grill-fast

Vault Location

Vault path: Read from ${CLAUDE_PLUGIN_ROOT}/brain-os.config.md

What it does

Skip the linear Q1→Q2→...→Qn march. Pre-grill internally, partition decisions into ARCH (branches the architecture) vs PARAM (single value pick), render ≥4 candidate architectures as side-by-side ASCII trees — at least 3 spanning the real design space plus a mandatory 4th tree challenging the shared framing axis — with key deltas highlighted, batch PARAM picks in one block, ask user to pick a tree (or hybrid). Locks in 2-3 user turns.

Use when topic is multi-component design with ≥3 ARCH decisions. Fall back to /grill when topic is a single decision, policy choice, or pure exploration with no candidate to draw — the ≥4-tree rule does NOT force four trees onto trivially small design spaces.

Flow

1. Vault research first — check RESOLVER.md + relevant zone READMEs + prior grills/research. Same as /grill. If the topic references GitHub issues, run gh issue view N on each FIRST — confirm it is still open and read its latest comments for prior decisions. The SessionStart open-tasks cache is stale; grilling an already-closed issue contaminates every downstream pick.
2. Decision partition — list every decision the topic implies; tag each as ARCH (changes the diagram) or PARAM (single value). If <3 ARCH decisions, abort and recommend /grill instead.
3. ≥3 axis-aligned candidate architectures (typically A/B/C) — span the real design space. Not 1 good + 2 strawmen. Each tree must be internally consistent and defensible. Common axes: strict-vs-permissive, narrow-vs-wide, ship-fast-vs-cover-everything.
4. Mandatory axis-challenger tree (4th, typically D) — identify the shared framing axis: the dimension trees A/B/C all silently agree on (e.g. "schema-driven dispatch", "entities live in tables", "pipeline as DAG"). Render an additional tree that explicitly violates that axis. The axis-challenger may be intentionally weak / impractical / strawman — its job is to surface the framing assumption, not to be a serious candidate. The strawman exception applies ONLY to the axis-challenger; A/B/C must remain defensible per Step 3. See § Challenge-the-axis tree below for full mechanics + canonical example.
5. Render each tree as a self-contained block: header + ASCII + inline pros/cons. Each tree gets a header strip naming the ARCH picks, the box-and-arrow diagram with ◄── callouts where trees differ, then a **Pros** / **Cons** block immediately under the diagram (3-5 concrete bullets each side — build cost, what it actually proves, what it misses, key failure mode). The reader scans each tree top-down without cross-referencing tables. Stack trees vertically with ═══ separators (side-by-side horizontal layout doesn't fit ≥4 trees at terminal width). Inline pros/cons is REQUIRED, not optional — a tree without it forces the reader to scroll-and-cross-reference, which defeats the "decide on the spot" goal.
6. Delta table — one row per ARCH decision, one column per tree (A/B/C/D...) showing per-tree pick. Reader sees all differences in one glance, including how the axis-challenger differs from the axis-aligned set.
7. Cross-tree comparison table — dispatcher complexity, ship velocity, future-extension cost, failure mode, etc. Complements the inline pros/cons in step 5 by giving a side-by-side scan across all trees on the same axes. NOT a replacement — both are required.
8. Parameter block — bullet list of PARAM picks (one recommendation each, no tree branching).
9. Recommendation — one pick (typically A/B/C) with rationale. Not "you decide" — Decide-and-recommend rule applies. If the user picks the axis-challenger 4th tree (rare but possible), the grill MUST re-grill the framing before locking — the shared axis was just rejected, so A/B/C are no longer the right design space; drop back to /grill with a fresh question batch around the new framing.
10. Failure-mode call-outs — for each non-recommended tree, list what to flag if user picks it anyway.
11. User pick via AskUserQuestion (default mode). After the file is written, present the trees as side-by-side preview options via the AskUserQuestion tool — vertical option list on the left, monospace preview pane on the right showing the focused option's content. Each option has:
    • label: tree letter + ARCH summary, e.g. "C: Live outcome shadow" (the recommended tree gets (Recommended) suffix per tool convention; when the recommendation is a hybrid, mark the strongest single tree as (Recommended — pair with X for hybrid) and rely on the auto-Other field for hybrid input or custom direction like "park as backlog").
    • description: 1-line tradeoff summary.
    • preview: compact ASCII tree (≤20 lines: data flow + key cost/risk annotations). Compact ≠ different — same logical structure as the file's full diagram, just trimmed for the preview pane. Question header ≤12 chars derived from topic. Recommended tree is option 1. Null-option rule: whenever "don't do this / drop it" is a live outcome (topic could legitimately be dropped, deferred, or was previously deprioritized), include it as an explicit option — never offer only sub-variants. An option set of sub-variants presupposes the action is settled and makes the question leading; a user in flow clicks the sensible sub-variant, and that click cannot bear the weight of reversing a prior considered decision. Skip this step ONLY when user has pre-stated their pick before the skill ran, or when running in a non-interactive context (e.g. /loop). After user picks, update file frontmatter (status: pass for A/B/C; status: working for D triggering re-grill; status: deferred for none/Other-deferred) and append the outcome log row. If trees > 4, present the top 4 in AskUserQuestion (recommended + 3 highest-value alternatives); the file retains all.

Output file

Save to {vault}/daily/grill-sessions/YYYY-MM-DD-<topic-slug>.md. Same path as /grill. Frontmatter tags: [grill-fast, mode-b] to distinguish.

Required sections: # Frame, # The candidate architectures (the ASCII — ≥4 trees including the axis-challenger; each tree's diagram MUST be immediately followed by an inline **Pros** / **Cons** block per step 5), # Key deltas, # Cross-tree comparison, # Parameter picks, # Recommended pick, # Failure mode call-outs, # Status gate. The shared framing axis the 4th tree challenges MUST be named explicitly in the # Frame or above the ASCII so the reader can audit the framing assumption.

When to abort to /grill

• <3 ARCH decisions identified after Step 2 (single-decision topic). The ≥4-tree rule does NOT apply here — fall back to /grill rather than fabricating axis-aligned trees just to satisfy the count.
• Vault context too thin to sketch any candidate (skip the pre-grill, ask 1-2 framing questions first).
• Decisions are PARAM-heavy with only 1-2 architectural branches (just batch the params + 1-2 questions).

Do not force a multi-tree render when the design space is genuinely 1-dimensional. The skill failure mode is "all the trees look the same" — when that happens with the axis-aligned set, the apparent sameness IS the shared framing axis: lean into the 4th tree to expose it. When it happens because the topic itself has no real design space, abort to /grill.

Challenge-the-axis tree (mandatory 4th)

The first 3 trees converge fast precisely because they share framing — that speed is exactly what makes Mode B 2-3 turns instead of 17. But shared framing becomes the weakness when the framing itself is wrong: Mode B will lock 2-3 turns deep on an axis nobody questioned. The mandatory 4th tree is the antidote.

How to identify the shared framing axis. After drafting trees A/B/C, ask: what assumption do all three silently agree on? The axis is the dimension trees A/B/C don't disagree on, even though they vary on others (NK threshold, strategy count, etc.). Common axes: schema shape ("schema-driven dispatch"), where entities live ("entities live in tables"), pipeline topology ("pipeline is a DAG"), control-flow direction ("top-down classification"), who decides X ("deterministic dispatch, never LLM judgment"). If you can't name the shared axis after 30 seconds, render A/B/C side-by-side and look for the row that's missing from the delta table — the dimension nobody varied on.

How to render the 4th tree. Negate the axis. Examples:

• A/B/C all assume "entities live in tables" → 4th tree assumes "entities live as denormalized strings inside other records."
• A/B/C all assume "pipeline is a DAG" → 4th tree assumes "pipeline is a feedback loop with mid-stage human intervention."
• A/B/C all assume "deterministic record identity" → 4th tree assumes "LLM judgment for record identity."

The axis-challenger may be intentionally weak / impractical / strawman — that's the point. The strawman exception applies ONLY to the 4th tree; trees A/B/C must remain defensible per Step 3.

Canonical example — the 2026-05-04 airtable replay. Trees A/B/C in daily/grill-sessions/2026-05-04-airtable-mode-b-replay.md (strict / permissive / wide) all silently shared the assumption entities live in tables. Had the ≥4-tree rule been in force, the mandatory 4th would have asked "what if entities don't live in tables?" — a text-extraction architecture that scans singleLineText fields inside records for entity references regardless of schema shape. That 4th tree would have surfaced the framing assumption explicitly. The assumption broke when 8 children of #237 had already shipped on a denormalized Payments base where Payer Name / Payee Name were singleLineText strings — exactly the case the 4th tree would have flagged before any code shipped.

Fallback when the user picks the axis-challenger. Rare but real. The shared axis just got rejected, so the original A/B/C design space is moot. Re-grill the framing before locking: drop back to /grill with a fresh question batch around the new framing, then re-enter /grill-fast with the new axis to render a fresh A/B/C/D set. Do NOT lock on the axis-challenger as-is — it was scored against the old framing and may have been deliberately weak.

Usage

/grill-fast <topic>

Outcome log

Follow skill-spec.md § 11. Append to {vault}/daily/skill-outcomes/grill-fast.log:

{date} | grill-fast | grill-fast | ~/work/brain-os-plugin | daily/grill-sessions/{date}-{slug}.md | commit:{hash} | {result} | args="{topic}" turns_to_lock=N pick=A|B|C|D|hybrid|none

• result: pass if user picks a tree (or hybrid) and grill closes ≤3 user turns; partial if: user requests Q-by-Q follow-up after seeing trees (Mode B incomplete, Mode A patched on top), OR user picks the axis-challenger 4th tree and re-grill of the framing is required before locking, OR user explicitly defers the session to backlog (pick=none, status: deferred in file frontmatter); fail if aborted to /grill or trees rejected wholesale
• turns_to_lock: count user turns from skill invocation to lock (success metric — target ≤3)
• pick: which tree user selected — A, B, C, D (axis-challenger), hybrid, or none
• deferred_to: optional — issue URL when pick=none and user defers to a backlog issue (e.g. deferred_to=https://github.com/.../issues/267)
• Log exactly once per session — after the final outcome is known (user pick, explicit deferral, or session abort). Do NOT append an intermediate row while AskUserQuestion is pending. If a session ends mid-flow before a pick, resolve and log in the session where the pick or deferral is confirmed.