We’re building Ever Works in public. Some things may be incomplete, missing, or broken while we continue improving the platform.We’re building Ever Works in public — expect a few rough edges.
Build & ship · Steer the AI
Steering the runtime should not require a fork. Ever Works exposes the prompts at every level as ordinary text you write and keep: seven per-step overrides on a Work, five definition files per Agent, Skills any Agent can pull in, and prompt defaults carried by templates. Nothing here is a hidden setting or a support ticket — you write a sentence, save it, and the next run behaves differently.
One field per step of the generation pipeline, set per Work on its Settings tab under the Standard Pipeline card. Each is appended to our prompt for that step under an “Additional User Instructions” separator, so the default behaviour stays and your rule rides on top — up to 2,000 characters each.
Decide what counts as on-topic before anything is written. “Only include tools that are actively maintained and have more than 100 GitHub stars” narrows every source the run is willing to consider.
Steers the items the AI proposes before the web search runs. Use it to fix the shape of an entry: “each item should be a specific library or framework, not a general concept or methodology”.
The itemExtraction field — what gets pulled out of a page once it has been fetched, and which metadata comes with it. “Always extract the programming language and license type as tags.”
Changes what the runtime goes looking for in the first place. “Include queries for GitHub repositories and NPM packages, not just blog posts” rewrites the whole discovery pass.
Pin your taxonomy so it stops drifting between runs. Name the categories that are allowed, forbid new ones, and every future run files items the same way.
Tune the merge threshold in words rather than numbers. “Consider items with different major versions as separate entries” keeps a distinction the default would quietly collapse.
Say which URLs may stand as an item’s official source. “Only accept GitHub repository URLs or official documentation sites; reject blog posts and tutorials.”
Prompt overrides steer one Work’s pipeline. These steer the workers, the recipes, and the templates everything starts from — all of it plain text or YAML you can read.
SOUL.md is who the Agent is, AGENTS.md its house rules, HEARTBEAT.md what to do on a scheduled tick, TOOLS.md which tools it leans on, and agent.yml its metadata. Edit them in the Agent’s Instructions tab — five pills, autosave, 64 KB a file — and every write is secret-scanned and logged with a diff.
A Skill is a SKILL.md: YAML frontmatter with a name and description, then the instructions. Give one an invocation slug and typing /plan at the start of a task-chat message injects that Skill into the run, with the composer autocompleting the slugs. Companion files ride along — up to 20 per Skill, 2 MiB each, as references, configs, scripts or assets.
When you fork a base template you can describe the look you want and an agent edits the theme for you. The commit is deliberately restricted to one file — apps/web/src/styles/theme.css — so the fork keeps compiling whatever the agent does. Three base templates carry a customisation prompt today: minimal, web and web-minimal.
A Mission template repository can carry a .works/mission.yml manifest with the defaults a spawned Mission inherits — a cron cadence, guardrails (max Works and budget per run, whether creation needs approval), an outstanding-Ideas cap, seed knowledge-base files copied into the new Mission’s repo at fork time, and the Work templates the Idea→Work scaffolder pre-picks. Your own create-time values win over the manifest, and the manifest is optional: a template without one spawns a Mission on plain defaults.
The whole loop, from typing a sentence to seeing it change a run.
Go to the Work you want to steer, open its Settings tab, and expand the Standard Pipeline card. The seven fields load on first expand, each with its own description and a worked example in the placeholder.
Pick the step you want to change and write a directive sentence or two — inclusion criteria, preferred sources, the categories you allow. Short and specific beats long and hedged, and 2,000 characters is the ceiling.
One button writes the whole set. Editors and above can change prompts, viewers can read them, and the same state is reachable over the API with PUT /api/works/:id/advanced-prompts.
The next generation run appends your text to that step’s prompt, and the Work’s History tab records what the run added, updated and removed — so you can tell whether the instruction did what you wanted before you put it on a schedule.
None of these fields overwrite our prompts — they are appended to them. That is what makes them safe to experiment with: a badly worded override narrows or widens a step, it does not strip out the behaviour underneath, and clearing the field puts you back on the default in one save. The same principle runs through the Agent files and Skills — your text is added to a working system, kept where you can read it, and reversible.
Author recipes the runtime can execute. Reusable, versioned, your library of agent know-how.
Learn more →Named AI employees — a CEO, a CTO, a Researcher — that run your Missions, Ideas, and Works around the clock.
Learn more →Write copy, posts, listings, and pages at scale — with editorial voice, sources, and structure.
Learn more →Agents read the topic, gather sources, and produce briefs the writing pass uses.
Learn more →Every catalog Ever Works ships — website and Work templates, Mission playbooks, twelve starter Agents, a public Skills library and more than thirty prebuilt companies.
Learn more →Describe what you want to build. Ever Works researches it, ships it, and keeps it improving — content and code, owned in your own Git.