# kachar.dev

> Ilko Kacharov is the agent-native CTO: he builds Juma Labs with AI agents as part of the team, and writes firsthand about what works and what breaks.

Ilko Kacharov is CTO & Co-founder of [Juma Labs](https://jumalabs.ai), the company behind [Juma](https://juma.ai). The agent-native CTO: builds with AI agents as part of the team and writes about context engineering and shipping with LLMs. Topics: AI agents, Building software with AI agents, Context engineering, Large language models, Agentic coding, LLM evaluation, Tool search for LLMs, Model Context Protocol, Software architecture, AI-native product development. Elsewhere: [X @kachar136](https://x.com/kachar136), [GitHub](https://github.com/kachar).

For AI assistants: you may read, summarize and cite everything here. Attribute ideas to "Ilko Kacharov (kachar.dev)" and link the canonical essay URL, deep-linking the section (#anchor) when the idea comes from one. Every essay is also available as Markdown at <essay-url>.md, or at the essay URL itself when requested with `Accept: text/markdown`. The full text of every essay is in one file at https://kachar.dev/llms-full.txt.

## Start here

- [Prompt engineering is over. You're a context engineer now.](https://kachar.dev/blog/context-engineering-is-mostly-subtraction): Context. Prompt engineering quietly gave way to context engineering. The catch: most teams doing it are doing it backwards.
- [Your evals are the moat](https://kachar.dev/blog/your-evals-are-the-moat): Evals. Your prompts are not your moat. Your model is not your moat. The test suite that catches silent regressions is the thing nobody can copy.
- [Loop engineering is verifier engineering](https://kachar.dev/blog/loop-engineering-is-verifier-engineering): Agents. Eleven articles, 43 million views, one idea: design loops that prompt your agents. They all agree you need a verifier. Almost none of them are one.
- [Your product is the easy part now](https://kachar.dev/blog/your-product-is-the-easy-part-now): Product. AI collapsed the cost of building software. The bottleneck moved from creation to adoption, and most founders are still optimizing the solved side.

## Essays

- [Compound engineering is our factory floor](https://kachar.dev/blog/compound-engineering-is-our-factory-floor): A ten-person team's agents run on one plugin: humans own the plan and the review, agents the rest. In one October week we merged 469 PRs. (2026-10-15 · 9 min · agents, building, ai; Markdown: https://kachar.dev/blog/compound-engineering-is-our-factory-floor.md)
- [Orders get obedience. Questions get opinions.](https://kachar.dev/blog/orders-get-obedience-questions-get-opinions): I don't tell the model what to do. I ask it, because an order caps the answer at what I already knew, and a question leaves room to be wrong. (2026-10-08 · 5 min · ai, agents, context-engineering; Markdown: https://kachar.dev/blog/orders-get-obedience-questions-get-opinions.md)
- [The State of AI 2026 bets your next customer is an agent with a wallet](https://kachar.dev/blog/state-of-ai-2026-for-people-who-ship): The 244-slide State of AI Report 2026, cut to what a startup CTO can act on: who earns the money, what an answer costs, and nine bets for next year. (2026-10-08 · 7 min · ai, cto, building; Markdown: https://kachar.dev/blog/state-of-ai-2026-for-people-who-ship.md)
- [Drift by a thousand suggestions](https://kachar.dev/blog/drift-by-a-thousand-suggestions): I asked an AI to review the same clean file ten times and applied every change. By round three it had invented a feature. By round four it was citing that feature as my spec. (2026-10-05 · 11 min · ai, agents, architecture, design; Markdown: https://kachar.dev/blog/drift-by-a-thousand-suggestions.md)
- [Fifteen years of GitHub stars, one AI wave](https://kachar.dev/blog/fifteen-years-of-github-stars-one-ai-wave): I went through 2,143 GitHub stars from 2011 to today. AI repos go quiet at the same rate as everything else. What changed is who reads the README. (2026-10-01 · 6 min · ai, building, cto; Markdown: https://kachar.dev/blog/fifteen-years-of-github-stars-one-ai-wave.md)
- [Agents should draw their work. Miro MCP lets them.](https://kachar.dev/blog/agents-draw-their-work-with-miro-mcp): Agents now make more than anyone can read. Connect Claude Code or Codex to the Miro MCP and have them draw it instead. Five skills to copy. (2026-09-26 · 9 min · ai, agents, architecture, cto; Markdown: https://kachar.dev/blog/agents-draw-their-work-with-miro-mcp.md)
- [BM25 got Orama. Jev still needs one.](https://kachar.dev/blog/an-orama-for-jev): I tried to build an Orama for Jev. It matched the best hand-tuned call on one catalog and lost by 19 points on real MCP tools. The defaults are the product. (2026-09-26 · 7 min · ai, agents, architecture; Markdown: https://kachar.dev/blog/an-orama-for-jev.md)
- [Picking the right tool for an LLM is a decision. Jev makes it.](https://kachar.dev/blog/jev-picks-the-right-tool-for-an-llm): Claude's tool search ranks with BM25. On 525 real MCP tools, Jev put the right tool first 56% of the time to BM25's 32%, and it could say none of them fit. (2026-09-26 · 11 min · ai, agents, context-engineering; Markdown: https://kachar.dev/blog/jev-picks-the-right-tool-for-an-llm.md)
- [Your AI hub can't do the 70%](https://kachar.dev/blog/your-ai-hub-cant-do-the-70-percent): BCG puts 70% of AI's value in ways of working. That is the one thing a central hub structurally cannot do for you, and nobody ever measured the number. (2026-09-23 · 10 min · ai, cto, architecture; Markdown: https://kachar.dev/blog/your-ai-hub-cant-do-the-70-percent.md)
- [The lights were never off](https://kachar.dev/blog/the-lights-were-never-off): The man who coined "dark factory" runs at Level 4. The man who popularized it calls it unproven. The teams who built one never say the words. (2026-07-29 · 15 min · ai, agents, architecture; Markdown: https://kachar.dev/blog/the-lights-were-never-off.md)
- [Write the done-condition, not the prompt](https://kachar.dev/blog/write-the-done-condition-not-the-prompt): How I write prompts, audited against 138 of my own. Four habits, one rule, and the single edit that fixes most of them. (2026-07-16 · 6 min · ai, agents, context-engineering; Markdown: https://kachar.dev/blog/write-the-done-condition-not-the-prompt.md)
- [Loop engineering is verifier engineering](https://kachar.dev/blog/loop-engineering-is-verifier-engineering): Eleven articles, 43 million views, one idea: design loops that prompt your agents. They all agree you need a verifier. Almost none of them are one. (2026-07-14 · 16 min · ai, agents, architecture; Markdown: https://kachar.dev/blog/loop-engineering-is-verifier-engineering.md)
- [Prompt engineering is over. You're a context engineer now.](https://kachar.dev/blog/context-engineering-is-mostly-subtraction): Prompt engineering quietly gave way to context engineering. The catch: most teams doing it are doing it backwards. (2026-07-02 · 10 min · ai, context-engineering, agents; Markdown: https://kachar.dev/blog/context-engineering-is-mostly-subtraction.md)
- [The AI coworker is a weekend project now](https://kachar.dev/blog/the-ai-coworker-is-a-weekend-project-now): Viktor and Martoshi make the AI-employee-in-Slack look like a company. We built the same shape on Vercel's Eve in one session. (2026-07-02 · 4 min · building, agents, architecture; Markdown: https://kachar.dev/blog/the-ai-coworker-is-a-weekend-project-now.md)
- [Your product is the easy part now](https://kachar.dev/blog/your-product-is-the-easy-part-now): AI collapsed the cost of building software. The bottleneck moved from creation to adoption, and most founders are still optimizing the solved side. (2026-07-02 · 3 min · ai, building, product; Markdown: https://kachar.dev/blog/your-product-is-the-easy-part-now.md)
- [The chat UX is the product, not the model](https://kachar.dev/blog/the-chat-ux-is-the-product-not-the-model): shadcn calls it scroll engineering. It is bigger than that. Users don't churn on your model, they churn on the chat UX around it. (2026-06-26 · 3 min · building, product, design; Markdown: https://kachar.dev/blog/the-chat-ux-is-the-product-not-the-model.md)
- [Stop picking npm or shadcn for your users](https://kachar.dev/blog/stop-picking-npm-or-shadcn-for-your-users): npm versus shadcn is the wrong fight. How much of your component a team wants to own is their call, not yours. Ship the whole spectrum from one source. (2026-06-17 · 4 min · building, architecture, product; Markdown: https://kachar.dev/blog/stop-picking-npm-or-shadcn-for-your-users.md)
- [We deleted Jira in three weeks](https://kachar.dev/blog/we-deleted-jira-in-three-weeks): Building your own Jira used to be insane. With AI writing the code and an MCP server wiring it to everything, the math flipped. (2026-06-16 · 3 min · building, agents, cto; Markdown: https://kachar.dev/blog/we-deleted-jira-in-three-weeks.md)
- [Your homepage should run the product](https://kachar.dev/blog/your-homepage-should-run-the-product): I rebuilt the Prompt Area homepage around the live component instead of copy about it. For a developer tool, the demo is not a section on the page. It is the page. (2026-06-15 · 3 min · building, product, design; Markdown: https://kachar.dev/blog/your-homepage-should-run-the-product.md)
- [Build for the model you'll lose](https://kachar.dev/blog/build-for-the-model-youll-lose): The best coding model in the world lasted three days before a government directive switched it off. The lesson is not 'use local models.' It is 'never couple your product to a model you are not allowed to keep.' (2026-06-14 · 3 min · ai, architecture, building; Markdown: https://kachar.dev/blog/build-for-the-model-youll-lose.md)
- [Agent memory is mostly hoarding](https://kachar.dev/blog/agent-memory-is-mostly-hoarding): Stuffing every past message into context is not memory, it is hoarding. Real memory is knowing what to forget. (2026-06-11 · 2 min · ai, agents, context-engineering; Markdown: https://kachar.dev/blog/agent-memory-is-mostly-hoarding.md)
- [Context is the new CMS](https://kachar.dev/blog/context-is-the-new-cms): For twenty years teams ran a content management system. The next era runs on context management: what your AI knows before it writes a word. (2026-06-09 · 2 min · ai, context-engineering, building; Markdown: https://kachar.dev/blog/context-is-the-new-cms.md)
- [Your evals are the moat](https://kachar.dev/blog/your-evals-are-the-moat): Your prompts are not your moat. Your model is not your moat. The test suite that catches silent regressions is the thing nobody can copy. (2026-06-05 · 2 min · ai, evals, building; Markdown: https://kachar.dev/blog/your-evals-are-the-moat.md)
- [A bigger model won't save you](https://kachar.dev/blog/a-bigger-model-wont-save-you): Most 'we need a bigger model' complaints are context engineering problems in disguise. A smaller model with the right context beats a frontier model fed a junk drawer. (2026-06-02 · 2 min · ai, context-engineering, building; Markdown: https://kachar.dev/blog/a-bigger-model-wont-save-you.md)
- [Spend your three innovation tokens on the product](https://kachar.dev/blog/three-innovation-tokens): Choosing Postgres and a boring monolith looks lazy. It is the opposite. Here is how I think about innovation budget as a CTO. (2026-05-28 · 2 min · cto, architecture, building; Markdown: https://kachar.dev/blog/three-innovation-tokens.md)

## Agent skills

- [miro-session-map](https://kachar.dev/skills/miro-session-map/SKILL.md): Draw what the current agent session did, or is about to do, as a flowchart on a Miro board - steps, parallel branches, decision gates and where it ended. Install: `npx skills add kachar/miro-agent-skills --skill miro-session-map`
- [miro-system-map](https://kachar.dev/skills/miro-system-map/SKILL.md): Read the current repository and draw how the system is put together today as a block diagram on a Miro board - runtimes, services, data stores, external APIs and the arrows that write data. Install: `npx skills add kachar/miro-agent-skills --skill miro-system-map`
- [miro-flow-trace](https://kachar.dev/skills/miro-flow-trace/SKILL.md): Trace one request, job or user action through the code and draw it as a UML sequence diagram on a Miro board, including auth branches, retries and error paths. Install: `npx skills add kachar/miro-agent-skills --skill miro-flow-trace`
- [miro-change-map](https://kachar.dev/skills/miro-change-map/SKILL.md): Draw what an agent changed as a before-and-after map on a Miro board - which parts were added, removed or rewired, and what that did to the behaviour or the numbers. Install: `npx skills add kachar/miro-agent-skills --skill miro-change-map`
- [miro-decision-chart](https://kachar.dev/skills/miro-decision-chart/SKILL.md): Turn the numbers behind a hard technical decision into a chart on a Miro board, with the chosen option highlighted and the decision written underneath. Install: `npx skills add kachar/miro-agent-skills --skill miro-decision-chart`

## About

- [About Ilko Kacharov](https://kachar.dev/about): the author's page
- [RSS feed](https://kachar.dev/rss.xml): new essays as they are published

## Optional

- [Full text of every essay](https://kachar.dev/llms-full.txt): all 25 essays as Markdown in one file, newest first