An Evidence-First Review Routine
A fictional sequence to adapt and measure—not a promised schedule or productivity result.
Run each tenant separately. Duration, usefulness, and review effort depend on account size, platform availability, data quality, rate limits, attribution lag, workflow design, and operator experience.
/daily-report separately for each tenant. A named reviewer checks sources, definitions, claims, and recipients before export or send; client data is never joined or reused across tenants.Record a pre-AI baseline, then measure source accuracy, correction rate, reviewer acceptance, review effort, latency, full cost, incidents, and business guardrails. Report the observed window and limitations; do not generalize one team's result into a fixed time-saving claim.
Chaining Components Without Chaining Trust
Chained steps can prepare evidence and drafts, but each side effect crosses a typed, policy-enforced approval boundary.
One component's output is untrusted input to the next. Validate schemas, provenance, tenant scope, and claims at every boundary. The model remains read-only and proposal-only and never receives credentials. Writes use a separate identity through a deterministic executor that enforces server-side reviewer and account binding, full immutable diff, entity and spend allowlists, exact approval, stable idempotency, rate limits, audit logs, monitoring, stop, and rollback.
Example: Fictional Campaign Proposal
More Workflow Ideas
| Workflow | Agents Used | What Happens |
|---|---|---|
| Daily Ops Check | Monitor + Budget | Check authorized metrics, draft budget change-set, generate summary |
| Creative Refresh | Monitor + Creative | Flag a possible fatigue pattern, rank alternatives, draft reviewed variations, queue |
| Scaling Protocol | Monitor + Budget + Campaign | Rank candidates, draft a capped canary in PAUSED state, await exact approval |
| Weekly Report | Monitor + Report | Pull tenant-scoped data, draft sourced report, human-review recipients and send |
| Market Observation | Monitor + Creative + Campaign | Corroborate permitted observations, draft substantiated creative and a reversible canary |
Building Custom Skills (Advanced)
For experienced users documenting narrow, repeatable, low-risk assistance under explicit permissions and evaluations.
This section is for people who've mastered the pre-built agents. If you're just starting, Parts 4-6 have you covered. Come back here when you're ready to customize.
A suitable bounded procedure can be documented as a Skill when its purpose, inputs, permissions, failure behavior, tests, and prohibited actions are clear. A Skill is instruction content, not a security boundary or automatic production authorization. Check the current official Agent Skills documentation.
---
name: my-custom-skill
description: One line describing what this skill does.
---
# My Custom Skill
## Purpose
What this skill accomplishes and why.
## When to Use
- User says "/my-command" or asks about X
- Specific trigger conditions
## Prerequisites
- Approved tenant-scoped read service and minimum fields
- Synthetic fixtures, task-specific evals, named reviewer
- No secrets or write credentials in the Skill or model context
## Compatibility & Failure Protocol
If an API contract or write fails:
1. Stop writes; record request ID, pinned version, and redacted evidence
2. Review the pinned official reference and changelog
3. Make a human-reviewed, version-controlled code/schema change
4. Run offline evals and contract tests in a test account
5. Canary with tenant scope, PAUSED, cap, idempotency, approval, monitoring, rollback
6. Never mutate production schema or retry a write from model output
## Instructions
### Step 1: [First Action]
### Step 2: [Second Action]
### Step 3: [Decision Point]
## Configuration
- Non-secret identifiers and policy references only
- Secrets remain in the independent server-side service
## Example Usage
User: /my-command with these parameters
SOP to Skill Conversion
| Feature | SOP (Document) | Skill (Executable) |
|---|---|---|
| Format | Any text document | skill-name/SKILL.md with supported frontmatter |
| Trigger | You tell Claude to follow it | Selected when invoked or relevant under the current product rules |
| Compatibility | Manual update | Reviewed docs, versioned change, evals, contract tests, canary, rollback |
| Reusable | Copy-paste each time | Versioned instructions with regression evaluation |
A reviewed skill library can make recurring analysis more consistent, but it does not guarantee capacity or results. Measure quality, error rate, review time, and business outcomes; keep client tenants isolated and every side effect behind deterministic controls.
MarTech: A Cross-Functional Practice
Marketing technology connects domain expertise with data, software, governance, and operations.
MarTech (Marketing Technology) is the practice of designing and operating technology that supports marketing outcomes under clear data, security, accessibility, financial, legal, and platform constraints.
No framework can predict demand, compensation, career progression, or individual value. The durable skill is accountable cross-functional work: understanding the marketing decision while partnering with engineering, analytics, security, privacy, legal, finance, creative, and operations.
What MarTech Practitioners Do
- Implement AI and automation into marketing operations
- Build systems that connect advertising, analytics, CRM, and store platforms
- Work as technical partners with performance teams and agencies
- Evaluate whether a technology improves a defined baseline without violating guardrails
- Turn authorized evidence into reviewable proposals and controlled systems, not unsupported claims
Mazin Zaki — Marketing Technology Advisor & eCommerce Scaling Consultant. I leverage and implement the highest level of technologies and AI into marketing activities to improve and scale eCommerce businesses. I work as a Technical Partner with performance teams and agencies to provide a higher edge in the market.
Follow @itsmazinzaki on Instagram for more frameworks, insights, and advanced content. Follow avamartech.io for future releases.
Tenant-Safe Learning Opportunities
Useful systems can be built without pooling, reusing, or exposing client data.
Access to client data is a duty, not permission to create a shared asset. Isolate every client's credentials, storage, vector indexes, logs, prompts, reports, evals, and exports by tenant and purpose. Enforce authorization server-side on every read and write.
Do not train on, benchmark against, compare, or reuse one client's data for another client or your own product without a specific lawful agreement, documented purpose, data-minimization plan, retention limits, and required consent. “Anonymized” or aggregated does not automatically make reuse safe or lawful.
Safer Sources of Learning
- Per-client evals using that tenant's authorized, minimized data and account-specific policy
- Synthetic fixtures for contract, safety, duplicate, partial-failure, and rollback tests
- Official public data used within published terms and with source, date, and uncertainty
- Client-approved research under a separate agreement with purpose, access, retention, and deletion controls
- Operational metadata that measures reliability without exposing campaign or customer content
Product ideas should be validated through opt-in research and clean, purpose-built datasets. Existing clients are not automatic test subjects; obtain separate approval before experimentation and never move their data into a pooled model or shared memory by default.
Trustworthy isolation, reproducible evals, clear provenance, and client-controlled use create a stronger long-term capability than opaque data pooling. Make “no cross-client reuse by default” an enforced architecture rule.
Disclaimer: Test Before You Trust
AI is powerful but not infallible. Here are the rules.
Follow these rules without exception:
- Define risk-based acceptance criteria. A run count is not a safety gate. Specify offline eval thresholds, required test coverage, maximum harm, owners, and evidence for each workflow.
- Test the contract and failure modes. Use synthetic fixtures and test accounts for permissions, stale data, prompt injection, malformed output, duplicates, partial failures, spend overflow, approval expiry, tenant escape, and rollback.
- Separate analysis from execution. Models propose typed change-sets; a tenant-scoped service enforces account/entity allowlists,
PAUSED, spend caps, idempotency, exact signed approvals, audit logs, and rate limits. - Release through a capped canary. Start with the smallest reversible production slice that meets the approved risk criteria, preserve the prior configuration, and prohibit unrelated accounts or entities.
- Observe predefined signals. Monitor continuously for policy violations, cost, duplicates, quality, latency, and business guardrails—not merely for a fixed number of weeks.
- Stop and roll back automatically. A kill switch and tested rollback must work without the model. Keep manual platform access and credential revocation available.
- Promote only with evidence and approval. Re-run evals and contract tests for every model, prompt, policy, schema, API-version, or credential-scope change.
A bug could create duplicates or misstate a budget. Prevent that structurally: typed schemas, maximum-count and spend constraints, PAUSED-only creation, account allowlists, idempotency keys, exact approvals, audit logs, canaries, monitoring, and automatic rollback. Prompt instructions alone are not controls.
AI is a tool of extraordinary power. Like any powerful tool, it requires respect, care, and responsible use. The framework in this document gives you the knowledge — but the responsibility is yours.
Built by @itsmazinzaki — AVAMARTECH
Educational review — 31 August 2026