| name | aiwfx-whiteboard |
| description | Open-work synthesis ritual โ answers direction questions like "what should I work on next?", "give me the landscape", "where should we focus?", "what's the critical path?", "synthesise the open work", "draw the whiteboard". Loads tree state via `aiwf status` / `aiwf list` / `aiwf show` / `aiwf history`; produces a tiered open-work landscape, a recommended sequence, a first-decision fork, and an optional Q&A gate over pending decisions. Read-only over the planning tree โ no mutation, no commit; writes a gitignored `WHITEBOARD.md` cache (regenerated each run, never committed). |
aiwfx-whiteboard
Synthesises the open-work landscape into a tiered view, a recommended sequence, a first-decision fork, and a Q&A gate over pending decisions. The output is conversational, not authoritative โ the operator decides; the skill surfaces.
Tier classification rubric
Classify each open item by leverage on future work, not by chronology of when it appeared. The criteria below are reproducible; tier contents may vary at the margin (LLM judgement on borderline items is acceptable; the criteria themselves do not move).
Tier 1 โ compounding fixes
Criterion: the item, once closed, removes friction from every future planning or implementation session. Typical shape: a kernel asymmetry, a chronic warning class, a missing-verb gap that forces a workaround at every use.
Example archetypes:
- A kernel rule that is lifecycle-blind โ fixing it removes a standing warning class that recurs on every build.
- A missing writer verb for milestone
depends_on โ every multi-milestone epic re-derives the workaround.
- A missing-verb gap โ every scope-change moment loses the fix option.
Tier 2 โ architecturally foundational
Criterion: the item is a proposed-status ADR whose ratification + implementation epic gates downstream work. Typical shape: data-shape, lifecycle, or namespace decisions whose absence forces every consumer to invent a local convention.
Example archetypes:
- A
proposed entity-id-minting-policy ADR โ foundational for parallel-branch work and downstream decisions.
- A
proposed uniform-archive-convention ADR โ makes high-volume kinds tractable for read verbs.
- A
proposed new-entity-kind ADR โ substrate for AC-closure chokepoints and cycle-time findings.
Tier 3 โ workflow rituals
Criterion: the item is a missing or under-defined ritual that the operator currently re-derives in conversation each time. Smaller leverage than Tier 1/2, but codification removes ad-hoc thrash.
Example archetypes:
- An under-defined branch-model ritual โ no canonical mapping from epic/milestone hierarchy to git branches.
- A loosely-defined patch ritual โ small fixes lack a canonical shape.
- A missing activation ritual โ the sovereign act has no preflight today.
Tier 4 โ operational debris
Criterion: small, isolated fixes that don't compound but are cheap to batch. Typical shape: a one-line .gitignore change, a single config nudge, a typo. Leverage is per-item, not per-session.
Example archetypes:
- A
.gitignore miss โ rendered site/ output not ignored.
- A stray binary in the repo root not ignored.
- An
aiwf init nudge that hardcodes the user-scope CLI form.
Tier 5 โ defer until a forcing function shows up
Criterion: the item is open but no current consumer or workflow forces it. Premature work here costs design effort that will be re-derived once the forcing function lands.
Example archetypes:
- A
aiwf doctor --format=json flag โ defer until a JSON consumer appears.
- A TDD-advisory skill โ couples to the agent-orchestration substrate.
- A discoverability-policy gap โ misses dynamic finding subcodes; activates when the finding kind lands.
Output template
The output is a single conversational message containing four named blocks, in this order. The action-shaped blocks (sequence, fork, pending) lead; the tiered landscape comes last as the supporting reference data.
(a) Recommended sequence โ numbered prose
Numbered list, one entry per concrete next action. Each entry uses explicit before / after / parallel framing relative to the existing in-flight work, e.g.:
- Before E-NNNN's M-NNNN starts โ fix G-NNNN; closes the warning baseline. (Cost: wf-patch.)
- After M-NNNN wraps โ ratify ADR-NNNN.
- Parallel low-priority track โ Tier 4 operational debris as a single wf-patch, any time.
Sequence is reproducible across runs given the same tree state; only the lean phrasing varies with LLM judgement.
(b) First-decision fork โ option list
The next concrete sequencing question presented as concrete options, typically A/B/C, each with pros / cons and a lean:
A. Fix Tier 1 standalone first, then start the next milestone. Pros: cleaner baseline. Cons: 2 small milestones of delay.
B. Roll Tier 1 fix into the next milestone. Pros: marginal scope, immediate warning cleanup. Cons: scope creep.
Lean: B. Reasoning: โฆ
The lean is named explicitly so the operator can agree, redirect, or re-weigh.
(c) Pending decisions โ list
Numbered list of open Q&A items implied by the synthesis. None should be blocking the next concrete action; if one is, surface it as the first-decision fork instead. Each item names what it would unlock if answered.
(d) Tiered landscape โ table
A markdown table, one row per open item across the relevant kinds (epic, milestone, gap, adr). Required columns:
| Column | Content |
|---|
| Item | id + short title (e.g. G-NNNN โ a lifecycle-blind kernel rule) |
| Kind | gap / adr / epic / milestone |
| Cost | rough sizing โ tiny, wf-patch, small milestone, medium milestone, epic, multi-epic |
| What it unblocks | one-line description of leverage on future work |
Group rows by tier (Tier 1 first, Tier 5 last). Each tier is a sub-heading above its rows. The landscape comes last because it's the supporting reference data โ the action-shaped blocks above lead with what to do; the table backs them with the full inventory.
Output cache (WHITEBOARD.md)
After rendering blocks (a)โ(d) into the conversation, write the same four blocks to WHITEBOARD.md in the consumer repo's root. The file is gitignored by convention (see .gitignore entry; the consumer repo's aiwf init / aiwf update should add this if not present). Subsequent invocations overwrite the file in place.
The cache lets the operator re-read the last synthesis without re-invoking the skill โ useful when the chat-session context has scrolled past the rendered output. The cache is not authoritative: the live tree is the truth, and WHITEBOARD.md is a snapshot that drifts from the tree the moment any planning entity changes status. Treat WHITEBOARD.md like STATUS.md: a regeneratable view, not the source of truth.
Q&A gate
After rendering blocks (a)โ(d), the skill emits exactly one gate prompt and waits:
"Walk through the pending decisions one at a time, or is the recommendation enough?"
The operator picks one of three paths:
- Walk through (Q&A). The skill walks each pending decision one at a time per CLAUDE.md Working with the user ยงQ&A format โ context, options with pros/cons, lean, numbered choice, wait. Move to the next decision only after the operator answers the current one. Never batch.
- Recommendation is enough. The skill exits cleanly with a one-line summary (
"Recommendation captured. Next: <first-decision lean>.") and stops. No follow-up questions.
- Operator names a different follow-up. The skill does not silently extend; if the follow-up is out of scope (refactor advice, design review, cross-team blocking), respond with "That sounds like its own skill โ should we file one?" and stop.
The one-at-a-time discipline is non-negotiable: batched-question rendering breaks the operator's documented preference and makes the skill's output authoritative-but-brittle, which is exactly the failure mode the gate exists to prevent.
Anti-patterns
These are the failure modes this skill exists to avoid. If a draft response drifts toward any of them, stop and reshape.
1. Replacing the operator's judgement (instead of surfacing and gating)
The skill's job is to surface structure (tiers, sequence, decisions) and gate on the operator's choice. It does not pick on the operator's behalf. Phrases like "the right answer is โฆ", "you should โฆ", "the obvious choice โฆ" are red flags. Use "the lean is โฆ" and "option B trades X for Y" instead. The operator decides; the skill recommends.
2. Inventing verbs that don't exist on the kernel surface
Every verb invocation in the skill body or its rendered output must resolve to a real aiwf command available today. If the synthesis would benefit from a verb that doesn't exist, file a follow-up gap (aiwf add gap --title "..." --discovered-in M-NNNN) and surface the gap in the output โ do not encode a hand-edit workaround or pretend the verb exists. The kernel surface is authoritative; the skill is advisory; the verb-invention failure mode confuses that hierarchy.
3. Persisting the synthesis to a checked-in file
No whiteboard.md, landscape.md, or any synthesis snapshot committed to the tree. A checked-in snapshot goes stale within hours of the next planning act and becomes a second source of truth that disagrees with the live tree. Gitignored local caches are different and OK โ they regenerate on each invocation, don't share team-wide drift, and don't tax git history. The skill writes such a cache to WHITEBOARD.md (see Output cache above); STATUS.md is the precedent (a persisted artefact regenerated on every commit by the pre-commit hook).
4. Scope creep beyond direction-synthesis
The skill answers "what should I work on next, and what decisions are pending?" โ that is the bounded scope. Adjacent functions belong in their own skills:
- "Should I refactor X?" โ not this skill; suggest a code-review or design-review skill.
- "Is this design good?" โ not this skill; suggest a design-review or an architectural-decision conversation.
- "Who's blocked on what across teams?" โ not this skill; that's a coordination-layer concern, not a planning-tree synthesis.
When the operator's follow-up looks adjacent, say so explicitly: "That sounds like its own skill โ should we file one?". Don't silently extend.
