agentleFS
mhxiemhxie/Atelier/skills/dine/SKILL.md

dine

Skill5 starsChanged 35 days agoMITView on GitHub

Restaurant recommendation flow using local context and credit-burn signals.

The file

Show its rules and text

We split this skill into the rules it gives an agent

23 rules. About 0 are common, 11 are distinctive, and 12 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈11

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 12

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    D. Establishment Update: update one physical branch's address or lifecycle without creating a meal-history row.

    line 11
  2. Distinctive:

    If args present, parse them as initial filters.

    line 37
  3. Very specific:

    Read profile/diet.md once, if present, to resolve workplace slugs, catering folders, and ## Catalog files.

    line 43
  4. Very specific:

    Do NOT route to C if the message is forward-looking: contains 推荐 / 去哪吃 / 想吃 / what should I eat / where to eat / 今晚吃什么 → fall through to A.

    line 59
  5. Distinctive:

    Missing-file fallback: if any of these is absent, skip it silently and note the gap in the closing line ("scored without [missing source]").

    line 93
  6. Very specific:

    Log-side aggregates (deterministic): run uv run scripts/dine_rank.py --avoid-days <window> first.

    line 95
  7. Very specific:

    Do not re-read the meal-history tracker row by row.

    line 99
  8. Very specific:

    Treat (餐厅, 分店) as establishment identity.

    line 116
  9. Distinctive:

    If filter returned <3 candidates, note relaxation taken (e.g., "loosened distance to 🚗🚗")

    line 150
  10. Distinctive:

    Do NOT auto-book; just surface candidates.

    line 157
  11. Very specific:

    If an arg is a .pdf path → use it directly.

    line 163
  12. Very specific:

    Read the PDF (Read tool). Extract per-day sections (Mon/Tue/Wed/Thu/Fri).

    line 170
  13. Very specific:

    Read attendance pattern from profile/diet.md (the section matching the resolved <slug>, key: Attendance days).

    line 174
  14. Very specific:

    If profile/diet.md is absent or has no entry for <slug> → ask the user once, do not assume a default.

    line 178
  15. Very specific:

    Read dietary picking priorities and flag taxonomy from profile/diet.md (the <slug> section).

    line 182
  16. Very specific:

    Add a 1-2 line cross-day note if profile/diet.md defines cross-day rules (e.g., protein rotation, 油脂 balance).

    line 198
  17. Distinctive:

    Use this path for address or lifecycle maintenance without inventing a visit.

    line 223
  18. Very specific:

    Run python3 scripts/dining_audit.py --json.

    line 229
  19. Distinctive:

    Do not add a meal-history row, change ratings, or infer that another branch shares this lifecycle.

    line 231
  20. Distinctive:

    Read-only on catalog docs (under Intent A): do NOT modify the regional dining catalog or the credit-perks catalog when handling a …

    line 236
  21. Distinctive:

    Always show credit-burn opportunity: if relevant under the live private benefits tracker.

    line 238
  22. Distinctive:

    Read-only on the PDF: never modify the catering PDF

    line 244
  23. Distinctive:

    Per-day skip on parse failure: if any one day's section fails to parse, skip that day with a logged warning.

    line 247

≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Purpose

Four intents (auto-detected from args):

  • A. Restaurant Recommendation (default): pick 3 restaurant candidates based on user-supplied context, historical preferences, and credit-burn opportunities. Read-only on catalog docs under this intent.
  • B. Workplace Catering Tracker: parse a weekly catering PDF from the folder mapped to a workplace slug in profile/diet.md, choose health-aware picks for the user's attendance days, and surface a confirmed table for the user to record themselves (the system does not write to daily notes).
  • C. Meal Log Capture (ad-hoc): log a meal the user just ate. Parses receipt images (HEIC/JPG/PNG/PDF) when provided, cross-references catalogs for missing slots, asks ONE compact question for what cannot be derived, shows a draft row + side-effect plan, and appends to the private meal-history tracker on confirm. An explicitly trip-associated capture may also add a date-only meal-history reference to that trip note. Co-equal capture path with /hi Dining Pulse.
  • D. Establishment Update: update one physical branch's address or lifecycle without creating a meal-history row.

Quick start

Intent A examples:

  • /dine → ask all context
  • /dine 工作日午餐 → use as scene hint, ask remaining
  • /dine 朋友 4 人 川菜 dinner → use as filters, ask remaining
  • /dine <city> burn credit → location hint + flag credit-burn priority

Intent B examples (first arg = workplace slug mapped in gitignored profile/diet.md):

  • /dine <slug> → find the latest PDF in the mapped catering folder covering this week, pick per profile/diet.md attendance pattern
  • /dine <slug> <pdf-path> → explicit PDF
  • /dine <slug> all → all 5 weekdays (override)
  • /dine <slug> <day-codes> → custom attendance set (override; any day-code combination works)

Intent C examples (logging a meal you just ate):

  • /dine log 今天午饭吃的 <restaurant> 两人 $<amount> → free-text meal report
  • /dine 今天晚饭吃的是 /path/to/receipt.heic → receipt image (HEIC auto-converted before Read)
  • /dine 昨天 <restaurant> dinner $<amount> → dated free text (respects late-sleep rule)
  • /dine log /path/to/receipt.pdf → receipt PDF outside any catering folder

Intent D examples (maintaining a physical branch):

  • /dine status <restaurant> <branch> closed
  • /dine status <restaurant> <branch> moved <new-address>

If args present, parse them as initial filters; only ask for slots not derivable.

Step 0: Intent detection

Parse args. Precedence: D → B → C → A (most specific match wins; ambiguous → ask user one line before routing).

Read profile/diet.md once, if present, to resolve workplace slugs, catering folders, and ## Catalog files. Do not assume the private mappings from examples in this command.

Route to Intent D when the leading subcommand is status.

Else route to Intent B if any of:

  • First arg matches a workplace Slug declared in profile/diet.md
  • Any arg is a .pdf path under a catering Folder declared in profile/diet.md
  • Args contain the literal token catering

Else route to Intent C if any of:

  • Leading subcommand log (e.g., /dine log <freetext>)
  • Args contain past-tense / reporting markers: 记录 / log / 吃了 / 吃的 / 刚吃完 / "今天X吃的" / "昨天Y" / "just had"
  • An image path (.heic / .jpg / .jpeg / .png) is provided
  • A .pdf path is provided outside any $OV/*/catering/ folder
  • Free text mentions a specific restaurant + amount/party (e.g., <name> 两人 $<amount>) without a forward-looking verb

Do NOT route to C if the message is forward-looking: contains 推荐 / 去哪吃 / 想吃 / what should I eat / where to eat / 今晚吃什么 → fall through to A.

Otherwise route to Intent A (continue to Step 1 below).

For Intent B, jump to the "Intent B: Workplace Catering Tracker" section near the bottom and skip Steps 1-5. For Intent C, read protocols/dining-capture.md and follow it; skip Steps 1-5. For Intent D, jump to the "Intent D: Establishment Update" section near the bottom and skip Steps 1-5.

Step 1: Gather context

For missing slots, ask via AskUserQuestion or sequential 1-line prompts (whichever fits faster). Required slots first; optional slots only if useful.

SlotOptionsRequired
LocationHome region / nearby city / travel destination / otherY
PartySolo / Partner / Family (N) / Friends (N) / Mixed workY
MealLunch / Dinner / Brunch / Late nightY
Time budgetQuick (<30min) / Standard (1-2h) / Leisurely (2h+)Y
Mood / cuisineFree text or options declared in profile/diet.mdN (default any)
Health filteroptions enumerated in profile/diet.md ("Health filter input options" section) plus no preferenceN (default no preference)
Budget capUser-supplied cap / no capN
Avoid recentLast 30 / 60 / 90 daysN (default 30)

Step 2: Load data (parallel)

Resolve the following roles through profile/diet.md § Catalog files. Paths are relative to $OV unless absolute. If the profile or a mapping is absent, use structural discovery under <paths.travel>/ and <paths.finance>/ as a fallback and disclose the gap.

  • Regional dining catalog (rotation + Michelin wishlist + 场景索引 + 门店索引), under <paths.travel>/
  • Meal-history tracker (history with 评分 + 再去 + recency), under <paths.travel>/
  • Credit-perks dining catalog (eligibility + city catalogs), under <paths.travel>/
  • Benefits tracker (current cycle credit status, for burn signal), under <paths.finance>/
  • Prepaid-balance tracker (balances per restaurant, for a soft "use it" signal), under <paths.finance>/
  • For travel destinations: use the matching city section plus the corresponding local guide under <paths.archive>/practical/travel/

Missing-file fallback: if any of these is absent, skip it silently and note the gap in the closing line ("scored without [missing source]"). The recommendation still produces; the user can decide whether to recreate the catalog.

Log-side aggregates (deterministic): run uv run scripts/dine_rank.py --avoid-days <window> first; it returns per-restaurant visit counts, last-visit recency, 评分 averages, 再去 state, the log-derived score component, and the avoid-window exclusions. Do not re-read the meal-history tracker row by row; combine the returned log_score with the catalog-side factors below.

Integrity preflight: when running from the Atelier repo, execute python3 scripts/dining_audit.py --json. Use its establishments rows as the address/lifecycle source. If it reports errors, exclude the affected rows or source from scoring and disclose the degraded input. Warnings such as missing legacy values do not block recommendations.

Step 3: Filter + score

Hard filters (eliminate non-matches):

  • Location matches user's region
  • Cuisine NOT in avoid list
  • Estimated price ≤ budget cap (allow 20% margin)
  • Drive time fits time budget (heuristic: 🚗 count × 15min one-way)
  • For Quick lunch: ⌛ ≤ 1
  • For "Special occasion": Michelin OR Exclusive Tables only
  • Skip restaurants visited within avoid recent window (from the meal-history tracker)
  • Skip physical branches whose lifecycle is closed or moved; allow unknown only with an explicit lifecycle warning

Treat (餐厅, 分店) as establishment identity. Join visit ratings only to an exact branch-specific restaurant name; never assign a chain-wide or legacy ambiguous score to a branch by inference.

Soft ranking (order the survivors). dine_rank.py owns the log-side score (log_score, itemized in log_score_parts: 评分 averages, 再去, rusty recency); take it verbatim. Then apply the catalog-side priorities, strongest first:

  1. Credit-burn priority: an Exclusive Tables restaurant whose relevant cycle has unused credit and ≤ 60 days to deadline outranks everything else.
  2. Michelin star match when the mood is "Special occasion".
  3. Catalog 评 (higher first) and a 场景索引 match in the regional catalog.
  4. Novelty: never visited when the mood is "Surprise" or "探索"; old-favorite revisit when rotation 评 ≥ 2 and days_since > 60 from dine_rank.
  5. Health filter active: apply the "Health-filter scoring rules" section of profile/diet.md (recent-visit penalties, clean-style bonuses, cumulative-load adjustments) as exclusions and tie-breakers.

Exact weights, if wanted, belong in dine_rank.py beside log_score.

Step 4: Output

Top 3 candidates as a table:

| # | 餐厅 | 类型 | $ | 距离·等待 | Why | Credit signal |
|---|---|---|---|---|---|---|
| 1 | <restaurant-A> | <cuisine> | <$range> | <distance·wait> | <reason from catalog/log: 评 N + scene fit + recency> | n/a |
| 2 | <restaurant-B> | <cuisine> ⭐ | <$range> | <distance·wait> | <reason: Michelin tier + last log rating + want-revisit>; **<credit-card> <perk-program> <half> deadline <MM/DD> ($<amount>)** | 🔥 burn |
| 3 | <restaurant-C> | <cuisine> ⭐ | <$range> | <distance·wait> | <reason: novelty + Michelin tier + perk-eligible>; <perk-program> 候选 | <credit-card> <half> ✓ available |

Brief reasoning paragraph (2-3 lines) below the table:

  • Mention the top filter constraints applied
  • Flag any credit-burn 紧迫性 in plain text
  • If filter returned <3 candidates, note relaxation taken (e.g., "loosened distance to 🚗🚗")

Step 5: Close

End with one line:

"选哪个? (回 1/2/3) 我帮你在常用 booking platforms 查时段, 或者 /dine + 新约束 重排"

Do NOT auto-book; just surface candidates.

Intent B: Workplace Catering Tracker

B.1 Resolve PDF

  • If an arg is a .pdf path → use it directly.
  • Else: resolve the slug's Folder from profile/diet.md, list <Folder>/*.pdf, and pick the one whose filename date range covers the current calendar week. Typical filename pattern: <Workplace> Catering_<Mon> <DD>-<Mon> <DD>.pdf. If multiple match (e.g., manual override), prefer the most recent mtime.
  • Optional date arg YYYY-MM-DD shifts the target week (Mon of that week).
  • 0 matches: report 本周菜单还没传到 <configured folder> and exit cleanly.

B.2 Parse menu

Read the PDF (Read tool). Extract per-day sections (Mon/Tue/Wed/Thu/Fri). Each day has a theme + items + dietary tags (v / vg / mwgci).

B.3 Determine attendance days

Read attendance pattern from profile/diet.md (the section matching the resolved <slug>, key: Attendance days). Override via the second CLI arg:

  • all → all 5 weekdays present in the PDF
  • <day-codes> → custom set (case-insensitive day codes; any combination)

If profile/diet.md is absent or has no entry for <slug> → ask the user once, do not assume a default. Map each chosen day code to an absolute date based on the resolved week.

B.4 Pick per day (reuse Step 3 health-filter logic)

Read dietary picking priorities and flag taxonomy from profile/diet.md (the <slug> section). Apply the policy verbatim — do not bake personal preferences into this committed file.

Generic fallback when profile/diet.md is absent: choose ONE protein + 1-2 veg sides per day, no specific oil/protein bias, and ask the user to confirm the picks before presenting.

The skill itself enforces only the structural shape (one row per attendance day, columns: protein + veg + sauce-note + flag). The semantic content is policy from the private file.

B.5 Preview

Show table (one row per attendance day; values fill from B.4):

| Date | Day | Theme | Pick | Flag |
|---|---|---|---|---|
| YYYY-MM-DD | <day> | <menu theme> | <protein> + <veg sides> + <sauce/dressing note> | <flag from profile/diet.md taxonomy> |

Add a 1-2 line cross-day note if profile/diet.md defines cross-day rules (e.g., protein rotation, 油脂 balance). Otherwise omit.

B.6 Present

Show the user the per-day picks as ready-to-paste lines so they can record them in their daily notes themselves. The system does not write to daily notes.

For each attendance day, output one line in the format: <Slug> <YYYY-MM-DD> — <Day> <theme> (<pick>, <flag>) where <Slug> is the user-provided slug capitalized (first letter only).

B.7 Report

/dine <slug> summary (<week-range>)
  picks:  N   (date list)

Intent C: Meal Log Capture

protocols/dining-capture.md owns this path: source resolution, catalog cross-checks, explicit trip context, auto-derivation, the side-effect plan, the confirmation gate, and the write and report steps. Read it and follow it. It is shared with /hi Dining Pulse, which is why it is not inline here.

Intent D: Establishment Update

Use this path for address or lifecycle maintenance without inventing a visit.

  1. Resolve the regional catalog from profile/diet.md, then find an exact (餐厅, 分店) row in 门店索引.
  2. Accept only lifecycle values active, closed, moved, or unknown. Require a complete street address for a new row. Preserve an existing address unless the user explicitly changes it.
  3. Show the current row and proposed row, then ask one confirmation: Apply establishment update? (yes / no / edit).
  4. On yes, update or append exactly one registry row, set 核验日 to the local effective date, record the source without copying private prose, and bump the catalog's Last updated: line.
  5. Run python3 scripts/dining_audit.py --json. Repair only the introduced row if validation fails; otherwise report one line: Updated: <restaurant> (<branch>) → <status>.

Do not add a meal-history row, change ratings, or infer that another branch shares this lifecycle.

Rules

Intent A:

  • Read-only on catalog docs (under Intent A): do NOT modify the regional dining catalog or the credit-perks catalog when handling a recommendation request. The meal-history tracker is also read-only under Intent A; appends route through Intent C (or /hi Dining Pulse).
  • 0 candidates after hard filter: relax most-restrictive constraint by 1 step, retry; surface 1-2 closest matches with flag "relaxed: <constraint>"
  • Always show credit-burn opportunity if relevant under the live private benefits tracker. Even if the eligible restaurant does not match the exact mood, surface a fourth line using only the currently relevant benefit details.
  • Match user language: Chinese-dominant if cuisine is Chinese; English if Western
  • Shape: the candidate table, a short reasoning paragraph, one close line. Nothing else.
  • No web search: cuisine + restaurant data comes from local catalog files only

Intent B:

  • Read-only on the PDF: never modify the catering PDF
  • Read-only on daily notes: daily notes are user-authored; the system surfaces picks for the user to record themselves and never writes to <paths.daily_notes>/
  • Does not touch the meal-history tracker: workplace catering is excluded by design (low signal density per memory)
  • Per-day skip on parse failure: if any one day's section fails to parse, skip that day with a logged warning; do not abort the whole batch
  • No web search: menu data comes from the PDF only

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.