Note on authorship. This article was written with AI assistance. Obviously. That’s literally what this article is about. The irony is intentional. The prompt that generated this was informed by months of context. Keep reading.
The Visible Tip#
When someone reads “this blog post was AI-assisted,” they often imagine:
Me: "write a blog post about authentication"
AI: *generates 2000 words*
Me: *clicks publish*And they think: “Well, that’s not really your work, is it?”
They’re seeing the tip of the iceberg—the visible prompt, the generated text, the published post.
What they’re missing is the 95% underwater: the weeks or months of context that made that prompt possible.
The Invisible Infrastructure#
Here’s what actually happened before I wrote my recent blog posts on AI agent authentication:
Weeks 1-4: The Struggle#
- Implemented OAuth Token Exchange for our multi-agent system
- Hit edge cases that didn’t make sense
- Argued with teammates about why it felt wrong
- Read RFC 8693 multiple times, increasingly frustrated
- Built comprehensive documentation trying to prove inadequacy
Weeks 5-8: The Research#
- Discovered Christian Posta’s work on agent identity
- Found the expired IETF Power of Attorney draft
- Watched the dismissive IETF committee response (painfully)
- Started building my own PoA delegation model
- Iterated through dozens of architecture variations
Weeks 9-12: The Context Capture#
- Documented decisions in team_home (more on this below)
- Wrote technical implementation details in RedBank demo
- Captured frustrations in CONSTITUTION.md files
- Built test scripts and validation frameworks
- Created diagrams, flowcharts, comparison matrices
Week 13: The Blog Post#
Me: "Let's write a blog post. Last week I tried to figure out..."
AI: *reads all the context above*
AI: *generates coherent 2500-word argument*
Me: *edits, refines, publishes*
Time: 30 minutes instead of 8 hoursWhat people see: “5-line prompt → blog post”
What actually happened: “3 months of work → compressed into retrievable context → 5-line prompt → blog post”
The AI didn’t create the ideas. It surfaced, structured, and articulated what was already there, buried in weeks of docs, code, and frustrated Slack messages.
The Team Home Advantage#
This wouldn’t work without infrastructure. Specifically, team_home—a project created by my manager, Dimitri Saridakis.
Team_home is a “context engineering” repository. It’s not just a collection of docs. It’s a deliberate knowledge architecture that makes context retrievable.
Here’s the structure:
team_home/
├── jira_process/ # OBSERVE: Process health + context
│ ├── context/ # Jira conventions, org structure
│ ├── reports/ # Health analysis, insights
│ └── data/ # Snapshots for reproducibility
├── team_workspace/ # ACT: Planning, execution
│ ├── planning/ # Feature breakdowns, decisions
│ ├── skills/ # How Claude operates
│ └── context/ # Process targets, team info
└── product/ # BUILD: Actual code repositories
├── redbank-demo-2/ # Implementation details
├── kagenti/ # Architecture decisions
└── akram.github.io/ # This blogWhen I work in this environment, everything I do builds context:
- Fix a bug? The commit message captures the “why”
- Make a decision? It goes in CONSTITUTION.md with reasoning
- Hit a blocker? Documented in docs/DEPLOYMENT_ISSUES_AND_SOLUTIONS.md
- Learn something? Added to context/ with “when to use this”
- Ship a feature? README.md explains the architecture
Six weeks later, when I write a blog post, the AI has access to all of this.
It’s not magic. It’s organized memory.
The Actual Workflow#
Let me show you the real process for my last blog post:
Step 1: The Prompt (5 minutes)
Me: "Let's write a blog post. Last week I tried to figure out
what could be the state of the art for Authentication and Access
Control for AI Agents... [3 paragraphs of context]"Step 2: The AI Response (2 minutes) The AI reads:
team_home/product/redbank-demo-2/docs/TOKEN_EXCHANGE_SETUP.mdteam_home/product/redbank-demo-2/docs/TOKEN_EXCHANGE_DEBUG.mdteam_home/product/redbank-demo-2/docs/DEPLOYMENT_ISSUES_AND_SOLUTIONS.md- My commit history from the past month
- CONSTITUTION.md files explaining architectural decisions
- The PoA delegation model I’d been iterating on
It generates a structured outline.
Step 3: Refinement (15 minutes)
Me: "Add a section on the IETF presentation dismissal.
Include the YouTube timestamp."
Me: "Emphasize the 'square wheels' metaphor more."
Me: "The conclusion needs to be more actionable."Step 4: Review & Edit (10 minutes) I read the full draft, fix technical inaccuracies, adjust tone, ensure it actually reflects my thinking (not generic AI voice).
Total time: ~30 minutes
Equivalent manual writing time: 6-8 hours
Speedup: ~20x
But here’s the key: I couldn’t have written a better post manually.
Why? Because the AI has perfect recall of everything in the context. I would have forgotten half the details, missed connections between concepts, and spent hours digging through my own notes.
What This Isn’t#
Let me be clear about what this is not:
❌ Lazy - Building retrievable context is harder than writing ad-hoc
❌ Inauthentic - The ideas are mine; the AI just articulates them
❌ Low-effort - The effort went into the preceding weeks of work
❌ Generic - The context makes it specific to my experience
❌ Unedited - I review and refine every generated paragraph
What This Is#
This is a fundamentally different way of working with knowledge:
✅ Front-load the effort - Build good context infrastructure
✅ Capture as you go - Document decisions, reasoning, failures
✅ Retrieve on demand - Generate artifacts from accumulated knowledge
✅ Iterate quickly - Explore ideas 20x faster
✅ Stay honest - The context ensures authenticity
It’s the difference between:
- Traditional: Think → Write → Publish (8 hours)
- AI-assisted: Think (weeks) → Capture (daily) → Prompt (5 min) → Refine (30 min) → Publish
The thinking still takes weeks. The writing takes minutes.
The Authenticity Question#
“But is it really your writing if AI generated it?”
Counter-question: Is it really your thinking if you typed it yourself?
What matters more:
- The origin of the ideas (definitely mine)
- The accuracy of the content (reviewed and verified)
- The value to readers (hopefully high)
- The honesty of attribution (clearly stated)
Or:
- Whether I personally pressed the keys that formed the sentences
I’d argue the former. The mechanical act of typing is the least interesting part of writing.
The hard parts are:
- Having experiences worth writing about ✅
- Developing opinions through struggle ✅
- Building context that makes insights possible ✅
- Verifying accuracy and coherence ✅
- Taking responsibility for what’s published ✅
AI helps with: Structuring arguments, finding clear phrasing, ensuring logical flow.
I still do everything else.
The Context Infrastructure Tax#
Here’s the catch: this only works if you build the infrastructure.
Random ChatGPT prompts don’t produce quality because there’s no context. You get generic, surface-level output because the AI has no depth to draw from.
Team_home gives me:
- ✅ Organized memory - Context structured for retrieval
- ✅ Captured reasoning - Not just what, but why
- ✅ Linked knowledge - Connections between concepts
- ✅ Process documentation - How we got here
- ✅ Historical depth - Months of accumulated work
This didn’t happen by accident. Dimitri designed it deliberately.
The infrastructure tax is real:
- More discipline in documentation
- Structured directory semantics
- Consistent commit messages
- CONSTITUTION.md files explaining decisions
- Context/ directories with “when to use this”
But the payoff is extraordinary:
- Blog posts in 30 minutes instead of 8 hours
- Technical docs that write themselves from code context
- Decisions that reference prior reasoning automatically
- New team members onboard from reading context, not asking questions
Why This Feels Refreshing#
Traditional blogging incentivizes:
- Short context - Can’t assume readers remember last post
- Self-contained - Every post starts from scratch
- Broad appeal - Dilute specifics to reach more people
- Slow cadence - Writing takes too long to publish frequently
AI-assisted blogging with rich context enables:
- Deep context - Build on weeks of accumulated knowledge
- Connected ideas - Reference and build on prior work
- Specific insights - Dive deep on narrow topics quickly
- Fast iteration - Publish when ideas are fresh, not months later
I can write a 2500-word deep-dive on token exchange architecture because I’ve already done the work. The blog post is just making it accessible.
Without AI assistance, that post would never get written. The effort-to-publish ratio would be too high.
The Inference From Context#
The title mentioned “inference from context.” Let’s unpack that.
When I write:
"Last week I tried to figure out the state of the art for
AI agent authentication..."The AI infers from surrounding context:
- I’ve been working on RedBank demo (product/redbank-demo-2/)
- I hit OAuth Token Exchange limitations (docs/TOKEN_EXCHANGE_SETUP.md)
- I discovered the PoA IETF draft (mentioned in commits)
- I watched the dismissive presentation (YouTube link in notes)
- I’m frustrated with the industry (evident from doc tone)
Inference = Taking sparse input + rich context → generating coherent output that reflects both
This is the 2026 way of working with AI. Not “generate from scratch,” but “infer from everything I’ve already captured.”
The 20x Multiplier#
People ask: “Is it really 20x faster?”
Let’s do the math:
Manual blog post:
- Research and gather notes: 2 hours
- Outline and structure: 1 hour
- First draft: 3 hours
- Edit and refine: 1.5 hours
- Final review: 0.5 hours
- Total: 8 hours
AI-assisted with team_home context:
- Prompt with context: 5 minutes
- Review and refine structure: 10 minutes
- Edit for accuracy and tone: 15 minutes
- Final review: 5 minutes
- Total: 35 minutes
Math: 8 hours / 35 minutes = 13.7x speedup
I rounded to 20x because I’m accounting for:
- Higher quality output (better organized)
- More ideas explored (I can try 3 angles in the time I’d write 1)
- Less writer’s block (context provides momentum)
It’s not just faster. It’s better.
The Meta Irony#
This blog post about AI-assisted blogging was:
- ✅ AI-assisted
- ✅ Based on months of team_home context
- ✅ Generated in ~30 minutes
- ✅ Edited by me for accuracy
- ✅ Published with full disclosure
The prompt I gave referenced:
- My previous blog posts (in
product/akram.github.io/content/posts/) - The team_home structure (documented in multiple README.md files)
- Dimitri’s role (mentioned in
team_workspace/context/team.md) - The infrastructure tax concept (lived experience documented in commits)
The AI didn’t invent any of this. It synthesized what was already there.
That’s the point.
Why You Should Try This#
If you’re writing technical content, consider:
- Build a context repository - Even a simple docs/ folder helps
- Document decisions as you make them - Don’t wait for “documentation time”
- Capture reasoning, not just facts - Why did we choose X over Y?
- Link related concepts - Create a knowledge graph, not isolated notes
- Use AI to surface, not create - Let it find connections you missed
You’ll find that:
- ✅ Writing becomes synthesis instead of creation from scratch
- ✅ Ideas stay fresh because you can publish quickly
- ✅ Quality improves because context ensures accuracy
- ✅ You publish more because the friction is lower
The Future of Technical Writing#
I think we’re at the beginning of a shift:
Old paradigm:
Individual expertise → Manual writing → Slow publishing → Limited reach
New paradigm:
Team context infrastructure → AI-assisted synthesis → Fast publishing → Compounding knowledge
The bottleneck shifts from writing to experiencing and documenting.
Which is good! Because experiencing and documenting are the valuable parts. The mechanical act of writing was always just a necessary evil.
Now it’s optional.
Conclusion#
When you read an AI-assisted blog post, you’re not reading “what AI thinks.”
You’re reading:
- Weeks of context compressed into minutes
- Documented struggles made digestible
- Accumulated insights surfaced and structured
- Human experience articulated by machine
The AI is the medium, not the source.
The value comes from having something worth saying and the infrastructure to say it efficiently.
Team_home gives me the infrastructure.
AI gives me the efficiency.
The experiences and ideas? Those are still mine.
And honestly, this feels more authentic than manually typing every word while forgetting half the context and taking 8 hours to finish.
The bottom line: You see a 5-line prompt. You don’t see the 12 weeks of work that made it meaningful. That’s the context iceberg. And it’s why AI-assisted blogging might be the most honest way to share technical knowledge in 2026.
P.S. - This post took 28 minutes to generate and edit. The context that made it possible took months. That’s exactly the point.
Acknowledgments#
Team_home was created by Dimitri Saridakis, my manager at Red Hat. The project provides the context infrastructure that makes this workflow possible. It’s an example of “process-as-code” done right—where the methodology becomes retrievable knowledge instead of tribal lore.