7 posts tagged with "ai"

Two years ago seems like decades in the world of AI. AI slop wasn't a term I was familiar with, and the project I started with AI was meant to be short-lived—a stopgap until systems and people caught up to support the work.

Stage 1 - time-savings over process design​

Because it was meant to be a short-term solution, I didn't care about the what or how as much as the output and time savings. I didn't spend much time on it and certainly didn't architect it in any meaningful way. It was a markdown file - a prompt.

Back then, there was little concern for context size and the only harnesses were custom-made or 3rd party like LangChain.

It read source code and output what I asked it to based on my ever-growing prompt file. I fiddled with it when I had time but usually just took the output and made manual edits as needed.

Stage 2 - shifting sands of source code​

There were a few bumps in the road that caused the prompt file to sputter and halt. The related source code repo moved. That was just one place in the prompt file - an easy change. Then the source code repo completely refactored into a monorepo - that was much harder. The prompt file was looking at specific folders, files, classes and property names. A much bigger change to the prompt file. And of course, I still had other things to do.

Then the source project became much more popular and the feature set grew. Lots of changes and additions. This resulted in a short cycle of my wondering if AI, or my prompt file needed another rewrite.

At this point, I thought regex and a script might be faster. Leave any LLM decision/transition until later.

This was becoming an experiment in how to gain efficiency in a rapidly changing AI world.

Stage 3 - back to reality​

While I was considering shifting to a code-based solutions, the source repo created a CLI to reveal its features. That was exactly what I needed: a dependable input to my process.

The AI SDKs had also progressed, so instead of regex and a script, I switched to .NET and the OpenAI SDK for chat completions. This time I designed the process deliberately, and Copilot Chat in VS Code was available, which helped accelerate development.

This time the architecture, project structure, maintainability, and testability were top considerations as I was still relying on this thing to help me.

There were a few iterations of this but it was beginning to hum along.

Stage 4 - the infrastructure​

Since I only had the OpenAI resource in Azure, I kept it live and never shut it down. However, I used keys instead of managed identity—not because I didn't know the best practices, but simply because I lacked the time to implement them.

I was doing a lot of Bicep work, not an expert but not a newbie either, so cooking up the ./infra and azure.yml for Azure Developer CLI wasn't hard. Adding the RBAC roles for managed identity helped. And then finally I had that repeatable infrastructure I told everyone else they should have.

Stage 5 - monkey on the keyboard​

The underlying project repo had matured significantly, the source code had stabilized, and my system was working well for its purpose too. However, as AI slop became an industry term, the underlying project matured alongside the industry specs it supported but my project had not.

I managed the project in chat one turn at a time, and the .NET app generated the output. However, every change to the system required several to many turns, redirections, and re-runs. AI kept making the same mistakes, and I spent my time minute by minute watching the output to see when it went off track. I needed a better process—a process to manage the process.

Stage 6 - hiring a squad​

It was around this time I started working with Brady Gaster's Squad. Having a team of agents accelerated progress and caught issues and inconsistencies I hadn't been able to spot before. I onboarded Squad to the repo and had them take on much of the work.

Stage 7 - industry grows up​

While I was learning Squad's inner workings, SKILL.md and agent.md specifications emerged. I dug into skills first to impose structure on my manual process management. I used WAZA and other frameworks to understand context size, triggers, and evaluations.

Building and maintaining skills locally gave me practical understanding of task boundaries, debugging, and chaining. Working with these units of AI many times a week gave me opportunities to fail.

Stage 8 - leaving Squad behind​

Squad is an amazing build and maintenance platform. However, I need skills that succeed independently. The plugin specification is now available, so I took my chained skills, wrapped them in an orchestrating agent to manage gates between skills and human review, and packaged everything as a plugin.

Stage 9 - the agent and the app​

What about the .NET app? It's now used only when the agent determines it needs that full pipeline. The agent and skills handle most of the work.

Conclusion​

You'll likely continue to hear a lot about AI slop. Just like any industry, AI is maturing and developing better specifications and expectations. Look for them and try them out.

Focus on the fundamentals available at the time. Build with principles and practices that support them. Grow with industry specifications that enable better results.

You've built a workflow that works. But automating it requires engineering. Now you're stuck.

You know exactly how the process should run—you do it repeatedly, correctly, and it saves your team real time. But the moment you try to automate it, you hit a wall: write a script (and maintain it forever), hire engineers (and lose control to the project backlog), or give up and do it manually. None of these are wins.

There's a better way. What if you could capture your workflow in plain language, test it out, refine it, and gradually promote it to automation—keeping ownership the whole time? No developers required until you're absolutely sure the workflow is stable.

This is the progressive promotion model. Domain expertise starts as a natural language skill. You run it, refine it, test it. When repeated correct outcomes prove the flow is deterministic, the stable parts move behind MCP tools—reusable logic anyone can call. Then the skill can run autonomously. Ownership stays with the person who understands the process the whole time.

The automation tradeoff changed​

The old tradeoff was simple: business value or engineering time. Only high-value workflows got built. Mid-tier work died in the backlog because no developer had bandwidth. The domain expert had to describe the process, hand it to a team, and wait months for the software to arrive.

This model shifts that cost to zero. The domain expert writes and runs the workflow in a skill right now, while they work. When the process proves reliable through repeated correct outcomes, the stable parts move behind MCP tools without a full rewrite. Cached tool definitions and typed contracts mean even small workflows can graduate.

The payoff is ownership. You hone your own process, keep control of your decisions, and run it yourself while it matures.

The four layers​

Here is the path at a high level. The model puts each concern in its own layer. MCP stands for Model Context Protocol. In this post, an MCP tool is the typed interface that lets a skill, agent, or CI job call code in a predictable way:

The script is the logic. The MCP tool wraps it in a typed interface. The skill decides when to call which MCP tools. The agent runs the skill without you. Each layer has one job.

The key move: pull the script out of the skill and put it behind the MCP tool. Now any consumer can call it—another skill, an agent, a CI pipeline, an external system. The script is no longer locked inside one skill.

The resulting stack looks like this:

The script gets written once, wrapped in a typed tool once, and then only the driver changes during promotion from interactive to autonomous.

How work naturally evolves​

Start in natural language. Let the domain expert hone the process. Promote only after repeated correct outcomes prove the flow.

Here's how that progression works in practice:

A new skill starts with the LLM doing everything inline. Your instructions might say "query the GitHub API for recent releases, then compare against our changelog." The first version is written in plain language, not code. You stay in control.

Correctness matters more than speed here. You run the skill, adjust the instructions, and decide if the outcome matches your judgment. Repeat it several times until it consistently produces the right result.

Phase B: Determinism emerges​

After a few runs, you spot a pattern. Step 2 is always the same. Same API call, same parsing, same output format. The LLM isn't adding judgment here—it's just following a mechanical procedure that you've already validated.

This is your signal to move. When the same API calls and parsing steps keep showing up, and the outcomes have been consistently correct, that part is ready to extract.

Phase C: Extract to MCP (not script-in-skill)​

Now make the move. Extract the deterministic logic into a typed MCP tool instead of keeping it inside the skill. You still control the workflow through the skill. The stable, reusable part moves behind a typed interface.

The skill now calls detect_releases instead of embedding the logic. The MCP tool has a JSON input schema, a JSON output schema, and error handling. It's independently testable. Any consumer—another skill, an agent, a CI pipeline, or an external system—can call it.

Phase D: Promote to agent​

When the process is reliable and you want it to run without you, promote to autonomous execution. The agent uses the same MCP tools. The only difference is who drives: you (interactive) or the agent (autonomous).

The MCP server does not change. The tools do not change. The scripts do not change. Only the driver changes.

Where my first design stopped​

My forty-skill portfolio showed me where this breaks if you stop too early. My original approach was:

1. Write a skill (natural language instructions)
2. Notice a pattern is deterministic (same inputs → same outputs)
3. Extract that logic into a script inside the skill
4. Done

The problem is step 4. The skill works. The script works. But only that skill can use it. No other skill can call it. No agent can use it. No CI pipeline can run it. When you need that logic elsewhere, you copy-paste the whole thing.

After forty skills, I had forty pieces of scattered process knowledge with scripts locked inside individual skills, no typed contracts, no reusability, and no clear path to autonomous execution. The scripts weren't going anywhere.

Why MCP tools instead of scripts-in-skills​

The decision to extract into MCP rather than keep scripts inside skills comes down to three things:

Reusability​

A script locked inside one skill is only callable by that skill. An MCP tool is callable by any skill, any agent, any CI pipeline, and any external system. Reuse changes everything.

Typed contracts​

A script takes string arguments. An MCP tool has a JSON input schema and JSON output schema. The LLM knows exactly what to send and what to expect back. No parsing surprises.

Prompt caching​

The cost reason is direct: MCP tool definitions live in the system prompt and get cached at a 50-90% discount. Every time you spawn an agent fresh, you lose that cache.

Using MCP tools instead of spawning fresh agents cut uncached tokens by roughly 90%.

The decision point​

When you find yourself writing a script inside a skill, ask one question:

Will anything other than this skill ever need to call this logic?

• If yes → extract to MCP immediately
• If maybe someday → extract to MCP (future reuse is cheaper than a later move)
• If truly never (one-off, will be deleted soon) → script-in-skill is fine

In my forty-skill portfolio, the answer was almost always yes.

What promotion looks like in practice​

I have a content pipeline called Echo that detects new SDK releases, generates documentation metadata, and produces content reports. It started as a Squad agent spawning fresh context every time.

After extracting to MCP + skill:

The scripts themselves didn't change. The structured JSON output envelopes they produced already matched MCP tool responses—same schema, different transport.

Before: Script → JSON file → next skill reads file from disk
After: Script → JSON → MCP protocol → any consumer gets it directly

The cost model across stages​

Each stage changes the driver but reuses the same tools. Costs drop because the driver changes:

As work matures, it needs less LLM reasoning per run, until CI/CD needs none at all.

Context occupation cost of MCP​

MCP tools have lower per-token cost when cached, but they occupy context window space every turn, even when unused. A 4-tool server adds ~600-1600 tokens to every conversation.

Control this with grouping and toggling:

The pattern: skill triggers workflow (zero idle cost) → skill activates MCP (cost only when needed) → tools do work (cached calls). The MCP stays enabled only when you're using it.

Running autonomously with Copilot CLI​

Once a skill is promoted to an agent, the next need is running it without a human session. Copilot CLI supports this today:

# Simplest autonomous run
copilot -p "Run the echo pipeline" --yolo --silent

# With specific agent and model
copilot -p "Execute" \
  --agent echo-pipeline \
  --autopilot --no-ask-user \
  --yolo --silent \
  --model gpt-5.4

# Sealed sandbox: only specific tools available
copilot -p "Sync releases" \
  --additional-mcp-config @workflows/echo-sync/mcp-config.json \
  --available-tools='content-pipeline-mcp/*' \
  --no-ask-user --autopilot --silent

The key flags:

For CI/CD, authenticate with a fine-grained PAT:

COPILOT_GITHUB_TOKEN=github_pat_xxx copilot -p "Run pipeline" \
  --agent echo-pipeline --yolo --silent --no-auto-update

The sealed sandbox​

When something runs autonomously, the context must be fully specified at launch and locked in place. The agent gets exactly the tools it needs and nothing extra. The glass bell jar is boring on purpose.

A sealed sandbox manifest specifies:

1. Identity: who the agent is
2. Available tools: exhaustive list—nothing else exists
3. Execution plan: exact steps, no deviation
4. Error handling: complete rules, no improvisation
5. Output routing: where results go
6. Boundaries: hard constraints (violation = immediate exit)

Copilot CLI does this through --available-tools and the "tools" allowlist in MCP config. The MCP server might have twenty tools. The agent only sees three.

{
  "mcpServers": {
    "content-pipeline": {
      "command": "node",
      "args": ["./mcp-servers/content-pipeline/index.js"],
      "tools": ["detect_releases", "generate_metadata", "analyze_impact"]
    }
  }
}

Related patterns​

Similar progressions appear in other domains under different names:

The pattern exists in pieces across many domains. What was missing: a practical "Skill → MCP → Agent → CI" progression with extraction checklists and validation gates, tailored specifically for domain experts who want to keep ownership.

The rule I use now​

When a process proves repeatable, reusable logic moves to an MCP tool. Scripts-in-skills are prototype code.

Ask three questions: Who owns this? How stable is it? What driver does it need now?

What's next​

Echo is the pilot. Once the content-pipeline MCP server wraps Echo's three scripts and the /echo-sync skill drives them interactively, validation has two parts: the process still produces the right output, and the token savings are real.

Then come Finn, the reporting tools, and the rest one by one.

The pattern is straightforward: Start with the person who owns the domain knowledge. Capture the workflow in a skill. Run it until the outcomes are consistently correct. Move the repeatable parts behind MCP tools. Change the driver only when autonomy helps.

Ownership stays with the person who understands the process. The system matures around that expertise.

Every session now leaves behind a timestamped note I can review, audit, or delete. No cloud sync. No hidden database. Just files I own.

I built the missing memory half from my personal context blog post.

In Portable Personal Context Across AI Client Surfaces, I separated personal context from memory. Context is the small set of reviewed facts I want tools to trust. Memory is what happened while I worked.

The gap was capture. I needed a way for Copilot CLI sessions to leave behind structured observations I could review later. So I built copilot-cli-log-to-file: a Copilot CLI extension that writes each finished turn to a timestamped file I own.

The result: a reproducible, human-controlled pipeline from raw evidence to trusted rules—without accidental promotions.

Capture the memory feed first​

The stronger angle is memory feed, not cross-computer sync.

Sync is useful, but it comes later. What I needed first was a raw feed: timestamped session evidence that could become an observation in the pipeline.

observation  →  candidate  →  [ratification gate]  →  context
 (memory)       (proposed)     (a human decision)      (canonical)

A capture file is not context. It can prove that I asked a question, that Copilot answered, and that a tool ran if I opted into tool capture. It should not silently become an instruction. I still want the human gate because one weird session should not rewrite my operating rules.

Extension output​

copilot-cli-log-to-file runs after a Copilot CLI turn finishes. By default, it writes a complete YAML document with timestamp, sessionId, prompt, and response.

The default filename pattern is {timestamp}-{prompt30}.yaml, so a log folder looks like this:

copilot-response-log/
  2026-07-13T11-21-45Z-list-all-my-files.yaml
  2026-07-13T14-05-02Z-explain-this-function.yaml

That filename format mattered more than I expected. I can sort by time, skim the prompt slug, and delete one turn without touching the rest of the history.

I kept enriched capture off by default. If I want more detail, I opt in by category: attachments, reasoning, tool calls, tool results, usage, model changes, skills, subagents, permissions, errors, lifecycle events, turn boundaries, schedules, or notifications. There is also a COPILOT_LOG_CAPTURE_ALL=true override, but I treat that as a deliberate choice because these files can hold private data.

Here is a small capture with usage and tools enabled:

timestamp: "2026-07-14T18:21:45.000Z"
sessionId: "9f4c2a7b12345678"
prompt: |-
  Summarize the current branch and suggest the next test to run.
tools:
  - toolCallId: "call_abc123"
    toolName: "git"
    arguments: '{"command":"status --short"}'
    success: true
    result: "M website/blog/2026-07-17-capture-layer-for-portable-context.md"
usage:
  - model: "gpt-5"
    inputTokens: 1840
    outputTokens: 420
    cacheReadTokens: 600
    cacheWriteTokens: 0
    duration: 1310
    finishReason: "stop"
response: |-
  The branch has one modified blog post. The next targeted validation is to re-read
  the file for frontmatter, link, and YAML-example accuracy before committing.

The toolCalls and toolResults toggles feed one merged tools: block. That was a small design choice, but it keeps the question I care about in one place: what tool ran, with which arguments, and what happened.

Two implementation details saved me from future debugging. Copilot CLI reserves stdout for JSON-RPC, so I use session.log(...) instead of console.log(...). The runtime also provides @github/copilot-sdk; I do not ship it as an npm runtime dependency. js-yaml stays dev-only for tests because the extension has a hand-rolled YAML emitter.

The capture layer transforms a CLI turn into structured data. A turn contains two key outputs: the assistant response (Copilot's answer) and events (the structured record of what happened—tool calls, model usage, and other metadata). Both flow into the timestamped YAML file.

Promote observations only after review​

The previous blog post's pipeline feels less abstract now that I have real files on disk.

copilot-response-log/*.yaml
        ↓
observation
        ↓
candidate
        ↓
[ratification gate]
        ↓
portable personal context repo

The YAML file is the observation. A candidate is the proposed durable fact I might extract from one or more observations. The context repo is where the approved rule lives.

For example, a few sessions might suggest that I prefer targeted validation before full-suite validation for docs-only work. That is still only evidence. I decide whether it belongs in process/quality-bar.md, decisions/_active.md, or nowhere.

Boring is good here. The capture layer does not promote anything by itself. That keeps mistakes, one-off exceptions, and prompt-injection garbage out of my canonical context unless I approve them.

Keep the feed in files I can inspect​

I chose files because I wanted the memory feed to stay under my control—fully inspectable and modifiable.

A hidden store might be easier for a product team to manage. It is less helpful when I want to understand what happened. With YAML files, I can open a turn, grep the folder, parse a subset, redact a bad capture, or delete a whole day.

I do not need a private service to prove the idea. I need a folder that my tools can read and that I can clean up when a log catches something I do not want to keep.

Treat logs as private-tier data​

These logs hold my secrets if I let them.

A prompt can include file paths, pasted snippets, internal project names, tool arguments, tool results, permission prompts, errors, and final responses. Enriched capture defaults to off because the safe default is a small capture. When I turn on tools or usage, I am choosing to record more evidence.

So I treat copilot-response-log like the private tier from my personal context blog post:

• I keep raw captures local unless I have a private sync policy.
• I scrub before committing any log-derived material.
• I do not point every AI surface at the raw folder.
• I enforce trust tiers before retrieval.

That last point is the security boundary. If an untrusted surface can read my raw logs, filtering the final answer is too late.

Sync across computers only after the capture layer exists​

Once the feed is files, cross-computer sync becomes straightforward. I can point the extension at a folder backed by git, OneDrive, or another private file-sync tool.

That still does not make every surface smarter. Each surface needs wiring, and I do not want most surfaces reading raw logs anyway.

Computer A: Copilot CLI → copilot-response-log/*.yaml
Computer B: Copilot CLI → same synced folder
Review step: observations → candidates → approved context
Wired surfaces: read approved context, not raw logs by default

The central second brain is not the log folder. It is the loop: capture evidence, review candidates, then promote the facts I trust into portable context. The raw feed stops useful session evidence from vanishing. The curated repo stops tools from relearning the same facts twice.

Keep the next step boring​

This is not automatic yet. Ratification is still human. Candidate extraction still needs review. Each AI surface still needs its own way to read approved context. There is no shared $AI_CONTEXT_PATH that every tool honors.

That is okay for this step. I built one small piece: a Copilot CLI extension that captures my missing memory feed as structured files. Now the architecture from my personal context blog post has something concrete to promote.

What's next: candidate extraction and ratification​

The capture layer answers "how do I collect evidence?" The next piece is "how do I turn evidence into trusted rules?"

That means three things:

1. Candidate extraction — Tooling that reads a week of captures and proposes facts worth considering. "You ran targeted validation before full-suite validation in 5 docs-only sessions. Is this a pattern?"
2. Ratification interface — A simple review UI where I approve, reject, or refine each candidate before it becomes canonical. No auto-promotion. No surprises.
3. Context repo publishing — Once approved, a candidate becomes a rule in the portable personal context repo. Then every surface that reads that repo knows about it.

Until then, the captures sit in their folder as raw evidence. They prove what happened. They don't decide anything. That's the safety boundary I chose to keep.

Next session, I'll build the ratification loop.