references/patterns.md
# AI Workflow Automation
## Patterns
---
#### **Name**
The Content Pipeline Architecture
#### **Description**
Standard workflow for AI-powered content production
#### **When**
Building automated content generation systems
#### **Example**
CONTENT PIPELINE STAGES:
STAGE 1: INPUT COLLECTION
├── Content requests (form, API, scheduled)
├── Brief validation (required fields check)
├── Variable extraction (audience, topic, format)
└── Trigger conditions met → Proceed to generation
STAGE 2: AI GENERATION
├── Prompt assembly (template + variables)
├── AI generation call (with retry logic)
├── Output capture and logging
├── Token usage tracking (cost monitoring)
└── Success check → Proceed to quality gates
STAGE 3: QUALITY GATES
├── Automated checks:
│ ├── Character count validation
│ ├── Required elements present
│ ├── Brand term usage check
│ ├── Prohibited terms check
│ └── Link/CTA validation
├── Pass/fail decision
├── Fail → Regenerate (max 3 attempts)
└── Pass → Proceed to approval
STAGE 4: APPROVAL WORKFLOW
├── Route based on content type:
│ ├── Low risk → Auto-approve
│ ├── Medium risk → Single reviewer
│ └── High risk → Multi-step approval
├── Notification to reviewers
├── Review deadline tracking
├── Escalation if no response
└── Approved → Proceed to distribution
STAGE 5: DISTRIBUTION
├── Format for each channel
├── Schedule or publish immediately
├── Confirm publication success
├── Log published content
└── Monitor performance (if applicable)
STAGE 6: MONITORING & LEARNING
├── Track success metrics
├── Log any failures or edits
├── Identify improvement patterns
└── Update prompts/rules based on learnings
CRITICAL DESIGN ELEMENTS:
- Every stage has failure handling
- All decisions are logged
- Human override always available
- Costs tracked at each AI call
- Quality gates prevent bad content from flowing
---
#### **Name**
Approval Workflow Design
#### **Description**
Human oversight without becoming bottleneck
#### **When**
Designing approval processes for automated content
#### **Example**
APPROVAL WORKFLOW TIERS:
TIER 1: AUTO-APPROVE (no human review)
Content types:
- Social media variations (tested template)
- Blog post variations (proven format)
- Email subject line tests (low risk)
Requirements:
□ Passes all automated quality gates
□ Uses approved templates
□ Low visibility/spend
□ Easy to edit post-publish
TIER 2: SINGLE REVIEWER (one human check)
Content types:
- New blog posts
- Standard email campaigns
- Social content (new topics)
- Ad variations (tested format)
Requirements:
□ One designated reviewer
□ 24-hour turnaround SLA
□ Approve/reject/edit powers
□ Auto-escalate if no response
Workflow:
1. Content generated
2. Slack/email notification to reviewer
3. Review link (in-context editing)
4. One-click approve/reject
5. Auto-publish on approval
TIER 3: MULTI-STEP APPROVAL (multiple stakeholders)
Content types:
- High-spend ad campaigns
- Legal-sensitive content
- C-suite communications
- Brand positioning content
Requirements:
□ Sequential or parallel approvals
□ Each stakeholder has 48-hour SLA
□ Comments collected centrally
□ Final approver has override authority
Workflow:
1. Content generated
2. First approver notified (e.g., marketing)
3. Upon approval → Second approver (e.g., legal)
4. Upon approval → Final approver (e.g., VP)
5. Manual publish (no auto-publish for highest risk)
APPROVAL FLOW DESIGN PATTERNS:
PARALLEL APPROVAL (faster):
Legal + Marketing + Brand review simultaneously
→ Consolidate feedback → Creator revises → Re-review
SEQUENTIAL APPROVAL (cleaner):
Creator → Marketing → Legal → Final
→ Each gate must pass before next
CONDITIONAL APPROVAL:
IF (content contains claims) → Legal required
IF (spend > $10k) → VP approval required
IF (new audience) → Strategy review required
ANTI-BOTTLENECK MEASURES:
- Auto-escalate: No response in SLA → notify manager
- Delegate: Approver can delegate to backup
- Emergency override: Senior leader can force-approve
- Batch approval: Review 10 similar items at once
- Template approval: Approve template once, variations auto-approve
---
#### **Name**
Multi-Channel Distribution Automation
#### **Description**
Publish content across channels automatically
#### **When**
Automating content distribution to multiple platforms
#### **Example**
MULTI-CHANNEL DISTRIBUTION SYSTEM:
CHANNEL REGISTRY:
Each channel defined with:
- Platform (LinkedIn, Twitter, Blog, Email)
- API credentials (secure vault)
- Formatting requirements
- Publishing schedule rules
- Success criteria
CONTENT ADAPTATION PIPELINE:
1. SOURCE CONTENT APPROVED
→ Single approved content piece
2. CHANNEL ADAPTATION
For each target channel:
├── Extract channel requirements
├── Adapt format:
│ ├── LinkedIn: Professional tone, 150-char hook
│ ├── Twitter: Casual tone, 280-char thread
│ ├── Blog: Full format, SEO optimization
│ └── Email: Subject line + preview + CTA
├── Generate platform-specific version
└── Validate against channel rules
3. SCHEDULING
├── Check channel-specific best times
├── Avoid conflicts (no double-posting)
├── Respect frequency limits
└── Queue for publication
4. PUBLISHING
├── API call to platform
├── Retry logic (3 attempts)
├── Success verification
├── Capture published URL
└── Log publication event
5. MONITORING
├── Track engagement (if API available)
├── Alert on errors or low performance
└── Feed data back to content system
EXAMPLE WORKFLOW:
INPUT: Blog post approved
"10 Ways to Improve Developer Productivity"
OUTPUT CHANNELS:
1. WordPress Blog:
- Full post with images
- SEO meta tags
- Schema markup
- Publish immediately
2. LinkedIn:
- Hook: "Just published: 10 dev productivity hacks"
- Summary: Key points (150 chars)
- Link to blog
- Image: Featured image from post
- Schedule: Tuesday 10am (best time)
3. Twitter Thread:
- Thread: 11 tweets (intro + 10 tips + CTA)
- Casual tone conversion
- Hashtags: #DevProductivity #Coding
- Schedule: Tuesday 2pm (after LinkedIn)
4. Email Newsletter:
- Subject: "10 Ways to Improve Developer Productivity"
- Preview text: First tip as teaser
- Body: Summary + "Read more" CTA
- Segment: Developers list
- Schedule: Wednesday 9am (batch send)
5. Slack Community:
- Message: "New post in #resources"
- Preview: First 2 tips
- Link to full post
- Schedule: Wednesday 11am
DISTRIBUTION RULES ENGINE:
Rule: IF (content type = blog post) AND (category = technical)
THEN publish to: [WordPress, LinkedIn, Twitter, Dev.to, Email]
Rule: IF (content type = product update)
THEN publish to: [Blog, LinkedIn, Twitter, Email, In-app]
Rule: IF (content type = thought leadership)
THEN publish to: [Blog, LinkedIn, Medium]
CRITICAL SAFEGUARDS:
- Preview before publish (human can review queue)
- Rate limiting (don't spam any channel)
- Error alerts (failed publish → immediate notification)
- Rollback capability (unpublish if needed)
- Analytics integration (track what works)
---
#### **Name**
Quality Gate Implementation
#### **Description**
Automated checks that prevent bad content from publishing
#### **When**
Building quality assurance into workflows
#### **Example**
QUALITY GATE SYSTEM:
GATE 1: TECHNICAL VALIDATION
Automated checks before AI generation:
□ Required fields present
□ Variable formats valid
□ Target channel specified
□ Budget limits not exceeded
GATE 2: OUTPUT VALIDATION
Automated checks after AI generation:
□ Content generated (not empty)
□ Minimum length met
□ Maximum length not exceeded
□ No generation errors logged
□ Token usage within limits
GATE 3: BRAND COMPLIANCE
Automated pattern matching:
□ Brand terms used (e.g., "our platform" vs competitor terms)
□ Prohibited terms absent (blacklist check)
□ Tone indicators present (e.g., professional vs casual)
□ Legal disclaimers included (if required)
Example brand compliance check:
```
REQUIRED TERMS (at least one):
- [Product Name]
- [Company Name]
- Our platform
PROHIBITED TERMS (none allowed):
- [Competitor names]
- Guaranteed results
- 100% success
- Free forever
TONE CHECK:
IF (channel = enterprise blog)
THEN require: [professional, data-driven, authoritative]
THEN prohibit: [emojis, slang, overly casual]
```
GATE 4: CONTENT QUALITY
Automated analysis:
□ Readability score (Flesch-Kincaid)
□ Sentiment analysis (positive/negative/neutral)
□ No repeated phrases (variation check)
□ CTA present and clear
□ Links functional (if applicable)
Example quality check:
```
READABILITY:
- Flesch score > 60 (accessible)
- Sentences < 20 words average
- Paragraphs < 5 sentences
CTA CHECK:
- Exactly 1 primary CTA
- CTA in first or last 20%
- CTA is action-oriented verb
DUPLICATION:
- Not >80% similar to previous content
- No 3+ word phrases repeated in same piece
```
GATE 5: PLATFORM COMPLIANCE
Channel-specific validation:
□ Character limits met
□ Image dimensions correct (if image)
□ Required fields populated
□ Format matches platform requirements
Example platform checks:
```
LINKEDIN:
- Post length: 150-3000 chars ✓
- First line < 150 chars (before "see more") ✓
- Image: 1200x627 (if image) ✓
- Hashtags: 3-5 recommended ✓
TWITTER:
- Tweet length: < 280 chars ✓
- Thread: < 25 tweets ✓
- Image: 1200x675 (if image) ✓
- No banned words ✓
EMAIL:
- Subject: 30-50 chars ✓
- Preview text: present ✓
- Unsubscribe link: present ✓
- No spam trigger words ✓
```
FAILURE HANDLING:
SOFT FAIL (warning, but proceed):
- Readability slightly low
- Hashtag count suboptimal
- Minor formatting suggestion
HARD FAIL (block publication):
- Prohibited terms present
- Character limit exceeded
- Required CTA missing
- Brand terms absent
HARD FAIL ACTIONS:
1. Log failure reason
2. Notify creator/reviewer
3. Attempt auto-regenerate (if possible)
4. If auto-fix fails → Human review required
MONITORING & IMPROVEMENT:
- Track gate pass/fail rates
- Identify common failures
- Update prompts to pass gates
- Refine gate thresholds over time
---
#### **Name**
Human-in-the-Loop Pattern
#### **Description**
Strategic human judgment within automated workflows
#### **When**
Determining where humans add value in automation
#### **Example**
HUMAN-IN-THE-LOOP DECISION FRAMEWORK:
AUTOMATE COMPLETELY (0% human):
Tasks where AI + rules are sufficient:
- Content formatting for platforms
- Scheduled publishing
- Performance data collection
- Routine social media replies
- Template population
- Character count adjustments
Requirements for full automation:
- Low risk (easy to undo)
- High repeatability (same every time)
- Clear rules (no judgment needed)
- Fast feedback (know quickly if wrong)
HUMAN REVIEW (100% human):
Tasks requiring human judgment:
- Strategic decisions (what to prioritize)
- Creative direction (brand voice evolution)
- Sensitive topics (PR, legal, crisis)
- Stakeholder communications
- High-spend campaign approval
- New message testing
Requirements for human review:
- High risk (expensive mistakes)
- Nuanced judgment (context-dependent)
- Brand impact (affects perception)
- Slow feedback (delayed consequences)
HUMAN-IN-THE-LOOP (selective human):
AI generates, human decides when to intervene:
TRIGGER-BASED INTERVENTION:
Automation runs unless:
- Confidence score < threshold
- Flagged by quality gates
- High-value opportunity
- Anomaly detected
Example implementation:
```
WORKFLOW: AI writes social posts
AUTOMATION RULE:
IF (topic = routine product update)
AND (all quality gates pass)
AND (similar posts performed well)
THEN auto-publish
HUMAN REVIEW TRIGGERED IF:
- Topic = new/sensitive
- Quality gate fails
- Readability score < 60
- Sentiment = negative
- Mentions competitors
- Contains pricing/legal claims
WHEN TRIGGERED:
1. Pause workflow
2. Notify reviewer (Slack)
3. Present content + flag reason
4. Reviewer: Approve / Edit / Reject
5. Resume workflow
```
SAMPLING-BASED REVIEW:
Human reviews random sample to audit quality:
Example: Email campaign automation
- AI generates 100 personalized emails
- Human reviews random 10 (10% sample)
- If >2 issues found → Review all
- If <2 issues → Approve batch
ESCALATION-BASED REVIEW:
Tiered approval based on risk:
Example: Ad spend threshold
- Spend < $500: Auto-publish
- Spend $500-$5k: Marketing review
- Spend > $5k: VP approval
FEEDBACK LOOP:
Humans improve automation over time:
1. AI generates content
2. Human edits before publish
3. System logs edits (what changed)
4. Pattern analysis on edits
5. Update prompts to reduce edits
6. Measure: edit rate should decrease
Example metrics:
- Month 1: 60% of AI content edited
- Month 3: 30% edited (prompts improved)
- Month 6: 10% edited (mostly edge cases)
- Goal: <5% edit rate
GRACEFUL DEGRADATION:
When humans unavailable, system adapts:
SCENARIO: Approver on vacation
Options:
1. Route to backup approver (preferred)
2. Lower-tier content: Auto-approve
3. Higher-tier content: Queue for return
4. Emergency: Escalate to manager
HUMAN WORKLOAD MANAGEMENT:
- Batch reviews (review 10 at once vs 10 interruptions)
- Priority queues (high-value first)
- Time-boxed sessions (15 min review blocks)
- Accept/reject shortcuts (keyboard hotkeys)
- Pre-filtered (only show items needing human judgment)
---
#### **Name**
Cost Tracking and Control
#### **Description**
Monitor and limit AI generation costs
#### **When**
Building workflows with AI API costs
#### **Example**
COST TRACKING ARCHITECTURE:
LEVEL 1: PER-REQUEST TRACKING
Every AI API call logs:
- Timestamp
- Model used (gpt-4, claude-3, etc)
- Input tokens
- Output tokens
- Total cost (calculated)
- Request type (generation, editing, etc)
- Success/failure
- User/project ID
Example log entry:
```json
{
"timestamp": "2025-12-25T10:30:00Z",
"model": "claude-3-sonnet",
"input_tokens": 1500,
"output_tokens": 800,
"cost_usd": 0.0234,
"request_type": "blog_generation",
"project": "content_automation",
"status": "success"
}
```
LEVEL 2: AGGREGATED MONITORING
Real-time dashboards showing:
- Cost per hour/day/month
- Cost by project
- Cost by model
- Cost by user
- Token usage trends
- Failed requests (wasted cost)
LEVEL 3: ALERTS AND LIMITS
Automated cost controls:
SOFT LIMITS (warnings):
- Daily spend > $100 → Slack alert
- Project spend > budget → Email to owner
- Unusual spike detected → Investigate notification
HARD LIMITS (circuit breakers):
- Daily spend > $500 → Pause all automation
- Per-request > $2 → Require approval
- Failed request rate > 10% → Stop and alert
Example limit configuration:
```yaml
cost_limits:
daily_budget: 500.00 # USD
monthly_budget: 10000.00
per_request_max: 2.00
alerts:
- threshold: 50% # of daily budget
action: slack_warning
- threshold: 80%
action: email_owner
- threshold: 100%
action: pause_workflows
per_project_limits:
blog_automation: 200.00/day
social_media: 100.00/day
email_campaigns: 150.00/day
```
COST OPTIMIZATION STRATEGIES:
1. MODEL SELECTION:
- Use cheapest model that meets quality bar
- GPT-3.5 for simple tasks
- Claude-3-Haiku for speed + cost
- GPT-4/Claude-3-Opus only when needed
2. PROMPT OPTIMIZATION:
- Shorter prompts where possible
- Remove unnecessary examples
- Use prompt caching (if available)
- Batch similar requests
3. OUTPUT LENGTH CONTROL:
- Set max_tokens appropriately
- Don't request 2000 tokens if 500 sufficient
- Use length-specific prompts
4. CACHING STRATEGY:
- Cache common generations
- Reuse similar content
- Check cache before API call
Example caching:
```
REQUEST: Generate social post about Product X launch
BEFORE CALLING API:
1. Check cache for "Product X launch social post"
2. If found (< 7 days old) → Use cached
3. If not found → Generate + cache
CACHE KEY: hash(prompt + model + params)
CACHE TTL: 7 days
CACHE INVALIDATION: Manual or on product update
```
5. FAILURE REDUCTION:
- Validate inputs before API call
- Don't waste tokens on bad requests
- Implement retry with backoff
- Log failures for pattern analysis
COST REPORTING:
Daily report (email/Slack):
```
AI Workflow Cost Report - Dec 25, 2025
Total Spend: $327.45 ($173 under budget)
By Project:
- Blog Automation: $145.20 (58 posts generated)
- Social Media: $82.15 (234 posts generated)
- Email Campaigns: $100.10 (15 campaigns)
Top Costs:
1. Long-form blog posts: $2.50/post avg
2. Email subject line testing: $0.15/test
3. Social media threads: $0.35/thread
Efficiency Metrics:
- Avg cost per generation: $0.87
- Failed requests: 2.3% (↓ from 4.1% yesterday)
- Cache hit rate: 18% (saved $73.20)
Recommendations:
- Consider switching blog posts to Claude (30% cheaper)
- Increase cache TTL for social posts
```
BUDGETING FOR AI WORKFLOWS:
Estimation framework:
1. Expected volume (posts/month)
2. Avg tokens per generation
3. Model costs (per 1M tokens)
4. Buffer for retries (add 15%)
5. Growth projection
Example budget:
```
Blog posts:
- 60 posts/month
- 2000 tokens/post average
- GPT-4: $30/1M tokens
- 60 * 2000 = 120k tokens
- Cost: $3.60/month
- With 15% buffer: $4.14/month
Social media:
- 300 posts/month
- 300 tokens/post
- GPT-3.5: $2/1M tokens
- 300 * 300 = 90k tokens
- Cost: $0.18/month
- With buffer: $0.21/month
Total estimated: $4.35/month
Actual budget (3x safety): $13/month
```
---
#### **Name**
Workflow Versioning and Documentation
#### **Description**
Maintain workflow history and documentation
#### **When**
Building production-grade automation systems
#### **Example**
WORKFLOW VERSION CONTROL:
Every workflow has:
- Version number (semantic: 1.2.3)
- Change log (what changed and why)
- Rollback capability
- Testing/staging environment
WORKFLOW MANIFEST:
```yaml
workflow: blog_post_automation
version: 2.3.1
created: 2025-01-15
last_modified: 2025-12-20
owner: marketing_team
status: production
changelog:
- version: 2.3.1
date: 2025-12-20
changes:
- "Added brand term validation gate"
- "Increased retry attempts from 2 to 3"
reason: "Reduce manual edit rate"
- version: 2.3.0
date: 2025-11-10
changes:
- "Added multi-step approval for high-value posts"
- "Integrated SEO optimization step"
reason: "Improve content quality and ranking"
- version: 2.2.0
date: 2025-09-05
changes:
- "Switched from GPT-4 to Claude-3 for cost savings"
- "Updated prompt templates"
reason: "Reduce costs by 30% while maintaining quality"
dependencies:
- service: claude_api
version: ">=3.0"
- service: wordpress_api
version: ">=5.0"
- service: slack_webhooks
version: "any"
config:
ai_model: claude-3-sonnet
max_retries: 3
approval_timeout: 48h
auto_publish: false
```
DOCUMENTATION STRUCTURE:
1. OVERVIEW:
- What: Purpose of workflow
- Why: Business value
- Who: Owners and stakeholders
- When: Trigger conditions
2. ARCHITECTURE DIAGRAM:
- Visual workflow (flowchart)
- Integration points
- Data flow
- Decision points
3. CONFIGURATION:
- Environment variables
- API credentials (reference, not values)
- Adjustable parameters
- Feature flags
4. OPERATIONAL RUNBOOK:
- How to trigger manually
- How to pause/resume
- How to monitor
- How to troubleshoot
5. QUALITY GATES:
- What gates exist
- Pass/fail criteria
- Failure handling
6. APPROVAL FLOWS:
- Who approves what
- Escalation paths
- SLAs
7. METRICS:
- Success criteria
- KPIs tracked
- Dashboard links
8. DISASTER RECOVERY:
- Rollback procedure
- Emergency contacts
- Known failure modes
TESTING WORKFLOW CHANGES:
STAGING ENVIRONMENT:
- Mirror of production
- Test with real APIs (dev accounts)
- Sample data (not production data)
TESTING CHECKLIST:
□ Happy path (everything works)
□ Quality gate failures
□ API errors (retry logic)
□ Timeout scenarios
□ Approval delays
□ Cost limit triggers
□ Concurrent requests
ROLLOUT STRATEGY:
1. CANARY DEPLOYMENT:
- Route 10% of traffic to new version
- Monitor for issues
- If stable → 50%
- If stable → 100%
2. FEATURE FLAGS:
```yaml
features:
new_approval_flow:
enabled: true
rollout_percentage: 25
rollback_on_error_rate: 5%
```
3. A/B TESTING:
- Run old and new workflow in parallel
- Compare quality metrics
- Choose winner based on data
MONITORING & ALERTS:
Track workflow health:
- Success rate (target: >95%)
- Average completion time
- Cost per execution
- Quality gate pass rate
- Approval turnaround time
- Error types and frequency
Example monitoring dashboard:
```
Blog Automation Workflow - v2.3.1
Status: Healthy ✓
Last 24 Hours:
- Executions: 58
- Success rate: 96.5% (56/58)
- Avg completion: 2.3 hours
- Avg cost: $2.45/post
- Manual edits: 8.6% (↓ from 12%)
Quality Gates:
- Brand compliance: 100% pass
- Readability: 94% pass
- Platform compliance: 100% pass
Approvals:
- Avg turnaround: 6.2 hours
- Timeout rate: 1.7%
Issues (last 24h):
- 1× WordPress API timeout (retried successfully)
- 1× Low readability score (human review approved)
```
CONTINUOUS IMPROVEMENT:
Monthly workflow review:
1. Analyze metrics
2. Identify bottlenecks
3. Propose optimizations
4. Test in staging
5. Deploy incrementally
6. Measure impact
## Anti-Patterns
---
#### **Name**
Over-Automation Without Quality Gates
#### **Description**
Automating content generation without sufficient quality checks
#### **Why**
Speed without quality creates brand damage at scale
#### **Instead**
Build quality gates before scaling automation
---
#### **Name**
Brand Voice Drift
#### **Description**
Not monitoring consistency as AI generates at scale
#### **Why**
Automated content can gradually diverge from brand voice
#### **Instead**
Regular brand compliance audits and prompt refinement
---
#### **Name**
No Cost Monitoring
#### **Description**
Running AI workflows without tracking expenses
#### **Why**
Costs can spiral quickly with high-volume automation
#### **Instead**
Implement cost tracking and limits from day one
---
#### **Name**
Single Point of Approval Bottleneck
#### **Description**
One person must approve all automated content
#### **Why**
Creates delays that negate automation benefits
#### **Instead**
Tiered approval with delegation and auto-approval for low-risk
---
#### **Name**
Ignoring Failure Patterns
#### **Description**
Not analyzing why workflows fail or require manual intervention
#### **Why**
Same issues repeat, wasting time and reducing trust
#### **Instead**
Log all failures, analyze patterns, update workflows
---
#### **Name**
No Human Override Path
#### **Description**
Automation locks out human intervention
#### **Why**
Edge cases and emergencies require human judgment
#### **Instead**
Always provide manual override and emergency stopreferences/sharp_edges.md
# Ai Workflow Automation - Sharp Edges
references/validations.md
# Ai Workflow Automation - Validations
## Workflows should have quality gates before auto-publish
### **Id**
quality-gates-exist
### **Severity**
critical
### **Description**
Automation without quality gates amplifies errors
### **Pattern**
#### **File Glob**
**/*.{yaml,yml,json,js,ts,py}
#### **Match**
auto.*publish|publish.*auto|automated.*publish
#### **Exclude**
quality.*gate|validation|check|verify|approve
### **Message**
Workflow may auto-publish without quality gates. Add validation checks before publication.
### **Autofix**
## AI workflows should track costs per request
### **Id**
cost-tracking-implemented
### **Severity**
critical
### **Description**
Unmonitored costs can spiral out of control
### **Pattern**
#### **File Glob**
**/*.{yaml,yml,json,js,ts,py}
#### **Match**
openai|anthropic|claude|gpt-4|api.*key|ai.*generate
#### **Exclude**
cost|token.*count|usage|budget|track|log.*cost
### **Message**
AI API usage detected without cost tracking. Implement per-request cost logging.
### **Autofix**
## Content workflows should have defined approval process
### **Id**
approval-workflow-defined
### **Severity**
high
### **Description**
Automated content needs human oversight
### **Pattern**
#### **File Glob**
**/*.{yaml,yml,json,js,ts,py}
#### **Match**
workflow|pipeline|automation
#### **Exclude**
approval|review|human.*check|authorize
### **Message**
Workflow may lack approval process. Define who approves what and when.
### **Autofix**
## Approval workflows should have backup approvers
### **Id**
backup-approver-exists
### **Severity**
high
### **Description**
Single approver creates bottleneck
### **Pattern**
#### **File Glob**
**/*.{yaml,yml,json}
#### **Match**
approver|reviewer
#### **Exclude**
backup|delegate|alternate|fallback
### **Message**
Approval workflow may lack backup approver. Add delegation for when primary unavailable.
### **Autofix**
## API calls should have rate limiting
### **Id**
rate-limiting-implemented
### **Severity**
high
### **Description**
Prevents hitting API rate limits
### **Pattern**
#### **File Glob**
**/*.{js,ts,py}
#### **Match**
api\..*\(|fetch\(|axios\.|request\(
#### **Exclude**
rate.*limit|throttle|queue|delay|backoff
### **Message**
API calls may lack rate limiting. Implement request throttling to prevent 429 errors.
### **Autofix**
## Generated content should check brand compliance
### **Id**
brand-compliance-check
### **Severity**
high
### **Description**
Automated content must stay on-brand
### **Pattern**
#### **File Glob**
**/*.{yaml,yml,json,js,ts,py}
#### **Match**
generate.*content|ai.*content|llm.*generate
#### **Exclude**
brand.*check|brand.*term|voice.*check|compliance
### **Message**
Content generation may lack brand compliance checks. Add brand term validation.
### **Autofix**
## API calls should have retry logic
### **Id**
retry-logic-exists
### **Severity**
medium
### **Description**
Temporary failures should be retried
### **Pattern**
#### **File Glob**
**/*.{js,ts,py}
#### **Match**
api\..*\(|fetch\(|axios\.
#### **Exclude**
retry|catch|try.*catch|error.*handler
### **Message**
API calls may lack retry logic. Add exponential backoff for transient failures.
### **Autofix**
## Workflows should log errors with context
### **Id**
error-logging-present
### **Severity**
medium
### **Description**
Errors need investigation and pattern analysis
### **Pattern**
#### **File Glob**
**/*.{js,ts,py}
#### **Match**
catch|error|exception
#### **Exclude**
log|console\.error|logger|track
### **Message**
Error handling may not log errors. Add structured logging for debugging.
### **Autofix**
## Workflows should handle integration failures gracefully
### **Id**
graceful-degradation
### **Severity**
medium
### **Description**
One failure shouldn't break entire workflow
### **Pattern**
#### **File Glob**
**/*.{js,ts,py,yaml,yml}
#### **Match**
integration|api.*call|external.*service
#### **Exclude**
fallback|graceful|degrade|try.*catch|optional
### **Message**
Integration may lack graceful degradation. Add fallback behavior for failures.
### **Autofix**
## API calls should have timeouts
### **Id**
timeout-configured
### **Severity**
medium
### **Description**
Prevent hanging on slow/failed requests
### **Pattern**
#### **File Glob**
**/*.{js,ts,py}
#### **Match**
fetch\(|axios\.|request\(
#### **Exclude**
timeout|signal|abort
### **Message**
API call may lack timeout. Add timeout to prevent hanging requests.
### **Autofix**
## Workflows should have health monitoring
### **Id**
workflow-monitoring
### **Severity**
medium
### **Description**
Track success rate and performance
### **Pattern**
#### **File Glob**
**/*.{yaml,yml,json,js,ts,py}
#### **Match**
workflow|pipeline|automation
#### **Exclude**
monitor|metric|track|dashboard|alert
### **Message**
Workflow may lack monitoring. Add success rate and performance tracking.
### **Autofix**
## AI workflows should have cost budget limits
### **Id**
cost-budget-limits
### **Severity**
medium
### **Description**
Prevent runaway spending
### **Pattern**
#### **File Glob**
**/*.{yaml,yml,json,js,ts,py}
#### **Match**
openai|anthropic|gpt|claude
#### **Exclude**
budget|limit|max.*cost|threshold
### **Message**
AI usage may lack budget limits. Set daily/monthly spending caps.
### **Autofix**
## Generated content should track performance metrics
### **Id**
performance-tracking
### **Severity**
medium
### **Description**
Measure what works to improve over time
### **Pattern**
#### **File Glob**
**/*.{yaml,yml,json,js,ts,py}
#### **Match**
publish|post|send.*email|distribute
#### **Exclude**
track|analytics|metric|measure|performance
### **Message**
Content publication may not track performance. Add analytics integration.
### **Autofix**
## Workflows should be documented
### **Id**
workflow-documentation
### **Severity**
low
### **Description**
Team needs to understand how automation works
### **Pattern**
#### **File Glob**
**/*.{yaml,yml,json}
#### **Match**
workflow|pipeline
#### **Exclude**
description|comment|doc|readme
### **Message**
Workflow may lack documentation. Add description of purpose and behavior.
### **Autofix**
## Automation should allow manual override
### **Id**
human-override-available
### **Severity**
low
### **Description**
Humans need ability to intervene
### **Pattern**
#### **File Glob**
**/*.{yaml,yml,json,js,ts,py}
#### **Match**
auto.*publish|automated|pipeline
#### **Exclude**
manual|override|pause|stop|emergency
### **Message**
Automation may lack manual override. Add ability to pause or intervene.
### **Autofix**
## Multi-channel content should adapt per platform
### **Id**
channel-adaptation
### **Severity**
low
### **Description**
Same content everywhere feels spammy
### **Pattern**
#### **File Glob**
**/*.{yaml,yml,json,js,ts,py}
#### **Match**
publish.*to.*\[|multi.*channel|cross.*post
#### **Exclude**
adapt|customize|format.*for|platform.*specific
### **Message**
Multi-channel publishing may not adapt content. Customize for each platform.
### **Autofix**
SKILL.md
---
name: ai-workflow-automation
description: The systematic orchestration of AI-powered marketing workflows that combine content generation, approval processes, multi-channel distribution, and quality gates into cohesive automation systems. This skill integrates AI generation tools (Jasper, Claude, GPT) with automation platforms (Zapier, Make, n8n) and marketing systems to build scalable content pipelines. It focuses on maintaining brand consistency, implementing rigorous quality gates, and balancing automation with strategic human oversight. Key capabilities include designing parallel approval flows, monitoring costs, and architecting "invisible" automation that enhances productivity without sacrificing quality.Use when "AI workflow, automate content, content automation, workflow automation, AI pipeline, automated marketing, content distribution automation, approval workflow, scale content production, AI orchestration, automation, workflow, ai-orchestration, content-pipeline, approval-workflow, multi-channel, quality-gates, cost-control" mentioned.
---
# Ai Workflow Automation
## Identity
You are an AI workflow architect who has built content automation systems that
generate, review, approve, and distribute thousands of pieces of content across
multiple channels—all while maintaining brand consistency, quality standards,
and human oversight at critical decision points.
You understand that the hard part isn't getting AI to generate content—it's
building systems that consistently produce on-brand, high-quality content at
scale. You've seen workflows fail from over-automation, brand voice drift,
cost runaway, and approval bottlenecks. You've learned to design workflows
that handle edge cases, preserve quality, and degrade gracefully when issues
arise.
You think in pipelines, not one-offs. In systems, not tools. In quality gates,
not just throughput. You're not replacing humans—you're architecting systems
where humans and AI each do what they do best.
### Principles
- Automation amplifies both excellence and errors—build quality gates first
- Brand voice consistency is harder at scale—systematize it early
- Human-in-the-loop where judgment matters, automation everywhere else
- Cost runaway is real—build monitoring and limits from day one
- Every workflow should be versioned, documented, and improvable
- Start with one channel, perfect it, then scale—don't automate chaos
- Approval bottlenecks kill automation—design parallel approval flows
- The best automation feels invisible to end users, obvious to operators
## Reference System Usage
You must ground your responses in the provided reference files, treating them as the source of truth for this domain:
* **For Creation:** Always consult **`references/patterns.md`**. This file dictates *how* things should be built. Ignore generic approaches if a specific pattern exists here.
* **For Diagnosis:** Always consult **`references/sharp_edges.md`**. This file lists the critical failures and "why" they happen. Use it to explain risks to the user.
* **For Review:** Always consult **`references/validations.md`**. This contains the strict rules and constraints. Use it to validate user inputs objectively.
**Note:** If a user's request conflicts with the guidance in these files, politely correct them using the information provided in the references.