Date Validation

Date Validation

Overview

Prevent day-of-week errors, countdown mistakes, and timeline inconsistencies before they reach files.

Core principle: Trust but verify. Establish independent ground truth, validate every date systematically, block edits until all pass.

CRITICAL: This skill exists because baseline testing showed agents:

• Trust anchor dates without verification (propagates errors)
• Spot-fix obvious errors, miss related ones (incomplete fixes)
• Trust sequential patterns without individual validation (consistent ≠ correct)
• Only verify when suspicious (errors slip through)

When to Use

Auto-Triggers (Mandatory)

This skill MUST activate before Edit/Write when:

File patterns:

• *Planning-Hub.md
• *Timeline*.md
• *Checklist*.md
• calendar/day/*.md
• calendar/week/*.md
• Filename contains: "schedule", "deadline", "milestone"

Content patterns:

• Day-name + date: Wed Nov 12, Monday, November 11
• Relative dates: today, tomorrow, yesterday, in X days
• Countdowns: X days until, X days remaining, X days from today

Manual Use

Invoke when:

• User says "check the dates" or "verify timeline"
• You notice day-name + date combinations while reading
• After making date-related errors (systematic revalidation)

Ground Truth Protocol

CRITICAL: Verify "today" using TWO independent sources. If they disagree, BLOCK ALL EDITS.

Step 1: Get Both Sources

Source A - Environment Tag:

Parse: <env>Today's date: YYYY-MM-DD</env>
Example: "2025-11-12"

Source B - System Command:

date +"%Y-%m-%d %A"
# Output: "2025-11-12 Wednesday"

Step 2: Cross-Validate

Extract components:

• env_date: Date from <env> tag
• sys_date: Date from system command
• sys_day: Day-of-week from system command

If dates MATCH:

✓ Ground truth established
Today is: Wednesday, November 12, 2025

If dates DON'T MATCH:

❌ GROUND TRUTH MISMATCH - BLOCKING ALL EDITS

Environment says: 2025-11-11
System says: 2025-11-12 Wednesday

Reason: <env> tag likely stale from context compression
Action: User must restart session

STOP: Do not proceed with validations or edits

Why two sources: Baseline testing showed agents "trust user-provided context completely" without verification. If anchor is wrong, all calculations are wrong.

Step 3: Calculate Reference Values

Once verified:

• today_date: 2025-11-12
• today_day: Wednesday
• tomorrow_date: 2025-11-13 (today + 1)
• yesterday_date: 2025-11-11 (today - 1)

Validation Checklist

Run ALL validations before presenting errors. Don't stop at first error.

☐ Check 1: Day-Name ↔ Date Matching

What to find:

Patterns where day-name appears with date:

• Wed Nov 12
• Wednesday, November 12
• Thu 11/14/2025
• Mon Nov 11, Tue Nov 12, Wed Nov 13 (sequences)

Regex:

r'(Monday|Tuesday|Wednesday|Thursday|Friday|Saturday|Sunday)\s*,?\s*([A-Z][a-z]+)\s+(\d{1,2})'
r'(Mon|Tue|Wed|Thu|Fri|Sat|Sun)\s+([A-Z][a-z]+)\s+(\d{1,2})'

For each match:

1. Parse date (US MM/DD format for numeric dates)
2. Calculate actual day-of-week:

   from datetime import datetime
   date_obj = datetime.strptime("2025-11-12", "%Y-%m-%d")
   actual_day = date_obj.strftime("%A")  # "Wednesday"
   

3. Compare stated vs. actual

If mismatch:

❌ Error: Day-name mismatch
Found: "Wed Nov 12"
Truth: November 12, 2025 is Tuesday

Options:
1. Tue Nov 12 (correct day name)
2. Wed Nov 13 (correct date)
3. Something else (I'll type it)

CRITICAL: Validate EACH date individually. Baseline showed "sequential pattern feels right" even when all wrong.

☐ Check 2: Relative Date Calculations

What to find:

• today → should match ground truth
• tomorrow → ground truth + 1 day
• yesterday → ground truth - 1 day
• in X days → ground truth + X days

Validation: For each term, check surrounding text (~50 chars) for date mentions. Verify alignment.

Example error:

Text: "tomorrow (Wed Nov 13)"
Today: Wednesday, Nov 12
Tomorrow should be: Thursday, Nov 13 (not Wed)

If mismatch:

❌ Error: Relative date incorrect
Found: "tomorrow (Wed Nov 13)"
Truth: Tomorrow is Thursday, Nov 13

Options:
1. tomorrow (Thu Nov 13)
2. today (Wed Nov 12)
3. Something else

☐ Check 3: Countdown Math

What to find:

• X days until [date]
• X days remaining
• X days to [date]

Regex:

r'(\d+)\s+days?\s+(until|to|before)\s+([A-Za-z]+\s+\d{1,2})'
r'(\d+)\s+days?\s+remaining'

Validation:

1. Extract countdown number
2. Find target date
3. Calculate: actual = target_date - ground_truth_date
4. Compare stated vs. actual

If mismatch:

❌ Error: Countdown math incorrect
Found: "December 1, 2025 (18 days from today!)"
Truth: Dec 1 - Nov 12 = 19 days

Options:
1. (19 days from today!)
2. (18 days from tomorrow!)
3. Something else

☐ Check 4: Comprehensive Scan

CRITICAL: Baseline showed agents "fix obvious error" but miss others.

After finding first error, DON'T stop. Scan ENTIRE content:

• Check ALL lines with date references
• Include lines without day-names (might need them)
• Check headers, footers, tables, lists

Example from baseline: Agent fixed line 239 but missed line 74 (complex) and line 285 (no day-name).

Solution: Regex scan catches ALL occurrences before presenting errors.

Error Correction Flow

Step 1: Collect All Errors

Run all validations. Group results:

Found 3 errors:

CRITICAL:
- Error #1: Day-name mismatch at line 174
- Error #2: Countdown math at line 163

WARNINGS:
- Warning #1: Sequential inconsistency at line 240

Step 2: Present Interactive Corrections

For each error:

1. Show context
2. Show what was found
3. Show what ground truth says
4. Offer 2-3 correction options
5. Wait for user choice

Step 3: Apply and Re-Validate

1. Apply all corrections
2. Re-run ALL validations
3. If new errors → Return to Step 1
4. If all pass → Allow Edit/Write

Red Flags - STOP and Verify

If you catch yourself thinking:

• ❌ "Trust the user-provided context completely"
• ❌ "Would NOT naturally use a calendar tool"
• ❌ "Only check if something felt obviously wrong"
• ❌ "Fix the most obvious error"
• ❌ "Quick scan with shallow verification"
• ❌ "Might have trusted the sequential pattern"
• ❌ "Internal consistency feels right"
• ❌ "Not verified each date individually"
• ❌ "Without explicit contradiction, accept at face value"

All of these mean: Use two-source ground truth, validate systematically, check every date.

Common Rationalizations

These are from baseline testing - agents said these things when NOT using this skill:

Edge Cases

Ambiguous Dates

Problem: 11/12 could be Nov 12 or Dec 11

Solution:

1. Default: US format (MM/DD) = November 12
2. Check context for month names
3. If ambiguous: Prompt user

Past Dates in Future Tense

Problem: tomorrow (Nov 13) but today is Nov 15

Solution:

• Warning only (don't block)
• Show: ⚠️ Note: Nov 13 was 2 days ago
• Ask: Is this an old document? (y/n)

Year Boundaries

Problem: Dec 30 → Jan 2 crosses years

Solution:

• Show year explicitly in corrections
• Handle leap years with Python datetime

Success Criteria

Before allowing Edit/Write:

• ✅ Ground truth verified (env + system agree)
• ✅ ALL date patterns scanned
• ✅ EACH date validated individually
• ✅ All errors corrected by user
• ✅ Re-validation passed

Only after ALL criteria → Proceed with edit.

Example Session

User: "Update Move Planning Hub with Nov 13 events"

[Skill auto-triggers on "Planning Hub" filename]

Step 1: Ground truth
  ENV: 2025-11-12
  SYS: 2025-11-12 Wednesday
  ✓ Verified: Wednesday, November 12, 2025

Step 2: Scan content
  Found: 15 date references

Step 3: Validate
  ✓ Day-name matching (8/8)
  ❌ Relative dates (1 error)
  ❌ Countdown math (1 error)

Found 2 errors:

❌ Error #1: "TODAY - NOVEMBER 13"
Truth: Today is November 12
Options: 1. TODAY - NOVEMBER 12  2. TOMORROW - NOVEMBER 13

User choice: 2

❌ Error #2: "(18 days from today)"
Truth: Dec 1 - Nov 12 = 19 days
Options: 1. (19 days)  2. Keep 18

User choice: 1

Step 4: Re-validate
  ✓ All validations passed

✅ Proceeding with Edit...

Anti-Patterns from Baseline

DON'T:

• Trust anchor without two-source verification
• Fix one error and move on (scan comprehensively)
• Trust sequential patterns (validate each)
• Only verify when suspicious (always verify)
• Do mental math instead of bash date calculation
• Skip re-validation after corrections

DO:

• Establish ground truth first (env + system)
• Scan ALL content before presenting errors
• Validate EACH date individually
• Auto-trigger before edits (no manual check needed)
• Use Python datetime for calculations
• Re-validate after applying corrections