brew

name: brew description: Use when the user wants a Fellow Aiden coffee profile for a bag they name (roaster, origin, roast) or photograph. Also triggers on /brew.

/brew — Fellow Aiden profile factory

Turn a coffee bag into a well-tuned Fellow Aiden profile, then stop. You pick the volume, read the grams, and grind at the machine — the profile carries everything needed for that. Brewing and scheduling are not this skill's job: they live in the MCP (brew_now, create_schedule) and you reach them by saying "brew now".

When to use

• User names a coffee bag (roaster, origin, roast) and wants it dialed in.
• User shows a photo of a bag.
• User says /brew.

Required input

Just the bean — from text or a photo. Ask only if roast or origin is genuinely ambiguous; never guess them. Do not ask for a target volume — the profile is volume-agnostic (a profile_type=0 profile carries both single-serve and batch settings, and the device shows the dose once you pick millilitres).

Flow

1. Identify the bean (photo/text), incl. dimensions: varietal, process, roast, origin.
2. Check coffee://profiles for an existing match — roaster + coffee-name substring (e.g. "Intermission" + "Spring Bloom"). User profiles live in the Custom folder. If one matches, reuse it and say so; skip the rest. Only create fresh if there's no match or the user asks for a new one.
3. Clone a template drop from coffee://profiles via the decision tree in references/profile-heuristics.md.
4. Palate pre-flight — call query_palate(varietal, process, roast_level, origin, altitude_masl) for the user's tasting history of similar beans (see below).
5. Apply nudges: light-natural/Geisha overrides (below) + roast-delta, plus the palate correction from step 4 if its confidence is meaningful.
6. Derive the two grind numbers — single-serve and batch (below).
7. create_profile with profile_type=0, °C temps, and the grind-span title.
8. Print the recipe card. Stop.

Palate pre-flight (step 4)

query_palate returns {tendency, confidence, n, neighbours} — tendency is the similarity-weighted mean of the signed tasting axes (−2…+2) across the user's similar past brews. If n ≥ 2 and confidence is meaningful, translate the tendency into a template nudge via the fault→lever mapping and say why:

"Your last 3 Geisha naturals averaged sour (+1.5 acidity) — starting one grind finer and ratio 16.0 instead of the template's 16.5."

Lever directions: acidity + (trends sour) → grind finer / ratio tighter; bitterness + (trends bitter) → grind coarser; strength +/− → ratio looser/tighter. If n = 0 or confidence is low, proceed from the template silently — never invent history.

Non-negotiable MCP values

These are the traps a fresh agent falls into. Always use these.

A profile_type=0 profile sets both ss and batch pulse fields — that's why one profile serves any volume, and why you never ask which mode up front.

Template heuristic — clone a Fellow drop

Fellow's Fellow/ folder has three roast baselines (plocal0-2) and drops/ has ~10 curated real-world recipes. Always clone one as the starting point rather than inventing params. Full decision tree + template table: references/profile-heuristics.md.

Quick overview:

Fetch params at invocation via coffee://profiles — don't hardcode them; IDs change.

Roast-delta adjust vs the template:

Keep the template's pulse pattern (flat vs descending) and count.

Delicate light naturals & Geishas

These over-extract and shed florals easily, and the drop-clone defaults lean the wrong way for them. For a Geisha or fruity light natural, override toward:

• Flat pulses + hot bloom — bloom ~96 °C, then flat pulses ~93.5–94 °C, not a descending ramp. Descending ramps tame a darker/soluble coffee's bitter tail; a dense light roast doesn't have one. Only descend if a brewed cup turns astringent.
• Ratio 16.0–16.5 — clarity-forward; the extra dilution opens the florals.
• Bloom 2–2.5× — Fellow's default is 2×; reserve 3× for very fresh, heavily degassing roasts (above ~3× actually lowers extraction). Don't reach for 3× by reflex.
• Grind finer — see the Geisha row under Ode Gen 2 grind.

Ode Gen 2 grind — derive two numbers

The profile is volume-agnostic, but the right Ode Gen 2 grind isn't: single-serve grinds finer than batch (larger dose → deeper bed → coarser to avoid over-extraction). So derive both, and put both in the title. Scale 1.0–11.2, lower = finer.

• Single-serve is one stable number across its whole 150–450 ml range (shallow bed).
• Batch climbs with carafe size — start at the low end of its range and go coarser as the carafe fills. There's a step (not a slope) at the basket change: nobody grinds the in-between number.
• Delicate light naturals / Geishas grind FINER than the table: Fellow's own natural-Geisha call is single-serve 3.1–4.2 (use ~4), batch 6–9 (source). Corroborated by owners (single-serve ~4–6, batch ~8–8.5). So a Geisha title is 4-8.

These are native Gen 2 numbers — do NOT translate from Gen 1. Prefer Fellow's first-party per-coffee call when the user has it.

Water (BWT magnesium-forward filter)

Leo's brew water is BWT magnesium-mineralised — calcium stripped, magnesium added, soft, near-neutral pH.

• Magnesium-forward water amplifies perceived sweetness, fruit and floral — it flatters fruity light naturals / Geishas.
• "Magnesium over-extracts" is marketing, not science. Do NOT grind coarser to compensate for the water.
• The real risk is low alkalinity → sharp / sour ("bright front, empty middle"), because softening strips buffer. If a cup reads sharp or thin, fix it with a weaker ratio (~1:17) or by adding buffer (toward GH 50–70 / KH ~40 ppm) — never by changing grind.

Title rule — lead with the grind span

The first token of the title is the grind span <SS>-<batch>, so you can read the grind straight off the device's profile list:

4-8 Intermission Geisha Colombia
│ └── batch grind  (climb from here as the carafe fills)
└──── single-serve grind

• Left number when you brew single-serve; start at the right number and climb for batch.
• ASCII hyphen only. The fellow_aiden library validates with TITLE_REGEX = [A-Za-z0-9 !@#$%&*\-+?/.,:)(]+ and fullmatch — so ~, →, –, ·, _, [], {}, ", ' are all rejected before the request leaves the machine. Of the range glyphs, only -, /, : are legal; use -.
• Max 50 chars. Format: <SS>-<batch> <Roaster> <Origin> or … <Coffee Name>.

Recipe card output

After create_profile, print this and stop (no checklist, no brew-now nudge — those happen at the machine):

Created: 4-8 Intermission Geisha Colombia

Grind (Ode 2):  4 single-serve  ·  6-9 batch (climb with size)
Ratio:          1 : 16.5
Bloom:          45 s @ 96 °C  (2.5×)
Pulses:         3 flat @ 94 °C
Template:       d110 Square Mile (Brazil natural) + Geisha overrides

Synced to Fellow cloud. Pick ml at the machine.

Mode boundary (informs the grind numbers only)

Fellow's basket split is single-serve 0–450 ml, batch 451 ml+. The single-serve basket physically changes to the flat batch basket at that line — that's the step in the grind, and the only reason the boundary matters here. (You still never ask the volume.)

Error handling

Out of scope

This skill creates the profile and stops. Handled elsewhere:

• Brewing & scheduling — the MCP brew_now and create_schedule tools. Reach them by saying "brew now"; the server does duration, timezone and READY-time math.
• Device-state checks / pre-brew checklist — do at the machine.
• Target volume / dose / grams — the device shows the dose once you pick ml.
• Tasting-note iteration, persistent journal — planned for a v2 journal context.

Red flags

• About to pass profile_type=1 with pulse fields → stop, use 0.
• About to use °F → stop, Fellow is °C only.
• Guessing origin or roast from a bag → stop, ask the user.
• About to ask for a target volume / pick a mode → stop, the profile is volume-agnostic; you only derive two grind numbers.
• About to read coffee://device, render a checklist, estimate duration, or call brew_now as part of /brew → stop, all out of scope; this skill ends at the recipe card.
• Skipping the coffee://profiles dedup check → stop, reuse before creating.
• About to finalize a profile without calling query_palate for a bean whose dimensions you know → stop, check palate history first and pre-correct (step 4).
• Inventing pulse temps / ratios from nothing → stop, clone a Fellow drop first.
• Using the table's "light" grind (5 SS) for a Geisha / delicate natural → stop, those grind finer — Fellow's own call is ~4 SS.
• Giving a light natural / Geisha a descending pulse ramp → stop, flat ~94 °C suits them.
• Translating Gen 1 grind numbers to Gen 2 → stop, the table is already Gen 2.
• Grinding coarser "because the water is magnesium-forward" → stop, unfounded.
• Naming a profile without the grind span as its first token, or with a non-ASCII glyph (~, →, ·, en-dash) → stop, lead with <SS>-<batch> using a plain hyphen.