Most “AI isn’t working” stories are mode errors. Someone ran a chat when they needed a workflow, or they turned on an agent when a two-line prompt would have been safer.
This page is the fork in the road for the rest of Bye Prompt.
Definitions that survive vendor slides
| Mode | What it is | What it produces | When it is the right default |
|---|---|---|---|
| Chat | Turn-taking with a model in a thread | Text (and maybe files you copy out) | Exploration, rubber-ducking, one-off questions |
| Workflow | A saved sequence: brief → generate → critique → finish | A repeatable artifact | Anything you will do more than twice |
| Agent | A model that calls tools and continues until a stop condition | Side effects in the world (files, tickets, browsers, money) | Multi-step work you can bound and inspect |
If the vendor calls everything an “agent,” ignore the label and ask: does it take actions without me clicking each step? If no, it is chat or a workflow UI.
Chat — underrated, overused
Chat is the right tool when:
- You do not yet know the question
- The cost of a wrong answer is a shrug
- You want a sounding board, not a commit
Chat fails when:
- The thread is three days long and nobody can find the spec
- You paste a new dump of context every morning
- You needed a file, a PR, or a purchase — and you still have a conversation
Rule: if you would be embarrassed to show the thread to a colleague as the source of truth, it is not the source of truth. Lift the decision into a context pack or a doc.
Workflow — the missing middle
A workflow is a checklist the model participates in, not a personality.
Minimum viable workflow:
- Brief (what must be true / must never be true)
- Inputs (files, URLs, examples)
- Generate
- Critique against the brief (second pass, second model, or a rubric)
- Finish outside the generator (edit, test, export)
- Archive the brief + result so the next run is cheaper
That is the same skeleton Create By Prompt uses for images and video. It is not a creative-only idea. Research, ops, and buying use the same stages. See loops.
Workflows beat chat because the brief survives the tab. They beat naïve agents because you still gate the dangerous steps.
Agent — power with a blast radius
An agent is appropriate when:
- The job is many small steps (open, read, compare, write, re-read)
- You can name tools it may use
- You can name stop conditions (“open a PR,” “draft only,” “spend ≤ $X”)
- You will look at the diff
An agent is a bad default when:
- You cannot undo the action (wire, public post, production schema)
- You have not written what “done” means (eval)
- The environment is a pile of secrets (don’t paste them)
Computer-use demos look magical because they hide the blast radius. Your job is to draw the fence first: human in the loop, what not to automate.
A 30-second picker
Ask, in order:
- Is this a question or a job? Question → chat. Job → continue.
- Will I do this again? Yes → workflow (save the brief). No → chat unless there are side effects.
- Are there side effects in tools or money? Yes → agent only with caps. No → workflow is enough.
- Can I inspect the result in under five minutes? If no, the job is too big. Split it.
Worked examples
“Summarize this PDF.” Chat or a one-step workflow. Not an agent.
“Every Monday, draft a competitor note from these five URLs, then wait for me.” Workflow. An agent can fetch; a human ships.
“Find three laptops under $1,200 that fit this spec, fill a table, do not buy.” Agent-shaped research with a hard stop. Buying is a different page: pattern: buying and Buy By Prompt.
“Pay the invoice.” Not a prompt-box problem. Identity, caps, rails: Pay By Prompt.
How this maps to the rest of the site
- You still type: when you still type a prompt
- You assemble inputs: context packs
- You compile instructions: compiled specs
- You run cycles: loops
- You prove done: eval
If a product cannot tell you which mode you are in, assume chat with extra buttons until you see a tool call and a stop condition.