Verdad Heuristics Updater

VERDAD Heuristics Updater

Critical Configuration

• Supabase Project ID: dzujjhzgzguciwryzwlx — use this for ALL Supabase MCP calls (execute_sql, list_tables, etc.)
• Supabase Project Name: VERDAD
• Supabase URL: https://dzujjhzgzguciwryzwlx.supabase.co
• Repo path: Look for the prompts/ directory in the current working directory, or check /Users/j/GitHub/verdad/

On Invocation

When this skill triggers, immediately:

1. Determine your environment (Claude Code with file access? Claude.ai with only Supabase MCP? Cowork?)
2. Ask the user what they want to add — a Google Doc URL, raw notes, or a topic idea
3. Read the format templates in templates/ before generating anything

Architecture

The VERDAD pipeline has two stages that use heuristics, each optimized differently:

• Stage 1 (fast flagging): Scans transcriptions to flag potential disinformation. Heuristics are concise — just keywords, detection patterns, and examples. No subcategories in the main prompt.
• Stage 3 (deep analysis): Evaluates flagged snippets with web search verification. Heuristics are detailed — includes evidence grounding, cultural context, legitimate discussions, red flags, and full subcategories.

Each stage has two files: a user prompt (imported to Supabase DB, used at runtime) and a standalone heuristics.md (reference file, not imported).

File Locations (relative to verdad repo root)

Insertion Points

• Stage 1 detection_user_prompt.md: Before ## Additional Instructions
• Stage 1 heuristics.md: After the last category's ---
• Stage 3 analysis_prompt.md: Before ### **Additional Instructions**
• Stage 3 heuristics.md: After the last category's ---

Workflow

Step 1: Ingest Source Material

Accept input in any form:

• Google Doc URL: Read via gws docs CLI or Google Docs MCP. Extract text content, strip formatting artifacts and URLs.
• Raw text: Pasted notes, TikTok examples, field observations, radio transcript excerpts.
• Seed idea: Even just a topic name like "dental health misinformation" — research and generate comprehensive heuristics from domain knowledge.

Step 2: Check Current State

Find the current highest category number so you assign the next one correctly:

grep "### \*\*[0-9]" prompts/stage_1/main/heuristics.md | tail -1

Or via Supabase MCP (project ID dzujjhzgzguciwryzwlx):

SELECT substring(user_prompt from '### \*\*(\d+)\.' ) as last_cat
FROM prompt_versions
WHERE stage = 'stage_1' AND sub_stage = 'disinformation_detection' AND is_active = true;

Step 3: Generate Heuristics

Read the format templates before writing:

• templates/stage1_format.md — concise format for fast detection
• templates/stage3_format.md — detailed format for deep analysis

Generate both variants of each new category simultaneously — they cover the same content but at different levels of detail.

Content Requirements

Every new category needs:

1. Bilingual content — Spanish AND Arabic keywords, phrases, and examples throughout. Source material is often Spanish-only; generate Arabic parallels for every item.
2. Subcategories — Break the category into specific subtopics (typically 5-15). Each gets its own narratives, evidence, and examples.
3. Evidence grounding (Stage 3) — Brief medical/scientific consensus for each subcategory explaining why these claims are disinformation.
4. Cultural context (Stage 3) — How this disinformation spreads differently in Spanish-speaking vs Arabic-speaking communities (information channels, traditional practices, trust factors).
5. Legitimate discussions (Stage 3) — Topics that look like disinformation but aren't, to reduce false positives.
6. Detection tactics (Stage 3) — Meta-patterns that signal disinformation regardless of topic: commercial incentives ("link en bio"), censorship claims ("me censuran"), platform migration ("grupo de Telegram"), performative authority ("doctor dice").
7. Red flags (Stage 3) — Split into "immediate high-confidence" vs "requires careful analysis."

Adapting Source Material

Raw input needs these transformations:

• Strip links and dates — Remove verdad.app URLs, TikTok links, timestamps. Use representative examples that capture patterns.
• Restructure — Convert flat notes into the structured format (Description → Keywords → Heuristics → Examples for Stage 1; full template for Stage 3).
• Add cross-category notes — Note overlaps with existing categories (especially 3/COVID, 6/Abortion, 11/Healthcare, 14/Conspiracies).

Step 4: Update Files

Insert into all four files, matching each file's format. When you have file access (Claude Code / Cowork), use the Edit tool to insert at the documented insertion points.

When you DON'T have file access (Claude.ai), construct the updated user_prompt content by:

1. Fetching the current prompt from the database
2. Finding the insertion point in the text
3. Inserting the new category content
4. Deploying the modified prompt back to the database

Step 5: Deploy to Database

The pipeline reads prompts from the Supabase prompt_versions table at runtime.

Supabase project ID: dzujjhzgzguciwryzwlx

Check current versions

SELECT id, stage, sub_stage, version, is_active, length(user_prompt) as chars
FROM prompt_versions WHERE is_active = true ORDER BY stage, sub_stage;

Create new version (copy current, insert as inactive)

INSERT INTO prompt_versions (stage, sub_stage, version, description, created_by,
  system_instruction, user_prompt, output_schema, is_active)
SELECT stage, sub_stage, '{VERSION}', '{DESCRIPTION}', 'heuristics_updater',
  system_instruction, user_prompt, output_schema, false
FROM prompt_versions WHERE id = '{ACTIVE_ID}'
RETURNING id;

Update the user_prompt with new content

The prompt content is too large (30-125KB) to pass inline via SQL. Use this workaround:

1. Create a temporary SECURITY DEFINER function:

CREATE OR REPLACE FUNCTION public.temp_import_prompt(p_id uuid, p_user_prompt text, p_token text)
RETURNS jsonb LANGUAGE plpgsql SECURITY DEFINER AS $$
BEGIN
  IF p_token != '{RANDOM_TOKEN}' THEN RAISE EXCEPTION 'Invalid token'; END IF;
  UPDATE prompt_versions SET user_prompt = p_user_prompt WHERE id = p_id;
  RETURN jsonb_build_object('success', true, 'id', p_id);
END; $$;
GRANT EXECUTE ON FUNCTION public.temp_import_prompt TO anon;

Generate a unique token per session (e.g., verdad_import_ + random hex).

2. POST the content via REST API:

import requests
requests.post(
    "https://dzujjhzgzguciwryzwlx.supabase.co/rest/v1/rpc/temp_import_prompt",
    headers={"apikey": "{ANON_KEY}", "Content-Type": "application/json"},
    json={"p_id": "{ROW_ID}", "p_user_prompt": updated_content, "p_token": "{RANDOM_TOKEN}"}
)

3. Activate and clean up:

UPDATE prompt_versions SET is_active = false WHERE id = '{OLD_ID}';
UPDATE prompt_versions SET is_active = true WHERE id = '{NEW_ID}';
DROP FUNCTION IF EXISTS public.temp_import_prompt;

4. Verify:

SELECT id, stage, version, is_active,
  user_prompt LIKE '%{CATEGORY_NAME}%' as has_new_category
FROM prompt_versions WHERE is_active = true ORDER BY stage;

Repeat for both Stage 1 (disinformation_detection) and Stage 3.

Step 6: Push to GitHub (if file access available)

git add prompts/stage_1/main/heuristics.md prompts/stage_1/main/detection_user_prompt.md \
       prompts/stage_3/heuristics.md prompts/stage_3/analysis_prompt.md
git commit -m "Add category {N}: {Name} disinformation heuristics"
git push origin main

Gotchas

1. Stage 1 user prompt has NO subcategories. Only Description / Keywords / Heuristics / Examples. Subcategories go in the standalone heuristics.md only.
2. Stage 3 analysis_prompt.md is huge (~2600+ lines). The insertion point is between the last heuristic and ### **Additional Instructions**.
3. Dollar-quoting for SQL. When passing prompt content via SQL, use PostgreSQL dollar-quoted strings ($VUP$...$VUP$). Verify the tag doesn't appear in the content first.
4. Arabic text. Renders correctly in markdown. Don't wrap it in special formatting or RTL markers.
5. Always check current category numbers. Don't assume the last category is 21 — it changes as categories are added.
6. RLS blocks the anon key. The prompt_versions table has row-level security. Use the temp SECURITY DEFINER function for updates, then always drop it after.
7. Cross-category overlap. Health categories (20, 21) overlap with COVID (3), Abortion (6), Healthcare Reform (11), Conspiracies (14). Note these overlaps in new categories too.
8. Four files must stay in sync. Two user prompts (imported to DB) + two standalone heuristics (reference). All four need the new category.
9. Version format is semver. Check current active versions before choosing new numbers. Bump minor version for new categories (e.g., 2.1.0 → 2.2.0).
10. Always drop temp functions. The SECURITY DEFINER function bypasses RLS — leaving it is a security risk.

Source: PublicDataWorks/verdad — distributed by TomeVault.