Beyond Chat
Building Agentic Workflows with Claude Code
Most AI usage stops at “ask a question, get an answer.” What happens when you treat the AI as a team member with persistent context, codified judgment, and operational autonomy?
A walkthrough of a real production system — not a demo
Every Session Starts at Zero
You open a new session. The AI knows nothing about your project.
You spend 10 minutes re-explaining conventions, file locations, and preferences.
It generates something generic. You fix it manually. Repeat tomorrow.
The real cost isn’t the AI subscription. It’s the re-onboarding tax you pay every single session — and the drift that happens when context is lost.
Chat vs. Agentic: What’s the Difference?
| Dimension | Chat Pattern | Agentic Pattern |
|---|---|---|
| Context | Starts from zero each session | Loads project config automatically |
| Knowledge | Generic best practices | Your conventions, your patterns |
| Workflow | Single Q&A exchange | Multi-step with tools and side effects |
| Quality Control | You review everything manually | Codified standards enforce themselves |
| Trust | Verify every output | Permission envelope grows over time |
The key insight: An agentic workflow isn’t smarter AI. It’s the same AI with better infrastructure around it — memory, judgment, tools, and trust boundaries.
Five Layers of an Agentic System
Each layer solves a specific problem. Together, they compound.
Project memory that loads automatically
Quality standards the AI enforces
Chain models by their strengths
Permissions that grow incrementally
Same system, different entry points
None of this requires a custom framework. It’s configuration files, markdown documents, shell scripts, and CI/CD. The tools already exist — the insight is in how you compose them.
Session Bootstrap: CLAUDE.md
A single markdown file that loads at the start of every session. The AI arrives pre-configured.
What goes in it
→ Directory map with purpose annotations
→ Complete workflow templates (step-by-step)
→ Frontmatter schemas and conventions
→ Off-limits files requiring explicit approval
→ Common commands and build steps
project/
├── astro-build/
│ ├── src/content/blog/
│ ├── public/images/
│ └── scripts/
├── docs/ # Voice guide
└── .claude/ # This config
Think of it as onboarding docs for a team member who has perfect recall but joins fresh every morning. Invest 30 minutes once — it pays off in every session.
Slash Commands: Mode-Switching
Custom commands that activate specific operational modes. Same AI, different posture.
New Content Mode
Reads the voice guide first, then walks through: identify the hook, choose a structure template, draft following patterns, create frontmatter, self-review against checklist.
Refinement Mode
Reviews against voice authenticity, structural integrity, and tonal balance. Critically: includes explicit “What NOT to Fix” instructions — preserve rough edges, don’t add polish.
Audit Mode
Scores on three dimensions (1-10 each): Voice Authenticity, Structural Patterns, Tonal Consistency. Flags specific red-flag phrases.
Quick Screen
Five yes/no questions. Returns a binary: SOUNDS LIKE NINO: Yes/No/Mostly. Takes 30 seconds. Used for fast iteration.
Each command is a short markdown file — 15-30 lines. The power isn’t in complexity. It’s in giving the AI a clear role to inhabit.
Codified Judgment: The Voice Guide
A 156-post empirical analysis turned into a living editorial standard. Not aspirational — descriptive.
What it captures
- →Voice dimensions: Public practice, meta-awareness, vulnerable competence, pattern recognition
- →Structure templates: Reflection, Technical Deep-Dive, Leadership Insight, Origin Story
- →Anti-patterns: Retired phrases that became “tells” through overuse
- →Formatting rules: Paragraph length, divider usage, component limits
The Cardinal Rule
”This guide describes patterns, not templates. If you copy phrases from this document, you’re using it wrong.”
Every example is evidence, not instruction. The AI should understand the spirit, then find its own words.
The “What NOT to Fix” list is the most valuable part. Any AI can add polish. The hard part is teaching it to leave the rough edges that make writing feel human.
Freshness Tracking: Preventing Staleness
Even authentic phrases become hollow through repetition. The guide tracks recency and enforces cooldowns.
| Pattern / Phrase | Last Used | Cool Until |
|---|---|---|
| ”Here’s where my head is” | Jan 21 | Feb 21 |
| Opening with “How many…” question | Jan 21 | Feb 21 |
| Graveyard/death metaphor | Jan 21 | Mar 21 |
Any phrase appearing 3+ times in six months gets permanently retired. No exceptions.
”Ask me again in six months” — “Here’s where I’ve landed, for now” — “This is what I think today”
Multi-Stage Pipelines: Image Generation
One model thinks. Another model draws. Anti-cliche guardrails sit between them.
Gemini Flash
Reads the full post. Generates a specific, non-generic visual concept.
Prompt explicitly rejects: “no robots, lightbulbs, handshakes, puzzle pieces, or generic tech imagery”
GPT-5 Image
Takes the concept, applies category-specific style profile. Generates 1200x675 illustration.
10 visual styles mapped to content categories
Sharp + Frontmatter
Compresses to WebP at 85% quality. Auto-updates the post’s frontmatter with the image path.
No manual steps. Post references the image automatically.
The principle: Chain models by their strengths, not by vendor. One model for reasoning, another for generation, with your quality standards as the glue between them.
Category-Specific Visual Identity
Each content category has a distinct illustration style. The AI doesn’t choose — the system assigns.
Electric cyan circuit patterns, flowing data streams
Golden amber, journal marginalia aesthetic
Architectural linework — lighthouses, bridges
Organic network diagrams, mycelium patterns
Escher-like impossible geometries
Whiteboard energy, sketch-in-progress
Visual consistency at scale. Every post gets a unique illustration, but they all feel like they belong to the same publication. The AI handles uniqueness; the system handles coherence.
Trust Boundaries: The Permission Envelope
Autonomy isn’t granted wholesale. It’s earned incrementally through accumulated permissions.
Approved (170+ permissions)
✓ Git operations (add, commit, push, branch)
✓ npm build pipeline (dev, build, preview)
✓ Node automation scripts
✓ Browser automation (Playwright, DevTools)
✓ WebFetch for curated domains only
Off-Limits (Explicit Approval)
✗ .env files (secrets, API keys)
✗ Build configuration (astro.config.mjs)
✗ Unapproved external domains
✗ Destructive git operations
The permission list is a living document that grows through use. Each approval says: “I trust you to do this without asking.” Over time, the agent operates faster because the trust is pre-established.
Multiple Pathways: Same System, Four Entry Points
The same content system, the same quality standards — accessible from anywhere.
Claude Code Session
CLAUDE.md loads → slash commands activate modes → voice guide enforces quality → scripts generate images → git push deploys to Vercel
GitHub Issue Form
Structured issue template → GitHub Actions triggers → Claude API formats MDX → auto-commits → Mailgun cross-posts to Substack
Draft Preview UI
View draft on site → click “Generate Image” → API fires repository dispatch → CI generates image → Vercel redeploys
Visual Audit
Playwright/Chrome DevTools MCP → inspect live site at breakpoints → verify rendered output → audit visual design
The voice guide is the connective tissue. It doesn’t matter which pathway triggers content creation — the same quality standards apply regardless of entry point.
Publish From Your Phone
A structured issue form feeds directly into a CI/CD pipeline. No local dev environment needed.
The insight: Content from any source — a voice note transcribed by Gemini, a draft from your phone’s notes app, an idea at 2am — can enter the same quality pipeline.
Six Content Types, One Provenance System
Every piece of content can trace its lineage. Research notes link to posts. Posts link to whitepapers. Counterpoints challenge posts.
Conversational exploration. MDX components. Voice-guide enforced.
Formal analysis. Data tables. Plain markdown only.
Slide-based delivery. Exportable as standalone HTML.
Adversarial critiques of existing posts. AI-generated or self-authored.
Working documents. Red-team analyses, lit reviews, methodology.
Groups related posts into multi-part arcs.
Bidirectional linking: Posts declare supportedBy (which research backs this?). Research notes declare supportsContent (which posts use this?). The graph is navigable in both directions.
What Actually Worked
High ROI
- →CLAUDE.md — 30 minutes to write, saves 10 min per session forever
- →“What NOT to Fix” — The single most impactful editorial instruction
- →Slash commands — 15 lines of markdown creates a mode the AI inhabits
- →Freshness tracking — Prevents the staleness that makes AI writing feel templated
Honest Surprises
- →Permissions accumulate fast — 170+ entries. Needs periodic review.
- →Voice guide maintenance — It’s a living document, not write-once. Requires updating after every ~5 posts.
- →Multi-model pipelines are fragile — When one API changes, the whole chain breaks.
- →The counterpoints system — AI critiquing its own output is genuinely useful, not gimmick.
What You Can Do Monday Morning
You don’t need the full system. Start with Layer 1 and build up.
Write a CLAUDE.md (or equivalent) for your project
Directory map, conventions, workflow steps. Takes 30 minutes. Pays off immediately.
Create one slash command for your most repetitive task
Code review, PR description, ticket grooming — whatever you do weekly. 15 lines of markdown.
Document what the AI should NOT do
The most valuable instruction isn’t “do this.” It’s “don’t touch that.” Preserve what’s already working.
Let the permission list grow organically
Don’t pre-approve everything. Let the AI ask, approve what makes sense, and the trust boundary forms naturally.
The Bigger Picture
The gap between “using AI” and “working with AI” is infrastructure, not intelligence. The models are already capable. The question is whether your systems let them remember, enforce standards, use tools, and operate within trust boundaries.
Config files that form the system
Custom frameworks required
Posts through this system
Signal Dispatch | February 2026