Skip to main content
A work plan is an ordered sequence of steps Emma runs on request. Each step has an instruction, and steps carry state — pending, completed, skipped — so a plan advances deliberately instead of in one shapeless burst. Emma can already do most of this work conversationally. What a plan adds is order and checkpoints: the same steps, in the same sequence, with the same places where you get to look before she continues.

​
Start with the built-in plans

Go to New Chat → Select a workplan. These plans are built and vetted by professional marketers and founders — they encode how experienced operators actually sequence this work, not a generic template. If a built-in plan covers what you’re doing, run it before building your own. It’ll be better than a first draft you write from scratch, and you can always duplicate and adapt it later. Two things worth knowing about the picker:
  • It starts a new chat. A plan launched from the picker can’t see the conversation you were just having. If you want to build on the current thread, don’t launch from here — see the recommended creation path below.
  • The default view shows curated AgentWeb plans. The All tab shows everything you have access to, including plans you’ve created yourself and any shared with you.
Launching a plan while another is already running replaces it. If you’re mid-run, finish or abandon deliberately — the in-progress plan’s state doesn’t survive.
You can also launch by name in chat, but plan titles resolve on an exact match and several are near-identical. The picker is more reliable.
Answer the question Emma actually asked.As a plan runs, Emma asks for specific inputs — usually in bold. Those questions are load-bearing: each step is written to expect a particular answer, and the steps after it build on what you give.Answering a different question than the one asked — jumping ahead, adding unrelated context, or redirecting to something else you want — is the most common way a run goes off-script. Emma follows you off the plan, and the output quality drops well before it becomes obvious something went wrong.If a question doesn’t fit your situation, say so directly (“we don’t run paid ads”) rather than answering a nearby question instead. If you want to change direction, finish the step first, then redirect. Save the tangents for a normal chat.

​
Create a work plan

The best work plans come from work you’ve already done. When a multi-step conversation produces a result you’re happy with, say “make this as a workplan” in that same thread. Emma builds the plan from what just happened.
This is reconstruction, not recording. Emma reads back over the thread and composes what she judges the repeatable shape to be. Nothing captures your tool calls, your step order, or your mid-conversation corrections mechanically.The plan is also saved immediately, as private — there’s no draft state. So the workflow is save, then review, not review then save.
Check these four things right after conversion:
  • Your mid-thread corrections. “Shorter.” “Drop the emoji.” Those shaped the output but live in the conversation, not in the steps — unless Emma wrote them into the step text.
  • The model. If the thread’s quality depended on a particular model, that isn’t carried over. Set it on the plan.
  • Per-step tools and skills. These aren’t populated automatically from what the thread used. Add them where a step genuinely depends on reading real data.
  • Attachments. File links need to be written into the step text, or a later run can’t reach them.
What is carried correctly: your brand guidelines and market strategy are re-read every run, so the plan stays current rather than freezing stale context. This is also why you shouldn’t convert too early. A messy exploratory thread reconstructs into messy steps. Convert once the pattern has settled.

​
Manual: build one from scratch

Go to Settings → Workplans and create the plan directly. Better when you’re adapting an existing plan, or when you want precise control over per-step fields from the start. Either way, your plans appear under New Chat → Select a workplan → All.

​
Anatomy of a work plan

Plan level: title, description, welcome message, default model, autocomplete, visibility. Step level — each step carries its own instruction, plus:
Required function calls are a completion gate, not a permission allowlist. Naming a tool there forces Emma to call it before the step can finish. It does not restrict which other tools she may use, and adding names won’t make a run faster or more contained.Use it to guarantee grounding — requiring read_brand_guidelines so a copy step can’t proceed on vibes. If you want a step to stay narrow, write a narrower instruction.
Autocomplete decides whether Emma advances without pausing. On is fine for research and reports. Off — so she stops between steps — is what you want for anything that sends, posts, or spends.

​
Work plans vs. skills

These get conflated constantly. A skill is a capability: tools plus standing rules for a domain. A work plan is a sequence: ordered steps with gates. A skill answers can Emma do this; a plan answers in what order, with what checkpoints. Plans don’t grant tools — they install skills that do, declared per step. Mature capabilities often ship as both.

​
Writing steps that hold up

1

Write instructions, not topics

“Summarize competitor activity” is a topic. “Pull the last 30 days of activity for each competitor in Market Strategy, then summarize what changed and why it matters” is an instruction. Vague steps are the main cause of output that varies run to run.
2

Give step one a stop condition

If the plan depends on external data — an event page, an ad account, a search — say what Emma should do when there’s nothing there. Without it, plans burn tool calls retrying dead inputs instead of reporting the gap.
3

Gate the steps that need grounding

Add required function calls where a step must read real data before producing output.
4

Run it once and watch

The first run exposes which wording is too loose. Tighten before you share it with anyone.

​
Scheduling and unattended runs

Work plans have no schedule of their own. To run one on a cadence, create a scheduled chat whose prompt launches the plan.
Unattended runs behave differently. Scheduled runs proceed on a best-effort basis rather than stopping to ask for confirmation. A plan whose safety depends on you saying yes may never get the chance to ask.Before scheduling a plan, check whether any step sends, posts, or spends. If so, either restructure it to end at a draft and leave the send to you, or don’t schedule it.

​
Editing, duplicating, sharing

Editing is destructive. Saving a plan’s steps replaces the whole list — settings on steps you didn’t intend to touch can be lost, and there’s no version history. Duplicate before reworking anything you rely on. If you belong to an organization, plans are shared across it by default. A plan you create is visible to every member as soon as it is saved. Sharing is a single org-wide toggle at Settings → Organization → Org-wide Sharing → Work Plans, editable only by an org owner or admin — it cannot be set per plan. Because a plan’s full text is visible to your teammates, don’t put sensitive material — internal pricing, customer names, credentials — into a plan’s instructions. Put it in the Knowledge Base and have the plan reference it instead.

​
Troubleshooting

The picker starts a new chat by design. To build on an existing thread, use “make this as a workplan” in that thread instead of launching from the picker.
Titles resolve exactly and several are near-identical. Use the picker rather than launching by name.
Conversion reconstructs the workflow rather than replaying it, so corrections you made mid-thread may not have made it into the steps. Open the plan and add them as step text.
Usually an answer that didn’t match the question asked. Emma follows the conversation, so a tangent mid-run pulls the whole plan with it. Relaunch the plan and answer each bolded question directly; hold other requests until the run finishes.
The instruction is too vague. Specify fields, format, and length. If format matters, include a short example.
Usually an empty input the plan won’t stop on — a URL with no data, an ad account with no spend in the window. Add an explicit stop condition to the fetching step.
It likely has a required function call that hasn’t happened, often because the underlying account isn’t connected. Check the step’s required calls and the relevant integration.
Plans that read brand guidelines produce weak output when those guidelines are thin. Fill in Brand Hub first.
Work-plan sharing is on by default, so this usually means an owner or admin turned it off for the whole organization. Ask them to check Settings → Organization → Org-wide Sharing → Work Plans — see Organization sharing.

​
Next steps

Last modified on August 10, 2026