wa-campaign-premise-find

name	wa-campaign-premise-find
version	1.0.0
description	Find a specific Your Project campaign by premise — the trope/story the user remembers ("female resurrected OP person", "demon lord reincarnation", "where I was a swords bard in Thay", etc.). Searches BOTH Firestore (live data, paraphrased descriptions) AND the LLM wiki raw transcripts (full God Mode prompts + first scenes — where the user's own words actually live). Returns campaign_id + title + character + Firestore URL + verbatim God Mode excerpt.
when_to_use	Use when the user says: "find the campaign where I was [X]", "which campaign had [trope]", "I want to revisit the [character type] campaign", "find my [isekai/reincarnation/OP] campaign", "show me campaigns with [premise]", "do I have a campaign where [premise]". Also fires for "I think the campaign might be in the LLM wiki" (the dual-store decision rule).
allowed-tools	["terminal","file","memory"]
context	inline

wa-campaign-premise-find — premise-driven WA campaign lookup

The dual-store problem (this is the whole reason this skill exists)

WA campaign premises are duplicated across Firestore AND the LLM wiki, but they live in DIFFERENT forms:

Store	Path	What you find
Firestore	users/{uid}/campaigns/{cid}/story/{entry_id}	LLM-rewritten description blocks for first ~5 scenes — premise PARAPHRASED
Wiki raw transcripts	~/llm_wiki/raw/campaigns/<cid>/<Title>_<cid>.md / .txt	Full God Mode prompt + first scenes — user's ACTUAL words
Wiki source summaries	~/llm_wiki/wiki/sources/<slug>.md	Curated one-paragraph summary per ingested campaign
Wiki catalog index	~/llm_wiki/wiki/index.md	One-line entry per campaign — fastest signal

Firestore alone returns 14+ false positives for vague premises (every campaign has "reincarnation" or "demon" in its LLM-generated description even when that's not the user's premise). The wiki transcripts are where the user's own words live — the God Mode prompt and Scene 1-3 player input are unedited.

Phases — 5 phases total, run in sequence

Phase 1 — Confirm environment (one-time, ~10 sec)

ls -1 ~/serviceAccountKey.json \
  && python3 -c "import firebase_admin; print('firestore ok')" \
  && ls -1d ~/llm_wiki/raw/campaigns/ 2>/dev/null | head -1

If serviceAccountKey.json missing → fall back to wiki-only mode (Phases 3-4 only).

Phase 2 — Firestore scan (parallel with Phase 3)

import firebase_admin
from firebase_admin import credentials, firestore

cred = credentials.Certificate(os.path.expanduser("~/serviceAccountKey.json"))
firebase_admin.initialize_app(cred)
db = firestore.client()

UID = "vnLp2G3m21PJL6kxcuAqmWSOtm73"  # $USER's UID; verified 2026-07-28

# Scan all campaigns
camp_list = list(db.collection("users").document(UID).collection("campaigns").stream())

# For each, fetch the description block from the first 5 story entries
for c in camp_list:
    cid = c.id
    cd = c.to_dict() or {}
    title = cd.get("title") or cd.get("name") or "Untitled"
    # Get first 5 story entries — these have the LLM-generated description
    story = (db.collection("users").document(UID)
             .collection("campaigns").document(cid)
             .collection("story")
             .order_by("timestamp")
             .limit(5).stream())
    text = " ".join((s.to_dict() or {}).get("text", "") for s in story)
    # Score against user keywords (simple TF-IDF or keyword overlap)

CRITICAL: Firestore description blocks are LLM-generated paraphrases. Don't trust them as authoritative — use Phase 3 wiki transcripts as ground truth.

Phase 3 — Wiki fan-out (the actual ground truth)

# 1. Catalog index — fastest, curated one-liners
grep -i -E "<user-trope-keywords>" ~/llm_wiki/wiki/index.md

# 2. Source summaries — paraphrased by the ingest pipeline
grep -lir -E "<user-trope-keywords>" ~/llm_wiki/wiki/sources/

# 3. RAW TRANSCRIPTS — God Mode prompt lives here, user's own words
grep -lir -E "<user-trope-keywords>" ~/llm_wiki/raw/campaigns/

# 4. Concepts + entities — cross-references
grep -lir -E "<user-trope-keywords>" ~/llm_wiki/wiki/concepts/ ~/llm_wiki/wiki/entities/

Keyword expansion matrix (run multiple sets in ONE grep with | alternation):

User phrasing	Wiki regex
"demon lord" / "demon king"	demon[- ]?lord|demon[- ]?king|demonlord|overlord|dark lord
"reincarnation" / "reborn"	reincarnat|reborn|past[- ]life|previous[- ]life|isekai
"isekai" / "anime reborn"	isekai|anime[- ]?style[- ]?reborn|reborn in a new world
"OP" / "overpowered"	\bOP\b|over.?powered|god.?tier|level\s*[2-9]\d+|broken build|godlike
"resurrected" / "back from dead"	resurrect|return(ed)? from (the )?dead|back from the dead|raised (from the dead|back)
"female PC" / "make me female"	(make me|i am|let's make me).{0,30}(female|woman|girl|lady|witch|sorceress|goddess|succubus|demoness|queen|empress|princess)
"daughter from past life"	daughter.*past|child.*past|your daughter survives|my child.*past|my daughter.*previous life

Phase 4 — Hand-classify candidates (read God Mode openings)

For each top-N wiki match, read the FIRST 4000 chars of the .md or .txt file (God Mode block + Scene 1 + part of Scene 2). Look for the user's SPECIFIC premise phrases verbatim. Rank by how closely the literal text matches what the user described, not by frequency of trope words.

Phase 5 — Cross-reference back to Firestore (confirmation + URL)

For each wiki-confirmed match, fetch the Firestore doc to confirm it exists and grab the live campaign title + character name + entry count + last-played date.

Firestore URL pattern:

https://console.firebase.google.com/project/worldarchitecture-ai/firestore/data/users/{UID}/campaigns/{cid}

Output format

## Confirmed matches

### [title] — campaign_id=[cid]
- **Premise (God Mode verbatim):** "[first 300 chars]"
- **Character:** [name], [age], [class], [level start]
- **Setting:** [setting text]
- **Firestore URL:** https://console.firebase.google.com/project/worldarchitecture-ai/firestore/data/users/{UID}/campaigns/{cid}
- **Entries:** [N]
- **Match confidence:** [STRONG | PARTIAL | SPECULATIVE]

### [title] — campaign_id=[cid]
... (repeat)

## Disqualified candidates (looked promising, but God Mode didn't actually describe user's premise)

- **Aristocrat reborn V2** (cid=B1KlCh8DtmgcCqvp1nOs) — "Sylphina reincarnated as 7th daughter of a margrave"
  ≠ user's "demon lord reincarnation + investigating daughter from past life"

Contract

When this skill is invoked, the agent MUST:

1. Run BOTH Phase 2 (Firestore) AND Phase 3 (wiki fan-out) for any premise query with 2+ signal words. Do NOT trust Firestore alone — it returns 14+ false positives.
2. Hand-classify candidates by reading the first 4000 chars of each wiki match's primary file (God Mode prompt + Scene 1). Do NOT just count keyword frequency.
3. Output format: each match MUST have campaign_id, title, character, Firestore URL, and a verbatim God Mode excerpt (≥200 chars). The URL MUST be the clickable Firestore URL, never bare #123 or cid=....
4. Honest gap reporting: if NO campaign matches all stated signals exactly, say so explicitly and list the partial matches. Do NOT fabricate a "resurrected" angle that isn't there.
5. De-duplicate by campaign_id, not by title (Aizen, Nocturne, Nocturna each have 5-20 variants).

Output Format

A successful wa-campaign-premise-find invocation produces this shape:

{
  "user_query": "<the user's premise text>",
  "uid": "vnLp2G3m21PJL6kxcuAqmWSOtm73",
  "matches": [
    {
      "campaign_id": "dUfl4Adb3oH6foczNFSZ",
      "title": "Iseki v1",
      "character": "Renjiro (16yo Sorcerer, House Vane 5th son)",
      "setting": "Kingdom of Solis (Forgotten Realms, 1492 DR)",
      "firestore_url": "https://console.firebase.google.com/project/worldarchitecture-ai/firestore/data/users/vnLp2G3m21PJL6kxcuAqmWSOtm73/campaigns/dUfl4Adb3oH6foczNFSZ",
      "wiki_path": "~/llm_wiki/raw/campaigns/dUfl4Adb3oH6foczNFSZ/Iseki v1_dUfl4Adb.txt",
      "entries": 112,
      "last_played": "2026-06-27",
      "god_mode_excerpt": "Let's make me a reincarnation of a great demon lord...",
      "match_confidence": 
       
         
         
         
         
      
    
  
   
    
       
       
       
    
  
   

When posting in Slack, render the matches as a markdown table with columns [Campaign] | [CID] | [Character] | [Entries] | [Confidence] | [Firestore URL]. The URL MUST be a clickable https://... — never bare #123 or cid=....

Pitfalls (read these before running — most match the user's intent vs not)

1. The user's words live in wiki transcripts, NOT Firestore descriptions. Firestore description blocks are LLM-rewritten paraphrases. Always read the raw God Mode prompt as ground truth.
2. Female-PC disambiguation: many BG3-derived campaigns use female PC names (Shadowheart, Nocturne, Luna, Aizen — yes, Aizen Sosuke is rendered female in user's WA playthroughs, despite Bleach canon) but the original is male. Always check Character: field AND look for she/her/female markers in the description — do NOT assume gender from the character name.
3. Resurrected NPC ≠ resurrected PC. In "Luna post bg3" the player character Luna is female, but "Shadowheart (Resurrected)" is an NPC companion — NOT a match for "I am a resurrected..."
4. Re:Zero "Return by Death" = reincarnation loop. The PC doesn't literally die and come back; they loop on death. Counts as "reincarnation" trope but not "resurrected" — be precise.
5. "Female resurrected OP" is rare. Of 230+ campaigns, ZERO match all three exactly. Most match female + OP only. Report this honestly — don't invent a "resurrected" angle that isn't there. If the user asked for all three and we can't find any, the most likely answer is "No exact match — closest candidates are female + OP only, here's the inventory".
6. Re-running with different tropes requires re-running ALL phases. Don't cache prior results across tropes; the keyword sets are different.
7. Per-campaign de-duplication: Aizen, Nocturne, Nocturna, Saita each have 5-20 campaign variants (versions, copies, repro tests). Always present the canonical/original first.
8. Aizen Sosuke in WA = female PC. User's Aizen Sosuke campaigns are all rendered female despite Bleach canon being male. Don't apply canon lore to WA playthroughs.

Tests

• tests/test_dual_store_decision_rule.py — verify "wiki hint" triggers Phase 3 wiki fan-out
• tests/test_keyword_expansion_matrix.py — verify each user phrase maps to a working regex
• tests/test_female_pc_disambiguation.py — Aizen/NocSosuke must be classified female (WA canon) even though Bleach/BG3 canon say male
• tests/test_resurrected_npc_vs_pc.py — Luna post bg3 must NOT match "I am resurrected" (the resurrected character is Shadowheart NPC, not Luna PC)
• tests/test_god_mode_is_ground_truth.py — read raw .md file's God Mode block; verify text matches user's premise word-for-word

Cross-references

• memory-search/SKILL.md § "Pitfall — WorldArchitect campaign search by premise lives in TWO stores"
• memory-search/references/wa-campaign-premise-search.md — same pitfall, prior worked example (Iseki v1)
• download-campaign/SKILL.md — populates the wiki raw transcripts and source summaries
• wa-campaign-content-analysis/SKILL.md — analyzes content within a single campaign (Phase 1+ uses this skill's candidate list)