles." design-system additionally declares its own spine (adding parts-template.md + quality-gate.md to the standard six docs).
film hands off, it does not execute: the profile owns creative development and production management (logline → treatment → bible → lookbook → schedule → budget). It does not own the shot board — com.ikenga.studio holds the authoritative per-shot render/approval state, and 05-tracking.md is only a status mirror of it. When the two disagree, Studio wins.
status runs a profile-conformance check to catch drift in user-dropped profiles.
State, identity & idempotency
The contract that makes stateless actions safe over an existing folder:
.groundwork.json at the folder root is the identity + state anchor every action reads first.- Generated-region fences (
<!-- groundwork:auto:start ID --> … <!-- groundwork:auto:end ID -->) demarcate the only blocks an action may write. Everything outside a fence is hand-authored and never touched.
- Re-runs diff, not overwrite — an action recomputes a region, hashes it, and writes only when the hash differs.
- Stable IDs (
G-NN gaps, WP-NN work packages, G-<NAME> gates, D-NN designs) thread 01 → 05 → 09 → board; the review action computes the affected-doc set from IDs + hashes, not from guessing.
The full schema, fence grammar, hash-diff algorithm, and a worked example: lib/state.md. Read it before any action that writes.
Composition
groundwork composes existing skills rather than reimplementing them. Soft dependencies — every composed skill has a fallback:
| Capability | Preferred skill | Fallback |
|---|
| Plan-board artifact | ikenga-artifact-builder | Self-contained template at profiles/_shared/board/index.html |
| Plan-explorer artifact | — (no file-explorer archetype fits; always the static template) | Self-contained template at profiles/_shared/explorer/index.html (+ design-system / content overlays). Fully offline — React + marked + DOMPurify are inlined and the JSX is pre-transpiled (no Babel/CDN at runtime; rebuild from *.src.html via scripts/build-offline-artifacts.mjs) |
| Cross-plan dashboard | — (static template) | Self-contained template at profiles/_shared/plans-index/index.html; fully offline (inlined libs, pre-transpiled); written anchorless (write-region-plain) since it spans many plans |
Design-quality spine (every design run) | huashu-design (always engaged at step 2, regardless of surface) | Plain self-contained HTML under Claude's own craft direction |
| Interactive / data-bearing design mockup | ikenga-artifact-builder (layered with huashu) | Plain self-contained HTML (studio's mockups needed nothing more) |
| Scroll-driven narrative mockup | scrollytelling (layered with huashu) | Plain self-contained HTML |
| Hi-fi / production frontend | example-skills:frontend-design / web-artifacts-builder (layered with huashu) | huashu-design alone |
| Anti-slop polish / critique / audit pass | impeccable (layered with huashu) | huashu's own expert-review pass |
| Build handoff for Ikenga pkgs | ikenga-pkg-builder | 09-orchestration.md is the terminal deliverable |
| Multi-agent execution / fan-out (opt-in) |
The design action does not pick one skill from this list — huashu-design is the always-on quality spine and Claude layers any combination of the others on top, deciding the blend while building (no limits). If ikenga-artifact-builder is absent, refresh-board writes the self-contained fallback template. If no design skill at all is installed, design falls back to plain HTML.
Workflow (opt-in execution layer)
Three actions can deepen into a Workflow fan-out when the user asks for it — never by default, because Workflow is heavyweight, explicit-opt-in, and billed:
| Flag | Action | What it does |
|---|
--emit-workflow | orchestrate | Also writes artifact/orchestrate.workflow.js — the runnable twin of 09: waves → parallel() fan-out, freeze gates → adversarial sign-off barriers, WPs → worktree-isolated, schema-validated runs. Then offers to run it. |
--sweep | research | Fans out finders by angle (precedent / library-docs / codebase / constraints) → synthesis. |
--panel | review | Diverse-lens reviewer panel → dedup → perspective-diverse adversarial verify (kills plausible-but-wrong findings). Implements the old --multi stub. |
All three return schema-validated results (contracts in lib/schemas.md) and the calling session does every write — Workflow scripts have no filesystem, so fences, IDs, 05 checkboxes, Tasks, commits, and merges stay in the action's hands, exactly as in the single-agent path. Drop the flag (or run without Workflow) and behavior is byte-identical to today. Each agent declares a model tier (haiku/sonnet/opus) per the global subagent-model policy.
Portability
groundwork is a portable Claude Code skill. The Ikenga coupling is an additive layer:
| Capability | In Ikenga shell | Plain Claude Code / terminal | Any browser |
|---|
| Scaffold + actions | ✅ | ✅ | n/a |
| Research / review / orchestrate | ✅ | ✅ | n/a |
| Plan-board (view) | ✅ live status | ✅ open file | ✅ static |
| Copy brief from board | ✅ | ✅ | ✅ |
| Start session (click-to-implement) | ✅ (Phase 2) | ➖ copy-prompt | ➖ copy-prompt |
| Live tracking binding | ✅ (Phase 2) | ➖ reads 05 | ➖ |
Phase 1 (this skill) ships everything in the "Plain Claude Code / terminal" column and the static-board column. The host.startChatSession verb + live-tracking binding land in Phase 2 and are explicitly out of scope here — do not modify shell/ or contract/.
Critical files (this skill)
groundwork/
├── SKILL.md ← you are here (router only)
├── lib/
│ ├── state.md ← state machine spec — every action obeys
│ └── schemas.md ← Workflow result contracts + model tiers
├── actions/
│ ├── init.md
│ ├── research.md
│ ├── design.md
│ ├── subplan.md ← scaffold NN-*.md sub-plans (diff-plan / decision-doc / bug-doc)
│ ├── review.md
│ ├── clarify.md
│ ├── orchestrate.md
│ ├── refresh-board.md
│ ├── refresh-living-spec.md ← regenerate the spec-state fence in artifact/index.html
│ ├── explorer.md ← (re)generate artifact/explorer.html (file-tree + viewer + search)
│ ├── plans-index.md ← (re)generate <plans-dir>/_index.html (cross-plan dashboard)
│ └── status.md
├── agents/
│ ├── researcher.md ← brief template the research action spawns
│ ├── reviewer.md ← brief template the review action spawns
│ └── orchestrator.md ← brief template the orchestrate action spawns
├── scripts/
│ ├── groundwork_state.py ← byte-exact state executor (all subcommands)
│ ├── generate_explorer.py ← automate `explorer` for one plan / all plans (--all-under --missing-only)
│ ├── generate_plans_index.py ← automate `plans-index` for a plans/ dir
│ ├── watch.py ← live-refresh an explorer / index as files change
│ └── build-offline-artifacts.mjs ← rebuild the offline index.html from *.src.html (inlines libs, transpiles JSX)
└── profiles/
├── _shared/
│ ├── profile.json ← base vocab + spine list
│ ├── templates/ ← spine docs with {{vocab.*}} + fences
│ │ ├── 00-README.md
│ │ ├── 01-plan.md
│ │ ├── 02-research-external.md
│ │ ├── 03-research-internal.md
│ │ ├── 04-discussion.md
│ │ ├── 05-tracking.md
│ │ ├── subplans/ ← NN-*.md archetype templates
│ │ │ ├── diff-plan.md
│ │ │ ├── decision-doc.md
│ │ │ └── bug-doc.md
│ │ └── artifact/ ← scaffolded into <plan>/artifact/
│ │ ├── index.html ← living spec (Overview + Phasing/Decisions/Risks)
│ │ ├── manifest.json
│ │ ├── orchestrate.workflow.js ← runnable twin of 09 (--emit-workflow)
│ │ └── data/ ← optional external data sources
│ ├── board/index.html ← self-contained Mission Control + Wave-toggle
│ ├── explorer/ ← index.src.html (readable JSX source) → index.html (built, offline)
│ │ └── … ← file-tree + tabbed viewer + search; base, profile-adaptive
│ └── plans-index/ ← index.src.html → index.html (built, offline); cross-plan dashboard
├── software/ ← `extends: _shared`, produces_designs: true
│ ├── profile.json
│ └── templates/ ← thin overlays (only the docs that differ)
├── general/ ← `extends: _shared`, produces_designs: false
│ ├── profile.json
│ └── templates/
├── content/ ← `extends: _shared`, produces_designs: true (key art)
│ ├── profile.json
│ ├── templates/
│ └── explorer/index.html ← media-first explorer overlay
├── design-system/ ← `extends: _shared`, +parts-template + quality-gate docs
│ ├── profile.json
│ ├── templates/
│ └── explorer/index.html ← gallery-first explorer overlay
└── film/ ← `extends: _shared`, produces_designs: true (lookbook / key frames)
├── profile.json
└── templates/ ← 01-plan (pre-production bible) + 05-tracking (shot ledger, Studio mirror)
The reference instance — the canonical worked example of every artifact in this spine — is plans/studio/ in this workspace. Re-derive the spine from it on major spine_version bumps.
Caveat on the studio reference: studio predates two later locks — the living-spec / tracking-board split (commit a6bc357: studio's artifact/index.html is a single tabbed artifact, not the index.html living-spec + board.html tracking-board pair this skill scaffolds) and the .groundwork.json state anchor (studio has none). Read studio for the spine shape and the prose conventions (Round bodies, WP briefs, lifted-from headers, per-phase design discipline) — not as a literal current-shape template. The dogfood at plans/groundwork/ is the up-to-date worked example of the post-split, anchored layout.
What this skill does NOT do (in P1)
- No shell or contract changes.
host.startChatSession is Phase 2 — board actions are copy-prompt only.
- No live tracking binding. The board reads markdown; harness-Tasks mirroring is Phase 2.
- No board "Start session" button beyond copy-prompt. Same reason.
- No automatic renumbering of IDs. Once allocated, an ID stays — retire instead of free.
- No
groundwork kickoff action. The dogfood at plans/groundwork/KICKOFF.md is a hand-authored cross-session resumption snapshot (phase scope, locked decisions, what's-next) — useful, but a dogfood-specific artifact, not part of the scaffolded spine. A kickoff action that auto-snapshots the current phase boundary is a deferred stretch idea (Phase 2+, once host.startChatSession makes a snapshot worth seeding a session from); it is not in P1 and init does not scaffold a KICKOFF.md.
If the user asks for any of the above, say so explicitly and route them to the right phase.
