AI Workflow and Tips Daily

Culmen Pods

A short daily show for people who use AI at work. Each episode distills what practitioners are actually sharing — a new repo, a new skill, a workflow trick or tip — into one repeatable recipe you can run today, plus a quick tour of fresh finds worth bookmarking.

Episodes

  1. Aug 10

    Episode 8: The Name You'd Type Was Last Touched in 2022

    A six-day-old document converter with a package name that lies to `npm view`, a big company publishing its actual AI coding bill, and a government initiative whose most interesting feature is everything it declines to state. Then the recipe: serving the new stateless MCP spec and proving it works with a single call — including two things this show had written down wrong and corrected by running them. ## In this episode **Fresh Finds** - **anydoc** (Firecrawl) — a fast Rust document-to-markdown library. MIT, read off the LICENSE file, whose copyright line names Sideguide Technologies Inc. The library is free and runs entirely on your machine; the hosted parse tier is a separate, credit-metered product. - **"Managing AI Coding Costs at Scale"** (Databricks) — five bylined authors, five levers, and two percentages that are Databricks measuring its own internal usage with no denominator disclosed. The framing matters more than the numbers. - **The Genesis Open Models Initiative** (US Department of Energy, Arcee AI, Argonne) — a find because of what is *missing*. No licence named on either primary page, no model in existence yet, no price stated in either direction, and two official pages giving different deadlines. **The Recipe — serve the stateless MCP spec, and prove it with one call** Install the current SDK, expect an `npm audit` failure that is not your fault, stand up a roughly fifteen-line server, add the Origin and Host validation the SDK deliberately leaves to you, then verify with the Inspector. Includes two corrections this show made by re-running its own notes. **Two quick tips** - Treat everything your agent *reads* as untrusted input — including a pull request — with the UK AI Safety Institute's incident report as the evidence. - Run `npm view time.modified` before you install something an agent recommended. The gap between the conversation and the timestamp is the signal. **Do This Today** One command, one file out, under ten minutes. ## Sources - [anydoc on GitHub](https://github.com/firecrawl/anydoc) - [anydoc LICENSE](https://raw.githubusercontent.com/firecrawl/anydoc/main/LICENSE) - [anydoc repository metadata](https://api.github.com/repos/firecrawl/anydoc) - [@firecrawl/anydoc on npm](https://www.npmjs.com/package/@firecrawl/anydoc) - [firecrawl-anydoc on PyPI](https://pypi.org/pypi/firecrawl-anydoc/json) - [Firecrawl pricing](https://www.firecrawl.dev/pricing) - [Managing AI Coding Costs at Scale — Databricks](https://www.databricks.com/blog/managing-ai-coding-costs-scale) - [Omnigent on GitHub](https://github.com/omnigent-ai/omnigent) - [DOE launches the Genesis Open Models Initiative](https://www.energy.gov/undersecretaryforscience/articles/us-department-energy-launches-genesis-open-models-initiative) - [Arcee AI — Science](https://www.arcee.ai/science-1) - [Genesis Open Models portal](https://genesisopenmodels.anl.gov/) - [UK AISI — incident report on unsanctioned agent behaviour during cyber testing](https://www.aisi.gov.uk/blog/incident-report-unsanctioned-agent-behaviour-during-cyber-testing) - [Socket — AI agents and open-source malware](https://socket.dev/blog/ai-agent-open-source-malware) - [MCP streamable HTTP transport spec (2026-07-28)](https://modelcontextprotocol.io/specification/2026-07-28/basic/transports/streamable-http) - [MCP — build a server](https://modelcontextprotocol.io/docs/2026-07-28/develop/build-server) - [MCP SDK betas](https://blog.modelcontextprotocol.io/posts/sdk-betas-2026-07-28/) - [MCP 2026-07-28 release post](https://blog.modelcontextprotocol.io/posts/2026-07-28/) - [GHSA-frvp-7c67-39w9](https://github.com/advisories/GHSA-frvp-7c67-39w9) - [Sean Goedecke — LLMs reward expertise](https://www.seangoedecke.com/llms-reward-expertise/) ## Also from this network **The AI Daily Wire** — the day's AI news that actually matters, ranked, sourced, and one deep dive. Different host, same standards. ## Built by - **Culmen Games** — an indie game studio, an Android roster built solo. `studiosculmen@gmail.com` - **Handled** — done-for-you content and AI agent workflows for creators and time-poor professionals. `handled.anmol@gmail.com`

  2. Aug 8

    Episode 6: Half the Tests Failed and the Build Went Green

    `PROMPTFOO_PASS_RATE_THRESHOLD` takes a **percentage, not a fraction** — so writing `0.9` when you meant 90% gives you a **0.9%** threshold, and CI goes green with half your tests failing. Nothing warns you. This episode builds the eval gate that makes swapping a model a check instead of a vibe: one YAML file, one `echo` provider, and an exit code you can read. Plus three finds, two tips, and a ten-minute thing to do today. No git this episode — episodes 3, 4 and 5 lived in worktrees and hooks, and this one deliberately leaves them. ## In this episode - **Kimi K3 (Moonshot AI)** — 2.8T-parameter MoE, 104B activated per token, open weights, 1M context, natively multimodal (text, images and video). Weights went public **Monday 27 July 2026**. **The licence is the story.** Plenty of coverage carried "Modified MIT" forward from K2. It isn't. K2's file is titled *Modified MIT License*; K3's is titled *Kimi K3 License*, and **MIT is never named in it**. **Section 2** (MaaS revenue over **$20M across any consecutive 12 months** → separate agreement with Moonshot; and MaaS is defined narrowly, excluding models embedded in a specific product feature and excluding mere relaying to models hosted by others) and **Section 3** (**$20M per month** in revenue or 100M MAU → display "Kimi K3" in your UI) are **different units and must not be conflated**. **Section 4** exempts internal use and use via Moonshot's own products or certified inference partners — but "internal use" is defined as **not making the software, its outputs, or its underlying capabilities available to third parties**, which is narrower than "inside your own company". Weights are free; Moonshot **names no price** for the separate agreement, which is not the same as free. - **DeepSeek-V4-Flash-0731** — the official release, superseding the preview. `0731` is a **build date, not a version**. **304B parameters, read from the repo's own safetensors index**, not from press coverage. Active parameter count and context length are **not stated** on the model card page fetched, so the episode doesn't state them. **Licence: MIT** — verified from the repo's licence tag and the README. Weights free; **DeepSeek API pricing was not checked this run**. - **DwarfStar / `antirez/ds4` (Salvatore Sanfilippo, creator of Redis)** — **name mismatch: the repo and binary are `ds4`, the README calls it "DwarfStar"**. GitHub *repository* search for "DwarfStar" won't surface it (the name is in neither the repo name nor the description) — you'll get other people's wrappers instead. Search the path, `antirez/ds4`. A small native local inference engine tuned first for DeepSeek V4 Flash — and specifically the **0731 build**: the README names it repeatedly and ships `DeepSeek-V4-Flash-DSpark-support-0731.gguf`. Also GLM 5.2 and, on very-high-memory machines, DeepSeek V4 PRO. Metal / CUDA / ROCm. **MIT**, free, no paid tier named anywhere in the README or LICENSE. README credits **Georgi Gerganov** for the llama.cpp / GGML work. **Caveats in the maintainer's own words: "beta quality", DSpark "experimental" and opt-in, and — on 128 GB MacBooks — PRO q2 streaming "experimental but usable for inspection".** Metal's primary target is **Macs with 96 GB+**; smaller machines get `--ssd-streaming`. **Created back in May — ~20k stars and climbing, so trending, not new.** Its README also credits GPT 5.5, 5.6 and Claude Fable with "strong assistance", humans leading ideas, testing and debugging. - **The Recipe — put a pass/fail gate on your prompts before you change models.** promptfoo 0.122.0, **MIT — verified from the installed `package.json` and npm's registry entry.** (Neither OpenAI's post nor promptfoo's names a licence at all; OpenAI's says they'll "continue building the open-source project", which is a claim about openness, not a second confirmation of MIT.) **OpenAI announced its intent to acquire Promptfoo on Monday 9 March 2026 — announced, not closed: the post is titled "OpenAI *to acquire* Promptfoo" and says closing is subject to customary conditions.** Founders **Ian Webster** and **Michael D'Angelo** — worth knowing that an OpenAI-bound tool grades every vendor's models. **Price: Community is "Free Forever" including red teaming capped at 10,000 probes/month; Enterprise and On-Premise both say "Custom", which means promptfoo names no price — not the same as free.** Node floor is **`>=22.22.0`**, read off the installed package; the GitHub Action states the same floor and recommends Node 24 LTS. - **The steps** — `npx promptfoo@latest init --no-interactive` (**plain `init` is a wizard** — with nothing attached to answer it, it exits **130** and writes no files) → get a **zero-cost green run first** using the `echo` provider (seven deterministic assertion types passed, $0 spent) → then swap in the provider you actually ship → gate CI on the exit code. - **The exit-code contract, measured rather than quoted** — all pass: **0**. At least one assertion fails: **100**. Typo'd provider id, missing config, missing API key: **1**. So "promptfoo returns non-zero" is true and useless — a shell check for `$? -eq 1` catches the broken setup and **silently ignores every failing test**. (The 100 is the default; `PROMPTFOO_FAILED_TEST_EXIT_CODE` can override it, so check inherited CI before assuming.) - **Tip 1 — the threshold unit trap.** `PROMPTFOO_PASS_RATE_THRESHOLD` is a **percentage**, and it's inclusive (`>=`). Against a rigged 50% pass rate: `60` → exit 100, `50` → exit 0, **`0.9` → exit 0**, because it means 0.9%. **Write `90`, not `0.9`.** Precisely what's silent: the terminal *does* still print `✗ 1 failed (50.00%)` — what you don't get is any threshold message (that line only prints when the gate trips) and, critically, the **exit code is 0**. A human sees the failure; the machine gating your deploy sees success. The counter-intuitive bonus: **misspelling the variable name is safe** — `PROMPTFOO_PASSRATE_THRESHOLD` is ignored, the default applies, and the gate still fires. The name is forgiving; the value is not. - **Tip 2 — a typo'd assertion `type:` counts as an "error", not a "failure".** It still exits 100, so CI catches it, but the summary reads `0 passed, 0 failed, 1 error` — so a dashboard counting *failures* shows a clean green bar while nothing ran. Count errors too, or trust only the exit code. - **The honest counter-argument** — Hamel Husain and Shreya Shankar's evals FAQ, aired rather than softened. Note the attribution: the sharpest line ("all you get from prefab evals is you don't know what they actually do… at worst they create an illusion of confidence that is unjustified") is the FAQ **quoting one of their instructors, Eleanor Berger** — the authors' own version is "generic evaluations waste time and create false confidence". Theirs directly: "if you're passing 100% of your evals, you're likely not challenging your system enough", and you should be "extremely skeptical of this process being delegated". The harness is the easy part; the assertions are not. - **Stated on air** — none of this was run against a live frontier model this week. Everything was verified on Windows / Node 24.18.0 with the `echo` provider, so latency, cost accounting and `llm-rubric` model-graded behaviour are **not verified**. Also disclosed: **Claude Code requires a paid Claude plan (Pro, Max, Team, Enterprise, or API billing) — nothing on the free tier.** - **Do This Today** — one folder, one `promptfooconfig.yaml`, two test cases (one passing, one failing), then check the exit code. **You're looking for 100, not 1.** ## Sources **Fresh Finds** - [moonshotai/Kimi-K3 — Hugging Face](https://huggingface.co/moonshotai/Kimi-K3) - [Kimi K3 LICENSE](https://huggingface.co/moonshotai/Kimi-K3/blob/main/LICENSE) - [Kimi-K2-Instruct LICENSE (for the "Modified MIT" contrast)](https://huggingface.co/moonshotai/Kimi-K2-Instruct/blob/main/LICENSE) - [Simon Willison — Kimi K3](https://simonwillison.net/2026/Jul/27/kimi-k3/) - [deepseek-ai/DeepSeek-V4-Flash-0731 — Hugging Face](https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731) - [DeepSeek-V4-Flash-0731 — HF model API (parameter count)](https://huggingface.co/api/models/deepseek-ai/DeepSeek-V4-Flash-0731) - [antirez/ds4 — "DwarfStar"](https://github.com/antirez/ds4) - [ds4 README](https://github.com/antirez/ds4/blob/main/README.md) - [ds4 LICENSE](https://github.com/antirez/ds4/blob/main/LICENSE) **The Recipe** - [promptfoo — getting started](https://www.promptfoo.dev/docs/getting-started/) - [promptfoo — command line reference (exit codes)](https://www.promptfoo.dev/docs/usage/command-line/) - [promptfoo — the `echo` provider](https://www.promptfoo.dev/docs/providers/echo/) - [promptfoo — GitHub Action](https://www.promptfoo.dev/docs/integrations/github-action/) - [promptfoo — pricing](https://www.promptfoo.dev/pricing/) - [OpenAI to acquire Promptfoo](https://openai.com/index/openai-to-acquire-promptfoo/) - [Promptfoo — joining OpenAI](https://www.promptfoo.dev/blog/promptfoo-joining-openai/) - [Hamel Husain and Shreya Shankar — Evals FAQ](https://hamel.dev/blog/posts/evals-faq/) --- Sister show: **The AI Daily Wire** — the daily AI news brief from the same network. Different host, same standards. ## Built by - **Culmen Games** — a solo-built indie game studio. studiosculmen@gmail.com - **Handled** — done-for-you content and AI agent workflows for creators and time-poor professionals. handled.anmol@gmail.com

  3. Aug 7

    Episode 5: The Hook Git Fires That Nobody Documents

    `git worktree add` fires the `post-checkout` hook — and the `git-worktree` man page never mentions hooks at all. That one undocumented line is how every parallel agent gets its own ports, its own database and its own `.env` automatically, instead of by hand forever. Episode 3 gave every agent its own worktree; this is the layer underneath it. Plus three finds, two tips, and one ten-minute thing to do today. ## In this episode - **Cloudflare OS (5 Aug 2026)** — an agent workspace on Cloudflare Workers: isolated runtime, governance layer for internal data and services. **Apache-2.0** (licence file read directly), free to self-host on your own Cloudflare account. The planned managed dashboard product **names no price**. The README calls the project **"in a state of heavy development"** and an **"early access" release**. "Thousands of people across every function" using it internally daily is **Cloudflare's own count**, unaudited. - **Meta Muse Code (beta) + Muse Spark 1.2 (5 Aug 2026)** — Meta's terminal coding agent. **No licence stated anywhere**, so not "open source". **No price stated for the CLI**, which is not the same as free. **macOS and Linux only — no Windows.** Install host is `dev.meta.ai`, *not* `developers.meta.ai` — and the flag is `-fsSL`, with a **capital S and a capital L** (curl's flags are case-sensitive, and a lowercase `-l` is a different flag that fails quietly). Model API, per Artificial Analysis attributing Meta's API: **$1.25 / 1M input tokens, $4.25 / 1M output tokens** (standard tier). Meta's post shows benchmark **charts with no numbers printed in the text**. - **"How Castform + Neon Beats Frontier Models on Price and Efficiency"** — a 4B open-source model post-trained to match GPT-5.6 Sol on search retrieval at 100x less cost. The unit that matters: **~$0.03 per end-to-end multi-turn search request**. The honesty beat is what's missing — no base model named, no licence, no numeric accuracy scores, no benchmark named — and Hacker News said so. - **The Recipe** — give every parallel agent its own ports, database and env file automatically: slugify the branch, install one `post-checkout` hook, derive identity and ports, generate the env file, isolate Compose projects, isolate database names, and set per-worktree git config. - **Tip 1** — stop allocating ports at all. Publish a container port in Compose with **no host port**, let the kernel pick, and read it back with `docker compose port SERVICE PRIVATE_PORT`. Zero collisions by construction. - **Tip 2, and a correction we ran ourselves** — the docs say `git config --worktree` behaves as `--local` when `extensions.worktreeConfig` is unset, which sounds like a silent data leak across worktrees. It isn't. With a second worktree present git **hard-errors** (`fatal: --worktree cannot be used with multiple working trees…`, exit 128, nothing written). The silent fallback only happens in a repo with exactly one worktree — where there's nobody to leak into. Correct quote, wrong inference; running the command settled it in ten seconds. - **Spelling, said out loud** — the file is `.worktreeinclude`: one word, no hyphen, leading dot. The standalone binary `git-worktreeinclude` (MIT, by Satoshi Ebisawa) **does** have a hyphen, and you invoke it as `git worktreeinclude apply`. Note Claude Code requires a paid Claude plan — there is no free tier. - **The published counter-argument** — David Mytton (Arcjet) built this shape and rejected parts of it: dynamic ports don't propagate through the devcontainer spec, and per-workspace databases cost too much on a laptop. They run one shared infrastructure stack instead, and accept there's no per-workspace data isolation. - **Do This Today** — write a two-line `post-checkout` hook that fires on the null ref (forty zeros), and prove to yourself that `git worktree add` runs it. ## Sources **Fresh Finds** - [Cloudflare OS — announcement](https://blog.cloudflare.com/cloudflare-os/) - [cloudflare/cloudflare-os](https://github.com/cloudflare/cloudflare-os) - [Introducing Muse Code and Muse Spark 1.2 — Meta AI Research](https://research.meta.ai/blog/introducing-muse-code-and-muse-spark-1-2) - [How Castform + Neon beats frontier models on price and efficiency](https://neon.com/blog/how-castform-neon-beats-frontier-models-on-price-and-efficiency) - [Hacker News discussion](https://news.ycombinator.com/item?id=49186762) **The Recipe** - [Ally Piechowski — A setup script for git worktrees](https://piechowski.io/post/setup-script-git-worktrees/) - [githooks — official documentation](https://git-scm.com/docs/githooks) - [git-worktree — official documentation](https://git-scm.com/docs/git-worktree) - [git-config — official documentation](https://git-scm.com/docs/git-config) - [Mark Skelton — Using Git hooks when creating worktrees](https://mskelton.dev/bytes/using-git-hooks-when-creating-worktrees) - [Claude Code — worktrees and `.worktreeinclude`](https://code.claude.com/docs/en/worktrees) - [Conductor — `.worktreeinclude` reference](https://www.conductor.build/docs/reference/worktreeinclude) - [satococoa/git-worktreeinclude — Satoshi Ebisawa](https://github.com/satococoa/git-worktreeinclude) - [Docker Compose — specify a project name](https://docs.docker.com/compose/how-tos/project-name/) - [docker compose port](https://docs.docker.com/reference/cli/docker/compose/port/) - [PostgreSQL — CREATE DATABASE](https://www.postgresql.org/docs/current/sql-createdatabase.html) - [JD Wuarin — The missing piece of the Claude Code workflow: isolation](https://www.barnacle.ai/blog/2026-02-07-the-missing-piece-of-the-claude-code-workflow-isol) - [David Mytton, Arcjet — From devcontainers to VMs: parallel dev environments for AI agents](https://blog.arcjet.com/from-devcontainers-to-vms-parallel-dev-environments-for-ai-agents/) --- Sister show: **The AI Daily Wire** — the daily AI news brief from the same network. Different host, same standards. ## Built by - **Culmen Games** — a solo-built indie game studio. studiosculmen@gmail.com - **Handled** — done-for-you content and AI agent workflows for creators and time-poor professionals. handled.anmol@gmail.com

  4. Aug 6

    Episode 4: Your Five Agent Branches Are a Fan, Not a Stack

    Five agents finish overnight and leave you five branches. GitHub's stacked pull requests, in public preview since 30 July 2026, look like the answer — until you read GitHub's own docs, which say stacks can't have branching structures, and five siblings off `main` are exactly that. So this episode is the honest version: sort your branches into genuinely-dependent and independent first, and only the dependent list becomes a stack. Plus three finds, two tips, and one ten-minute thing to do today. ## In this episode - **Open Code Review (Alibaba Group)** — deterministic pipelines plus an LLM agent leaving line-level PR comments. Apache-2.0, but bring-your-own-key, so the running cost is your model provider's per-token rate. Its Claude Code comparison is self-graded on a benchmark Alibaba built. - **Soup (`soup-cli`, Alpamys Makazhan)** — the whole fine-tune config is one YAML file, and "layer streaming" fine-tunes an 8B on a 4 GB laptop GPU. Apache-2.0; the site says free forever. Layer streaming is marked BETA by the maintainer. - **Shieldstral-1.0-3B (Mistral AI)** — a 3B guardrail model that takes your safety policy as plain-language text at inference time, classifying text and images on one 16 GB GPU. Apache 2.0, weights free to download. - **All three finds' benchmark numbers are maker-measured** — Alibaba graded Alibaba, the maintainer graded Soup, Mistral graded Mistral. No independent replication found for any of them. - **The Recipe: sort first, then stack** — why five parallel agent branches are a fan and not a stack, the ten steps, what success looks like, and the failure modes. - **The version-floor correction** — episode 3 said `gh` 2.0+, which came from the extension's README. The feature's own quickstart requires **`gh` 2.90.0 or later and Git 2.20 or later**. Both numbers are GitHub's; state the higher one. - **The price, stated carefully** — across six GitHub-owned pages, GitHub names no price, no plan, no tier. That is not the same as GitHub saying it's free. The `gh-stack` extension itself is MIT, and that part is unambiguous. - **Tip 1** — turn on `git rerere` globally before you cascade-rebase anything. `gh stack init` enables it, but only inside that repo. - **Tip 2** — exit code 9 says "stacked pull requests not enabled for this repository," but issue #394 reports that message appearing when the real cause was the wrong active `gh` account. Run `gh auth status` first. - **Do This Today** — check your floors, turn on rerere, install the extension, then write two lists of your agent branches: genuinely dependent, and independent. ## Sources **Fresh Finds** - [Open Code Review — alibaba/open-code-review](https://github.com/alibaba/open-code-review) - [Open Code Review — project site](https://open-codereview.ai) - [Soup — MakazhanAlpamys/Soup](https://github.com/MakazhanAlpamys/Soup) - [trysoup.dev](https://trysoup.dev) - [Shieldstral — Mistral AI announcement](https://mistral.ai/news/shieldstral/) - [Shieldstral-1.0-3B — model card](https://huggingface.co/mistralai/Shieldstral-1.0-3B) **The Recipe** - [Stacked pull requests are now in public preview](https://github.blog/changelog/2026-07-30-stacked-pull-requests-are-now-in-public-preview/) - [Stacked PRs quickstart](https://docs.github.com/en/pull-requests/get-started/stacked-prs-quickstart) - [About stacked pull requests](https://docs.github.com/en/pull-requests/get-started/about-stacked-prs) - [Stacked pull requests reference](https://docs.github.com/en/pull-requests/reference/stacked-pull-requests) - [Stacked PRs CLI commands](https://docs.github.com/en/pull-requests/reference/stacked-prs-cli-commands) - [Roll out stacked PRs across your organization](https://docs.github.com/en/pull-requests/tutorials/roll-out-stacked-prs) - [Managing stacked pull requests](https://docs.github.com/en/pull-requests/how-tos/create-pull-requests/managing-stacked-pull-requests) - [Troubleshooting stacked pull requests](https://docs.github.com/en/pull-requests/how-tos/merge-and-close-pull-requests/troubleshooting-stacked-pull-requests) - [Stack AI-generated code in pull requests (Copilot tutorial)](https://docs.github.com/en/copilot/tutorials/stack-ai-generated-code-in-pull-requests) - [github/gh-stack](https://github.com/github/gh-stack) - [github/gh-stack — releases](https://github.com/github/gh-stack/releases) - [github/gh-stack — open issues](https://github.com/github/gh-stack/issues) - [GitHub CLI v2.90.0 release notes](https://github.com/cli/cli/releases/tag/v2.90.0) - [Managing a merge queue](https://docs.github.com/en/repositories/configuring-branches-and-merges-in-your-repository/configuring-pull-request-merges/managing-a-merge-queue) - [Tom Smith, DevOps.com — GitHub brings stacked pull requests out of the shadows](https://devops.com/github-brings-stacked-pull-requests-out-of-the-shadows/) - [Flavio Copes — GitHub stacked PRs](https://flaviocopes.com/github-stacked-prs/) - [GitHub community discussion #201439](https://github.com/orgs/community/discussions/201439) - [Lobsters — Stacked pull requests are now public](https://lobste.rs/s/pzl6f0/stacked_pull_requests_are_now_public) - [Mike McQuaid — Sandboxed agent worktrees](https://mikemcquaid.com/sandboxed-agent-worktrees-my-coding-and-ai-setup-in-2026/) - [Graphite pricing](https://graphite.com/pricing) - [AgenticFlict — Ogenrwot & Businge, arXiv 2604.03551](https://arxiv.org/abs/2604.03551) **Two Quick Tips** - [git-rerere — official documentation](https://git-scm.com/docs/git-rerere) --- Sister show: **The AI Daily Wire** — the daily AI news brief from the same network. Different host, same standards. ## Built by - **Culmen Games** — solo indie Android game studio — studiosculmen@gmail.com - **Handled** — done-for-you content + AI agent workflows for creators and time-poor professionals — handled.anmol@gmail.com

  5. Aug 5

    Episode 3: Your Agent Is Fixing a Bug That Doesn't Exist

    Two coding agents pointed at the same checkout will overwrite each other, and worse, an agent can't tell a port collision from a real bug, so it starts "fixing" code that was never broken. This episode is the full recipe for one worktree per agent: what a worktree actually is, the three vendors that shipped first-class support for it independently (and a fourth that documents the manual pattern), the exact steps, and the failure modes that were reproduced on a real machine rather than recalled from a blog post. Plus three finds, two tips, and one ten-minute thing to do today. ## In this episode - **QM** — Y Combinator open-sourced the multi-agent harness it uses internally. MIT, free and open source, but you bring your own model keys and your own hosting. - **Stacked pull requests** — GitHub's answer to the "five parallel agents, five branches" problem. In public preview and subject to change, rolling out to all repositories. GitHub names no price and no plan restriction. Needs GitHub CLI 2.90.0+ and Git 2.20+. - **TurboFieldfare** — Gemma 4 26B-A4B on Apple Silicon in about 2 GB of RAM, by streaming the mixture-of-experts weights from SSD. Apache 2.0. Apple Silicon only. - **The Recipe: one worktree per agent** — steps in order, what success looks like, the missing `.env`, port and database collisions, the "already used by worktree" error, and the branch that `remove` leaves behind. - **The myth corrected** — parallel agents do not fight over one `.git` index lock. Each linked worktree gets its own index. - **The counter-argument** — Alex Chaplinsky on why a worktree is a collision boundary, not a security boundary. - **Tip 1** — the core of Ankur Sethi's standing instruction for your agent, quoted verbatim (and honestly flagged: it's an excerpt, and he cites no studies). - **Tip 2** — JFrog found 54 of 55 SQLite advisories were fabricated, some scored 9.8 CRITICAL. The defence is one grep. - **Do This Today** — make one worktree, look at what's missing, write a `.worktreeinclude`, then tear it down including the branch. ## Sources **Fresh Finds** - [QM — yc-software/qm](https://github.com/yc-software/qm) - [Y Combinator's announcement](https://x.com/ycombinator/status/2083243960684908768) - [Stacked pull requests are now in public preview](https://github.blog/changelog/2026-07-30-stacked-pull-requests-are-now-in-public-preview/) - [github/gh-stack](https://github.com/github/gh-stack) - [Stacked PRs quickstart](https://docs.github.com/en/pull-requests/get-started/stacked-prs-quickstart) - [TurboFieldfare — drumih/turbo-fieldfare](https://github.com/drumih/turbo-fieldfare) **The Recipe** - [git-worktree — official documentation](https://git-scm.com/docs/git-worktree) - [OpenAI Codex — Git worktrees](https://learn.chatgpt.com/docs/environments/git-worktrees) - [Cursor — worktrees](https://cursor.com/docs/configuration/worktrees) - [Warp — how to run multiple AI coding agents](https://docs.warp.dev/guides/agent-workflows/how-to-run-multiple-ai-coding-agents/) - [Claude Code — worktrees](https://code.claude.com/docs/en/worktrees) - [Alex Chaplinsky — Git worktrees are not an isolation boundary for coding agents](https://fletch.sh/blog/git-worktrees-vs-clones-for-ai-agents/) - [Ally Piechowski — Your setup script should support Git worktrees](https://piechowski.io/post/setup-script-git-worktrees/) - [Mike McQuaid — Sandboxed agent worktrees](https://mikemcquaid.com/sandboxed-agent-worktrees-my-coding-and-ai-setup-in-2026/) - [copy-env — therohitdas](https://github.com/therohitdas/copy-env) - [Orca — Stably AI](https://github.com/stablyai/orca) **Two Quick Tips** - [Ankur Sethi — Prevent cognitive debt by manually retyping LLM-generated code](https://ankursethi.com/blog/prevent-cognitive-debt-by-manually-retyping-llm-generated-code/) - [Afek Berger, JFrog Security Research — SQLite Critical CVEs or LLM Slop?](https://research.jfrog.com/post/sqlite-critical-cves-or-llm-slops/) --- Sister show: **The AI Daily Wire** — the daily AI news brief from the same network. Different host, same standards. ## Built by **Culmen Games** — an indie game studio building an Android roster solo. studiosculmen@gmail.com **Handled** — done-for-you content and AI agent workflows for creators and time-poor professionals. handled.anmol@gmail.com

  6. Jul 30

    Episode 2: Let Your Agent Open the Browser

    Your coding agent doesn't have to guess about web bugs it can't see. Today we hand it a live Chrome browser with chrome-devtools-mcp, the free tool from the Chrome DevTools team, so it reads the real console errors and network requests instead of theorizing. Plus three fresh finds, two cost-and-safety tips, and one ten-minute thing to do today. ## In this episode - Fresh Find 1: Nano Banana 2 Lite, Google's cheapest image tier (about three and a half cents per 1K-resolution image, per one image, not per a thousand). - Fresh Find 2: Rowboat, an open-source desktop AI coworker with a memory of your work (Apache-2.0, free app, your own model keys). - Fresh Find 3: GPT Transcribe and GPT Live Transcribe, OpenAI's new speech-to-text (under half a cent a minute, and about 1.7 cents a minute for live). - The Recipe: let your agent debug a live web page with chrome-devtools-mcp. - Tip 1: route by task difficulty across one model family (OpenAI ~5x cheaper, Google ~8x cheaper, vendors' own list prices). - Tip 2: use Claude Code checkpoints and /rewind to roll back a bad agent edit (but keep committing to Git). - Do This Today: install chrome-devtools-mcp and point your agent at one real broken page, in under ten minutes. ## Sources - [chrome-devtools-mcp (Chrome DevTools team)](https://github.com/ChromeDevTools/chrome-devtools-mcp) - [chrome-devtools-mcp troubleshooting guide](https://raw.githubusercontent.com/ChromeDevTools/chrome-devtools-mcp/main/docs/troubleshooting.md) - [Google: Nano Banana 2 Lite announcement](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-omni-flash-nano-banana-2-lite/) - [Google Gemini API pricing](https://ai.google.dev/gemini-api/docs/pricing) - [Rowboat (Rowboat Labs)](https://github.com/rowboatlabs/rowboat) - [OpenAI release notes](https://openai.com/products/release-notes/) - [OpenAI API pricing](https://developers.openai.com/api/docs/pricing) - [Claude Code checkpointing docs](https://code.claude.com/docs/en/checkpointing) Sister show: The AI Daily Wire — the same network's daily AI news brief. Built by: - Culmen Games (indie game studio) — studiosculmen@gmail.com - Handled (done-for-you content + AI agent workflows) — handled.anmol@gmail.com

  7. Jul 28

    Episode 1: Welcome In - Four Terminal Agents and a Spend Cap

    The first episode of AI Workflow and Tips Daily. Why this show exists (short version: I spent years saving brilliant AI workflow posts I never had time to implement, and feeling behind because of it), then three finds from right across the ecosystem, a recipe that costs nothing, two tips, and one thing to do today that might save you real money. ## In this episode - **Fresh Finds** — SpaceXAI open-sourced **Grok Build**, its Rust terminal coding agent (Apache 2.0, and honest about what it costs to run); **Clawk**, which boots your agent into a throwaway Linux microVM with a default-deny network; and **i-have-adhd**, the skill file that makes any of eleven agent harnesses lead with the answer instead of the preamble. - **The Recipe** — **OpenWiki** from the LangChain team: repo documentation that generates itself, refreshes on a cron, and opens a pull request. Free if you point it at a local model. - **Two Quick Tips** — Google deprecated `temperature`, `top_p` and `top_k` starting with Gemini 3.6 Flash and 3.5 Flash-Lite, so audit your call sites; and write one `AGENTS.md` instead of eleven tool-specific config files. - **Do This Today** — set a *hard* monthly spend cap on your LLM API account, not just an alert. Notes on cost, because this show always says so out loud: **Grok Build's binary is free but running it is not** — it needs a browser login or an API key. **Claude Code** requires a paid Claude plan or API billing; there is no free tier. **Clawk, i-have-adhd and OpenWiki are all free and open source**, and OpenWiki runs at zero marginal cost against a local model. ## Sources - [Grok Build — GitHub](https://github.com/xai-org/grok-build) · [docs](https://docs.x.ai/build/overview) - [Clawk — Show HN](https://news.ycombinator.com/item?id=48892859) · [repo](https://github.com/clawkwork/clawk) - [i-have-adhd — GitHub](https://github.com/ayghri/i-have-adhd) · [install guide](https://github.com/ayghri/i-have-adhd/blob/main/INSTALL.md) - [OpenWiki — GitHub](https://github.com/langchain-ai/openwiki) · [LangChain announcement](https://www.langchain.com/blog/introducing-openwiki-an-open-source-agent-for-repo-documentation) - [Gemini API changelog](https://ai.google.dev/gemini-api/docs/changelog) · [scope of the deprecation](https://ai.google.dev/gemini-api/docs/latest-model) - [AGENTS.md](https://agents.md/) - [OpenAI hard spend limits — changelog](https://developers.openai.com/api/docs/changelog) - [Claude Code plan requirements](https://code.claude.com/docs/en/setup) ## Built by This show is made by a solo developer who builds things in his spare time. If any of it is useful to you, these are the projects behind it: - **Handled** — done-for-you content and AI agent workflows for creators and busy professionals. Get in touch: **handled.anmol@gmail.com** - **Culmen Games** — an indie Android game studio. Get in touch: **studiosculmen@gmail.com** Both are early-stage and built in the open. Questions, ideas, or just want to say hi? Either address above reaches me.

  • 7 Episodes

About

A short daily show for people who use AI at work. Each episode distills what practitioners are actually sharing — a new repo, a new skill, a workflow trick or tip — into one repeatable recipe you can run today, plus a quick tour of fresh finds worth bookmarking.

Information