{"slug": "munder-difflin-says-one-agent-redrew-119-blog-heroes-for-0", "title": "Munder Difflin says one agent redrew 119 blog heroes for $0", "summary": "Munder Difflin founder Chaitanya Giri said on August 19th that one agent rebuilt his project's blog, created hero illustrations for 119 posts, and deployed the redesign through CI, all for $0. The agent worked inside Claude Code, using reusable SVG components, 16 composition archetypes, and a media manifest, while Giri supplied feedback and approved the result. The redesign was prompted by Reddit criticism that the old blog looked unreadable and AI-written, and the agent was instructed to avoid red, leading to a magazine-style theme with serif headlines and yellow highlights.", "body_md": "# Munder Difflin says one agent redrew 119 blog heroes for $0\n\n**Founder Chaitanya Giri turned Reddit criticism into a Claude Code workflow built from SVG parts, 16 composition archetypes and one media manifest.**\n\nBy [RuntimeWire Staff](/author/runtimewire-staff)\n· Published\n\nPrimary source: [Munder Difflin Newsroom](https://munderdiffl.in/blog/an-agent-redesigned-this-blog/)\n\n## Why it matters\n\nGiri's redesign shows how a solo founder can turn an agent into production capacity by constraining it with reusable components, human review and a deployment pipeline. The durable output is the system behind the 119 images, not the images alone.\n\nOn August 19th, [Munder Difflin](https://munderdiffl.in/?ref=runtimewire) founder [Chaitanya Giri](https://in.linkedin.com/in/chaitanya-giri?ref=runtimewire) said [in a post](https://munderdiffl.in/blog/an-agent-redesigned-this-blog/?ref=runtimewire) that one agent rebuilt his project's blog, created hero illustrations for 119 posts and deployed the finished redesign through CI. The agent worked inside Claude Code, while Giri supplied feedback and approved the result.\n\nThe project began with an unusually useful piece of launch feedback: readers on Reddit said the old blog looked unreadable and AI-written. Giri responded by giving the agent a brief instead of a design file. It had to make the site more readable, give every post an image and avoid red.\n\nThat last instruction became a recurring human checkpoint. \"The accent is yellow, not red. No red anywhere,\" Giri told the agent, according to his account. The agent produced two functioning themes behind a switchable CSS flag, one bookish and one modeled on an editorial magazine. Giri selected the magazine version, then directed changes to its typography and palette.\n\nThe live result has serif headlines, a narrower reading column, a cream background and yellow highlights. The larger lesson sits underneath the styling: Giri used the agent to create a reusable graphics system instead of asking an image model for 119 unrelated pictures.\n\n### Reddit criticism became a product test\n\nGiri maintains Munder Difflin as an individual developer in India, according to the project's [privacy policy](https://munderdiffl.in/privacy.html?ref=runtimewire). His [public profile](https://in.linkedin.com/in/chaitanya-giri?ref=runtimewire) lists studies at Netaji Subhas Institute of Technology from 2021 to 2025 and an Applied Scientist internship at Amazon. It also lists wins in Amazon's 2023 machine-learning challenge and Morgan Stanley's Code to Give competition.\n\nHe started Munder Difflin after running several coding-agent terminals at once turned him into the person copying context between sessions and resolving conflicting edits. In the project's [origin story](https://munderdiffl.in/blog/why-we-built-munder-difflin/?ref=runtimewire), Giri described himself as the message bus, conflict resolver and long-term memory for his agents.\n\n[Munder Difflin's open-source application](https://github.com/chaitanyagiri/munder-difflin?ref=runtimewire) is his attempt to automate that coordination layer. It wraps command-line agents including Claude Code, Codex, Gemini CLI, Grok, Kimi Code and GitHub Copilot CLI, then gives them shared memory, mailboxes, isolated worktrees and an office-style interface. The local version is MIT licensed and runs the underlying tools on the user's computer.\n\nThe blog redesign gave Giri a contained way to test the same thesis outside routine coding work. The assignment required the agent to interpret criticism, propose alternatives, apply feedback across templates, generate assets and move the result into production. It also exposed where human direction remained essential. The agent initially produced a black-on-black category badge, chose a coral topic color after being told to avoid red and needed a no-cache development server after Giri's browser repeatedly showed the old build.\n\nThose corrections make the account more credible than a frictionless autonomy demo. The agent handled implementation volume. Giri made the taste decisions and caught visual failures.\n\n### The illustrations became software\n\nGiri says image-generation services would have cost between $5 and $20 for the initial batch, with additional charges for regeneration. He also expected generated bitmaps to reproduce the same generic appearance that prompted the redesign.\n\nThe agent instead built the illustrations in plain JavaScript and SVG. Its system has three layers: a reusable library of characters and office objects, 16 composition functions and a specification file assigning a scene and annotations to each post.\n\nThe character is adapted from the MIT-licensed Xiaohei illustration project and recolored yellow. Other reusable parts include desks, terminals, clocks, envelopes, coffee cups, shields and a kanban board. An SVG turbulence filter adds a seeded wobble, allowing each image to look hand-drawn without making the underlying geometry unpredictable.\n\nEach composition function arranges those parts into a recurring scene such as \"agents talking,\" \"night shift,\" \"versus\" or \"the routing desk.\" A headless browser opens the selected composition at 1600x900 and saves a screenshot to the relevant post directory. Giri says the agent rendered all 119 images in three batches without calling an image-generation API.\n\nThat structure matters more than the $0 headline. When a sleeping character appeared with open eyes, the agent changed the drawing function and regenerated the asset. A diffusion workflow would have required another prompt and produced a new image with less predictable changes. Code gave Giri a stable design vocabulary and a cheap revision path.\n\nA central `media.json`\n\nfile controls paths, alt text and readiness states for images, inline figures and video embeds. Templates show a designed placeholder until an asset is marked ready, avoiding broken-image states during production. Adding the illustration pipeline to a new article requires another manifest entry rather than a new round of manual template work.\n\n### A content team inside a one-person project\n\nThe redesign extends an earlier automation system. Giri has said in a separate [technical walkthrough](https://munderdiffl.in/blog/deploy-a-blog-writer-agent/?ref=runtimewire) that a writer agent drafts posts in an isolated worktree, while the Munder Difflin orchestrator acts as the single committer and a human approves deployment. The project says that pipeline produced the blog's hundred-plus posts.\n\nThe new illustration layer turns that operation into something closer to a small publishing system. A brief can produce a draft, metadata, internal links and a hero assembled from approved visual primitives. Giri still controls the topics, style references, final design choices and deployment gate.\n\nMunder Difflin remains an early prototype, a label used in its own [security policy](https://github.com/chaitanyagiri/munder-difflin/security?ref=runtimewire). Its autonomy claims should be read alongside the ordinary reliability work required to keep agents running. On August 22nd, [RuntimeWire reported that Munder Difflin's v0.4.5 release fixed broken cost accounting, Apple Silicon memory failures and unreliable worker communication](/article/munder-difflin-v045-ai-clones-pro-launch).\n\nThe redesign does not establish that an agent can independently discover good taste or run a publication without supervision. Giri's account shows a narrower and more practical capability: a founder can encode taste through examples, constraints and reusable components, then delegate the repetitive implementation.\n\nFor Giri, the timing also served the product. Criticism of an AI-written appearance arrived during launch week, when Munder Difflin needed to demonstrate that its agents could perform coherent work across a real repository.\n\nThe strongest part of the project is the decision to make the images maintainable. Giri did not leave himself with 119 opaque outputs. He ended up with a parts library, scene functions and one manifest that can keep producing assets after the original agent session is gone. That is the founder's real bet with Munder Difflin: useful agents should leave systems behind, not just completed tasks.", "url": "https://wpnews.pro/news/munder-difflin-says-one-agent-redrew-119-blog-heroes-for-0", "canonical_source": "https://runtimewire.com/article/munder-difflin-agent-redesigns-blog-svg-illustrations", "published_at": "2026-08-22 18:43:45+00:00", "updated_at": "2026-08-22 19:13:50.205277+00:00", "lang": "en", "topics": ["ai-agents", "generative-ai", "developer-tools"], "entities": ["Munder Difflin", "Chaitanya Giri", "Claude Code", "Reddit", "Amazon", "Morgan Stanley", "Netaji Subhas Institute of Technology", "GitHub Copilot CLI"], "alternates": {"html": "https://wpnews.pro/news/munder-difflin-says-one-agent-redrew-119-blog-heroes-for-0", "markdown": "https://wpnews.pro/news/munder-difflin-says-one-agent-redrew-119-blog-heroes-for-0.md", "text": "https://wpnews.pro/news/munder-difflin-says-one-agent-redrew-119-blog-heroes-for-0.txt", "jsonld": "https://wpnews.pro/news/munder-difflin-says-one-agent-redrew-119-blog-heroes-for-0.jsonld"}}