Plan Card

Plan Card — SMS/MMS Plan Image + PLAN.md

Deterministic pipeline that turns handoff profile data into (a) a branded NanoRhino plan card PNG for MMS delivery and (b) PLAN.md markdown content — no LLM involved. The openclaw-infra Twilio extension invokes the script programmatically (planImage.scriptPath in its deploy config must point to plan-card/scripts/plan-to-image.py).

CLI Contract (frozen — do not change)

python3 {baseDir}/scripts/plan-to-image.py \
  --input  --output  [--width 1080] [--max-bytes 614400]

Input JSON:

{
  "profile": { "sex": "male|female", "age_years": N, "height_cm": N,
               "weight_kg": N, "goal_weight_kg": N|null,
               "intent": "lose|maintain|recomp|gain",
               "daily_steps": N|null,
               "activity_level": "sedentary|lightly_active|moderately_active|very_active|null" },
  "tdee":    { "recommended": N, "low": N, "high": N, "bmr": N },
  "locale":  { "country": "US", "units": "imperial|metric",
               "language": "en|zh|..." },
  "qr":      { "kind": "sms_claim", "number": "+19152777888",
               "body": "" }
}

qr (optional, additive — part of the cross-repo "email plan uses plan-card" feature): when present, the card renders a QR code in the timeline/"unlock" area encoding SMSTO::. Scanning it opens the phone's SMS app with the claim token pre-filled — the same action as the web SMS CTA. kind must be sms_claim; number/body are required non-empty strings (the body is the TDEE handoff token, e.g. Crockford-8 or profile slug). The QR caption is localized via STRINGS like everything else. When qr is absent, output is byte-identical to the pre-QR renderer (the qrcode dependency is only imported on the QR path), and the page height stays 3200 px; with qr the page grows to 3540 px to fit the block. SMSTO: is used instead of the RFC 5724 sms:?body= URI because iOS Camera and Android scanners both pre-fill number AND body from SMSTO, while sms: body pre-fill is inconsistent across platforms.

locale.language (optional, BCP-47-ish lowercase, e.g. en, zh, zh-tw) selects the output language; absent/unknown → en. The upstream authority is USER.md > Language — the Twilio extension reads it and passes it through. This script does NO language inference of its own (consistent with docs/CONVENTIONS.md §10).

See examples/sample-input-with-goalweight.json, examples/sample-input-without-goalweight.json, examples/sample-input-zh.json, and examples/sample-input-with-qr.json.

stdout on success (single JSON line): {"ok": true, "png": "", "bytes": N, "plan": {...}, "plan_markdown": "..."}

On failure: non-zero exit, {"ok": false, "error": "..."} on stdout, traceback on stderr.

The caller is responsible for writing plan_markdown to the agent workspace as PLAN.md.

Methodology (canonical sources)

Plan content follows user-onboarding-profile/SKILL.md > Step 3：生成并 确认减脂方案 exactly:

• A single daily calorie target — never a band.
• Daily calorie deficit (~XXX).
• Weekly loss rate.
• A single completion month + year.

Card-only macro override (approved): the Step-3 "no macros at plan stage" rule is explicitly overridden for the CARD — it additionally shows daily macro ranges, the per-meal rhythm, a sample day, and a merged first-week block (see Card content below). PLAN.md stays Step-3 compliant: no macros, no meal split there — macros still belong to Step 4 (diet-mode selection) for everything downstream.

Energy anchoring (product decision — Jason, 2026-06-10)

The handoff tdee.recommended / tdee.bmr (the web TDEE decomposition the user already saw and trusts) is the energy authority for EVERY path — 以 tdee 的结果为准. This overrides the earlier trust-forward-calc instruction: forward-calc re-derives BMR/TDEE from the profile and can disagree with the web number (we saw a 2,032 Cal target against a web TDEE of 2,010 — a "target" above the user's shown burn reads as broken). forward-calc is therefore NOT called at all.

weight-loss-planner/scripts/planner-calc.py (subprocess; interface unchanged) still provides the methodology on top of the anchored TDEE:

• Pace: Step-3 pace table via recommend-rate when goal_weight_kg

is present (25 kg → 0.7); 0.35 kg/wk default without a goal.

• Target: calorie-target → daily target = handoff

tdee.recommended − deficit.

• Floor: safety-floor → max(handoff BMR, 1000). If the target falls

below the floor it is clamped to it, and the effective rate + completion date are re-derived from the actual achievable deficit.

• Timeline: weeks = total loss ÷ (possibly clamped) rate; completion

date recomputed from that.

• BMI: bmi with --standard asian when the locale country/language

is Chinese/Japanese/Korean (CN/TW/HK/MO/SG/JP/KR or zh/ja/ko), who otherwise.

• Macros + rhythm: macro-targets from the anchored daily target

(mode balanced; high_protein for recomp/gain), including the canonical 30/40/30 allocation.

• Guarantee: for intent=lose the daily target is always strictly

below the handoff TDEE — enforced in code (degenerate handoff data where TDEE ≤ floor fails loudly with an error instead of rendering a broken card).

• The Step-3 "50 岁以上偏向下限" pace note is a conversational judgment

call and is not part of planner-calc's deterministic math; this pipeline uses the pace-table default rate.

Intent handling

• lose + goal: full plan as above (pace table, timeline, goal BMI).
• lose, goal_weight_kg null: 0.35 kg/wk default deficit (≈385

kcal); the card and PLAN.md show "Reply with your goal weight to unlock your completion date" instead of a date — still a single target number.

• maintain: target = handoff tdee.recommended (single number),

no deficit, no timeline.

• recomp: target = TDEE − 200, no timeline.
• gain: target = TDEE + ≈275 (0.25 kg/wk lean gain); timeline only

if a goal weight above current weight is provided.

• Activity missing: derived from daily_steps

(<5000 sedentary, <8000 lightly, <12000 moderately, else very active), else defaults to lightly_active (the weight-loss-planner default). (Activity only drives the focus rules and the PLAN.md description — energy comes from the handoff TDEE, never from an activity multiplier.)

Card content (in order)

1. Header — NanoRhino wordmark (CamelCase, single color — no accent

split) + date.

1. Title + profile line — weight (→ goal), height, age.
2. Hero — the single daily calorie target.
3. Plan tiles — daily deficit + weekly pace.
4. Daily macros (as RANGES) — min–max is the primary tile value

(e.g. Protein 90–120g); protein is visually emphasized with its floor as subtext, fat/carbs show the target as subtext. From planner-calc macro-targets computed on the anchored daily target (mode balanced, or high_protein for recomp/gain).

1. Daily rhythm — per-meal calorie split from planner-calc's

canonical 30/40/30 allocation (breakfast/lunch/dinner with ~Cal each).

1. Sample day — one concrete meal example per rhythm slot

(breakfast/lunch/dinner) + a snack-ideas line + the plate-formula footer, derived from references/meal-templates.md (selection rules below; pure rules, no LLM).

1. Timeline — completion month tile, or the goal-weight unlock box.

When the input carries qr, the SMSTO QR claim block (QR + short localized scan prompt) renders here, below the tile/box.

1. Your first week — ONE merged block (replaces the former separate

"Focus this week" + "Week 1 checkpoints"): the SMS cadence checkpoints plus the personalized rule-based items, deduplicated (the protein anchor appears once — the 🍗 checkpoint absorbs the old palm-sized item), capped at 5.

1. Footer — "Text me anytime" coach line.

Your-first-week items (deterministic, capped at 5)

| # | Item | Source | |---|---|---| | 1 | 📸 Log every meal — just text a photo | cadence | | 2 | ⚖️ Weigh in Wednesday & Saturday morning | cadence | | 3 | 🍗 Hit the protein target at every meal | cadence (absorbs the protein anchor — deduplicated) | | 4 | → Movement slot, first match: steps < 5,000 OR sedentary → 10-min walks after lunch/dinner; lightly/moderately_active → 2 short strength sessions; very/extremely_active → protect sleep (7+ h) | personalized rule | | 5 | → Water first — 2 liters a day | personalized rule |

Sample-day selection rules (deterministic)

Content source: references/meal-templates.md (verbatim reference from Jason). Breakfast is selected by per-meal protein need = protein target ÷ 3 (per the reference's 25–40 g/meal guidance):

| Per-meal protein need | en breakfast | zh breakfast | |---|---|---| | < 32 g | Greek yogurt + berries + granola (~30g) | 鸡蛋 2 个 + 无糖豆浆 + 全麦面包（约 25g） | | 32–37 g | 3 eggs + turkey sausage + toast (~35g) | 鸡蛋 3 个 + 牛奶 + 燕麦（约 30g） | | ≥ 38 g | Whey smoothie with banana + oats (~40g) | （zh 取最高档，同上） |

Lunch / dinner / snacks / plate formula are the reference's fixed best defaults (en: chicken + rice + broccoli; salmon or lean beef + sweet potato; yogurt/cottage-cheese/egg snacks; "protein 6 oz + complex carb 1 cup + vegetables unlimited + fat 1 tbsp"). zh uses Chinese-food equivalents with the same protein-forward structure and metric portions (鸡胸肉 + 米饭 + 西兰花; 清蒸鱼或瘦牛肉 + 红薯 + 时蔬; 搭配公式：蛋白质 150g + 主食 1 碗 + 蔬菜不限量 + 油脂 1 勺) — not literal translations.

Localization

Every user-facing string on the card and in plan_markdown comes from the STRINGS table in scripts/plan-to-image.py (full en + zh today; adding a language = adding a table entry). Language-aware conventions: zh uses 大卡, kg/周 pace, kg / 斤 in PLAN.md, and YYYY年M月 dates.

Rendering

templates/plan-card.html (1080-wide portrait, page height 3200 px — 3540 px when the qr block is present — inline CSS only, no external assets) → WeasyPrint HTML→PDF → PyMuPDF PDF→PNG at --width, downscaled until it fits --max-bytes (default 600 KB MMS budget; full-size renders are ~225–310 KB). The QR SVG (pure-python qrcode SVG factory) is embedded as a data-URI image — no PIL, no binary deps.

PLAN.md structure (plan_markdown)

Mirrors the Step-3 presentation so downstream skills work unchanged: user info block (height / weight / goal weight / age / sex / colloquial activity) → "Your Plan"/"你的计划" with the four elements → 1–2 sentence pace explanation (no TDEE/BMR jargon). No macros, no meal split — the card-only macro override does not apply to PLAN.md.

Localized (en/zh) like the card. The calorie-target line stays parseable in both languages (Daily calorie target: / 每日热量目标： — matched by weekly-report's existing regex), and a language-independent machine anchor is always emitted near the top:

Dependencies

pip install -r {baseDir}/requirements.txt (weasyprint, pymupdf, qrcode — the latter pure python, used only when the input carries qr). WeasyPrint also needs pango/cairo/gdk-pixbuf system libraries — on EC2 (Amazon Linux): sudo dnf install -y pango cairo gdk-pixbuf2; on Ubuntu: sudo apt-get install -y libpango-1.0-0 libpangocairo-1.0-0 libgdk-pixbuf-2.0-0.

For zh cards the host also needs a CJK font (e.g. google-noto-sans-cjk-fonts), and the week-1 checkpoint emoji need a color emoji font (google-noto-emoji-color-fonts); without them CJK/emoji glyphs render as boxes.

Data dependencies

| Resource | Direction | Mechanism | |---|---|---| | Handoff profile JSON (--input) | read | provided by openclaw-infra Twilio extension | | weight-loss-planner/scripts/planner-calc.py | read/execute | subprocess (owner: weight-loss-planner) | | Plan card PNG (--output) | write | this script | | PLAN.md content | emit | plan_markdown in stdout JSON; caller writes the file |

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: NanoRhino
• Source: NanoRhino/weight-loss-skill
• License: MIT
• Homepage: https://nanorhino.com/

Install and usage instructions live in the source repository linked above.