workflow-creator
Turn a repeatable multi-step task into a runnable, reusable saved workflow through guided conversation, in either of two shapes — chaining existing skills/recordings (wiring each one's output into the next's input), or composing the workflow script's own primitives (agent/parallel/pipeline) directly when the task needs fresh sub-agent orchestration and no existing skill covers it. Figure out which shape fits, generate the Ruby workflow, dry-run it, and save it with workflow_save. Use when the user wants to combine / chain / orchestrate work into one repeatable flow — whether that's several EXISTING skills, or a from-scratch multi-agent script (parallel review, fan-out research, a pipeline over a list) — e.g. "chain these skills", "把这几个技能连起来", "串个流程", "做个 workflow", "编排一下", "build a workflow", "写个并行跑多个 agent 的脚本". Do NOT use to author a single new skill (that is skill-creator), to run one skill/agent a single time (just call it), or to design a recurring/self-triggering loop with its ow
open-octo · Static checks on file
Installation
Read the source and dependencies before running a command. Copying does not install or execute anything.
npx skills add open-octo/octo-agent --skill workflow-creatorAbout this skill
name: workflow-creator
system: true
description: Turn a repeatable multi-step task into a runnable, reusable saved workflow through guided conversation, in either of two shapes — chaining existing skills/recordings (wiring each one's output into the next's input), or composing the workflow script's own primitives (agent/parallel/pipeline) directly when the task needs fresh sub-agent orchestration and no existing skill covers it. Figure out which shape fits, generate the Ruby workflow, dry-run it, and save it with workflow_save. Use when the user wants to combine / chain / orchestrate work into one repeatable flow — whether that's several EXISTING skills, or a from-scratch multi-agent script (parallel review, fan-out research, a pipeline over a list) — e.g. "chain these skills", "把这几个技能连起来", "串个流程", "做个 workflow", "编排一下", "build a workflow", "写个并行跑多个 agent 的脚本". Do NOT use to author a single new skill (that is skill-creator), to run one skill/agent a single time (just call it), or to design a recurring/self-triggering loop with its own trigger and state file (use cron-task-creator to schedule a saved workflow instead — a saved workflow is just the one-shot script such a scheduled run calls, not the loop itself).
Build a saved workflow
A saved workflow is a small Ruby (mruby) script that runs on demand or on a schedule. Your job is to guide the user from a one-off description of what they want repeated to a saved, validated workflow. It takes one of two shapes, sometimes mixed in the same script:
- Skill chain — the script calls existing skills/recordings in order via
recording(...)/skill(...), passing each one's output to the next's input. You are composing things that already exist. - Primitive-composed — the script calls
agent(...)directly (alone, or fanned out withparallel/pipeline) to do fresh sub-agent work that has no matching existing skill — e.g. "review this diff across 3 dimensions in parallel" or "run this check over every file in the list." There is nothing to inventory here; you're writing the orchestration from scratch, just not the agent-level logic (that's still an LLM call insideagent(...), not new Go/tool code).
Figure out which shape (or mix) fits before drafting anything — see Step 0.
The pieces you use
agent(prompt, opts = {})— run one sub-agent to completion, returns a String.opts[:schema](a JSON-schema string) makes the reply come back as a JSON string matching it — parse it yourself withJSON.parse, unlikeskill()'s schema results, which arrive already parsed.skill(name, params = {}, opts = {})— runs one existing SKILL.md skill as a sub-agent and returns its result as native Ruby.opts[:schema]makes the reply come back structured (already parsed).recording(name, params = {})— replays one existing browser recording deterministically and returns its declared outputs as a RubyHash.args— the workflow's input, a RubyHash, so the saved workflow is parameterizable and reusable.parallel(items) { |it| ... }/pipeline(items, *stages)— for fan-out or staged flows, overagent()/skill()/recording()calls.log(msg)/phase(title)— progress output, no effect on scheduling.- The
workflowtool — runs a script (in the background: returns a run id). workflow_save— persists the script as a named workflow.
Steps
0. Decide the shape
First check this isn't actually a loop in disguise: if the user describes
something that keeps running on its own (daily/hourly cadence, "keep checking
until X", needs to remember state between runs, needs a done-condition or a
human gate) — use cron-task-creator to schedule a saved workflow or build
stateful persistence yourself; that's not this skill's job. A saved
workflow is a single deterministic script; it has no opinion on when or how
often it's invoked. Redirect there instead of scoping cadence/state yourself.
Otherwise, ask what the workflow should do — its steps and their order, not the broader scenario or how often it runs — then place it:
- Every step maps to a skill/recording the user already has → skill chain, go to Step 1.
- The work is fresh sub-agent orchestration with no matching skill (a review
fanned out across dimensions, a search run in parallel over several sources, a
per-item pipeline) → primitive-composed, skip straight to Step 3 and write
agent/parallel/pipelinecalls directly — there's nothing to inventory or infer an output schema for, since you're writing the prompts yourself. - Some steps are existing skills and others are ad hoc agent work → do both:
inventory only the skill steps (Step 1), and for the primitive steps just
decide the prompt and, if a later step needs structured output from it, a
schema.
1. Inventory what's available (skill-chain steps only)
Read the system prompt's # Available skills (SKILL.md skills) and
# Browser recordings (recordings) sections. List the candidates relevant to the
user's goal — with their params, and, for recordings, their outputs — and ask
which ones, in what order.
Check for a name collision: one name can exist as both a recording and a
SKILL.md skill. That's no longer an error — recording("x") always picks the
recording, skill("x") always picks the skill — but it IS a readability trap
for whoever reads the script later. Flag it, and prefer renaming one of them.
2. Wire outputs → inputs (skill-chain steps only)
For each adjacent pair, propose how the earlier step's output feeds the next step's params, then have the user confirm before you generate anything. Be honest that the two skill types give you different amounts to work with:
- Upstream is a browser recording — its
outputsare in the manifest ([outputs: files (file[])]). Wire directly, e.g.dl["files"]→ the next step'sinputs. This is manifest-driven and reliable. - **Upst