apply
Set chaff up in the folder: chaff.yaml for the chosen genre, today's findings shelved in .chaff-baseline.json, and, when asked, a…
Build the source's actions and ingests the spec decided to build as Worker features, and write down the ones left to a person.
$ npx -y skills add receptron/mulmoterminal --skill actions-cloudflare --agent claude-codeHow it fires
How this skill gets triggered: by you, by Claude, or both.
/actions-cloudflareContext preview
The summary Claude sees to decide when to auto-load this skill.
Build the source's actions and ingests the spec decided to build as Worker features, and write down the ones left to a person.
name: blueprint-from-collection-actions description: "Build the source's actions and ingests the spec decided to build as Worker features, and write down the ones left to a person."
Read `.blueprint/actions.json` (written with the spec, and agreed when the spec was approved). If it does not exist, the source had no actions or ingests: stop. Each entry names an action (`books.actions.tidy`) or an ingest (`books.ingest`), what it was (`kind`) and what was decided (`decision`). The original instructions are in `.blueprint/source/collections/<slug>/templates/`.
For every `feature`, build it as below, unless an earlier step already built it because a must-have needed it: then check it against its entry and give it its test here instead of building it a second time.
API's entry; its `set` is the change it makes; its `params` are the inputs the button asks for.
Where a step genuinely needs a language model (summarising, classifying, drafting text), call the Claude API from the Worker (`@anthropic-ai/sdk` runs there), with the model id in a `vars` entry of `wrangler.jsonc` and the key as the secret `ANTHROPIC_API_KEY`: locally in `.dev.vars`, in production set by the person with `yarn wrangler secret put ANTHROPIC_API_KEY`. Before anything goes in `.dev.vars`, make sure `.gitignore` has a line `.dev.vars` — add it if not; the check fails otherwise. Never write the key's value yourself: tell the person, in the README, where to put their own key. Everything that does not need a model stays plain code. The screen shows the result and any failure in words.
inserts or updates records by the `idFrom` it names, never duplicating. Run it from a Cron Trigger (`"triggers": { "crons": [...] }` in `wrangler.jsonc`, and a `scheduled` handler in the Worker) on the schedule the ingest declares; say in the README how to run it once by hand (`yarn wrangler dev --test-scheduled`, then open `/__scheduled`). An ingest of kind `agent` is built like an agent action.
Tests in `test/actions.test.ts`, one per `feature`, each with the entry's `name` in the title of its `it(…)` / `test(…)` (`it("books.actions.tidy: summarises", …)`) — with `it` / `test` imported from `vitest` in that file (`import { it } from "vitest"`); the check reads titles only from calls to those imports, so a global `it`, a helper of the same name, a comment or a title built at run time does not count, and neither does a `describe(…)` name (`it("tidy")` inside `describe("books.actions")` is not titled `books.actions.tidy`: write the whole name in every test's own title): walk the real flow through the Worker on the test D1 and read back what changed. Never call the network in a test: a model call goes through a client the test replaces with a stand-in, and an ingest job takes the fetch it uses as an argument, so the test hands it a fixture instead. The tests must pass with no API key set.
For every `manual`: a section in `README.md` whose heading names the entry (`## books.actions.tidy`) and which says, step by step, what a person does instead. For every `drop`: nothing to build; the spec already says why.
Done when the check passes: `.blueprint/actions.json` matches the source, every `feature` has a test titled with its name, every `manual` has a README heading naming it, a `.dev.vars` (if any) is ignored, and `yarn build` and `yarn test` succeed.
Code's trust prompt. The user adds git themselves after the build if they want it.
Run multiple coding-agent sessions in parallel — and see which one needs you. A browser terminal for parallel AI coding agents: several sessions side by side, each in its own cell, with the one that needs you marked in colour.
Set chaff up in the folder: chaff.yaml for the chosen genre, today's findings shelved in .chaff-baseline.json, and, when asked, a…
Measure what chaff reports on the named documents today, as the kind the person chose, and propose the setup — changing nothing…
Answer each question from the named documents only, quoting where the answer is written, or saying plainly that the documents do…
Add the approved answers to FAQ.md when the person chose to keep them, after what it already holds; change nothing otherwise.
Build the spec's operations as the Worker's API, with validation and tests.