LLaManchaAI Enablement
Checking access…
← All modules
Module 6Audience: all

From role/persona to agent workflow brief

Shows how role/persona cards can become workflow briefs for AI assistants, agentic tools, and future automation.

Outcomes

What you will be able to do

  • Decide when a workflow deserves a brief instead of a one-off prompt.
  • Scope a trigger, allowed actions, and prohibited actions precisely.
  • Design a human approval gate and an escalation path.
  • Recognize the common ways agent workflows fail.
Completion check

How this module is approved

Convert one persona workflow into a sanitized ten-field agent workflow brief.

Pass criteria

  • • Trigger is event + precondition + fallback
  • • Allowed and prohibited actions both present and specific
  • • Output format and validation checklist present
  • • Human approval gate and escalation path named
  • • Success measure stated
Loading personalization context…
Lesson35–45 minutes self-guided, or 60 minutes with a guided decomposition

What you should take away

By the end of this module, you will be able to take one workflow from your role persona and turn it into a workflow brief precise enough that an AI assistant — or a future agent — can help with it safely, with a human still holding the decision.

Prerequisites: Complete Module 5

Part 1

Start here: two ways to ask for the same automation

Read these two requests for the same outcome. The first is how people usually ask. The second is a workflow brief. Only the second is safe to hand to an assistant that can take actions.

  • Vague: 'Automate my weekly report — pull the numbers and send it out.'
  • Brief: 'Trigger: Monday 9am, when all five sanitized status notes are in the shared doc. Allowed: draft the narrative and risk list. Prohibited: send anything, edit the source notes, invent metrics. Output: draft in the doc, status = needs review. Approval: my manager posts it. Escalate: if a note is missing, stop and flag me.'
  • The vague version hides three dangerous instructions: send, pull, and a missing definition of done. The brief makes every one of them explicit and bounded.

Quick check · <30 sec

What makes the vague version unsafe to give an action-taking assistant?

  • A. It is too short
  • B. It grants 'send' and 'pull' with no boundary, approval gate, or stop condition
  • C. It does not name a tool
  • D. It is not polite
Show answer
Unsafe agent instructions are usually unsafe because of unbounded actions and a missing approval gate — not length or phrasing.

Part 2

When a workflow deserves a brief — the smallest safe workflow

Not every task needs a brief. A one-off, low-stakes ask is fine as a Module 2 prompt. A brief earns its cost when the workflow is recurring, multi-step, and could take an action that affects someone else. The heuristic: start with the smallest safe slice of the workflow — the part where AI assists and a human still presses the button.

  • One-off + reviewable → a prompt is enough.
  • Recurring + multi-step + has a side effect → write a brief.
  • Smallest safe workflow: cut the workflow at the last point before an irreversible or outward-facing action, and put the human there.
  • You can always widen scope later; you cannot un-send a message an over-scoped agent already sent.

Quick check · <30 sec

Where should the 'smallest safe workflow' boundary be drawn?

  • A. At the very start, before any AI involvement
  • B. Just before the first irreversible or outward-facing action
  • C. After the action, at the cleanup step
  • D. Wherever is fastest
Show answer
Put the human at the last point before an action becomes irreversible or visible to others. AI assists up to that line; the person crosses it.

Part 3

Scoping the trigger and the required context

A workflow brief has ten fields. The first two — Trigger and Required context — decide whether the rest can work. A vague trigger ('when needed') is one of the top agent failure modes: the assistant either never fires or fires constantly. A precise trigger names the event, the precondition, and what to do if the precondition is not met.

  • Trigger = event + precondition + missing-precondition behavior. 'Monday 9am, if all five notes present; if not, stop and flag.'
  • Required context = the sanitized inputs the workflow needs, named as categories, pulled from your persona card.
  • If you cannot state the trigger in one sentence, the workflow is not scoped enough to brief yet.

Quick check · <30 sec

Rewrite the trigger 'run it when the report is due' so it is brief-grade.

Show answer
A brief-grade trigger names the event, the precondition, and the fallback, e.g. 'On the last business day of the month, when the sanitized metrics sheet is marked final; if it is not final by 3pm, stop and notify me.'

Part 4

Allowed actions, prohibited actions, and the approval gate

This is the safety core of the brief. Allowed actions should be the shortest list that does the work. Prohibited actions should explicitly name the tempting-but-dangerous ones (send, delete, edit source, decide, spend). The approval gate names the human who turns a draft into a real action, and what they check.

  • Allowed: list only what is needed — usually 'draft', 'summarize', 'propose'.
  • Prohibited: name the dangerous verbs explicitly, even if 'allowed' doesn't mention them. Silence is not a boundary.
  • Approval gate: a named role + the specific thing they verify before the action happens.
  • Validation checklist + success measure: how the human knows the output is good, and how you'll know the workflow is worth keeping.
Activity · ~10 minworkflow brief design

Take one recurring workflow from your Module 5 persona card. You will decompose it into the ten brief fields. Keep everything sanitized.

  • Source: One recurring workflow from your persona card
  • Constraint: Allowed-actions list must be the shortest that still does the work

Your task

Fill in the ten fields for your workflow: (1) Trigger (event + precondition + fallback) (2) Required context (3) Allowed actions (4) Prohibited actions (name the dangerous verbs) (5) Suggested AI steps (6) Output format (7) Validation checklist (8) Human approval gate (who + what they check) (9) Escalation rules (10) Success measure. Then stress-test it: name one way this brief could still go wrong.

Show a hint
If your prohibited list is empty, you have not finished. Every useful workflow has a tempting dangerous action — name it.
Compare with a strong answer
Trigger: Monday 9am, all five sanitized status notes present; if any missing, stop and flag me. Context: the five notes, categories only. Allowed: draft narrative + risk list. Prohibited: send, post, edit source notes, invent metrics or owners. AI steps: cluster themes, draft 150-word update, list top three risks. Output: draft appended to the working doc, marked 'needs review'. Validation: every metric traces to a note; no invented names/dates. Approval: my manager posts to the leadership channel. Escalation: budget-impacting items go to finance first. Success measure: manager edits drop below ~10% over a month. Could still go wrong: a note that is present but stale would pass the trigger — add a freshness check.

Why this matters: Decomposition is where over-permissioning becomes visible. Writing the prohibited list by hand is the moment most people catch the dangerous action they were about to grant.

Quick check · <30 sec

Why must the prohibited-actions list name dangerous verbs even if the allowed list doesn't mention them?

  • A. For documentation completeness
  • B. Because an unstated action is not a bounded action — silence gets interpreted, not refused
  • C. To make the brief longer
  • D. It does not need to
Show answer
An action that is neither explicitly allowed nor explicitly prohibited is ambiguous. Naming the dangerous verbs turns ambiguity into a hard boundary.

Part 5

How agent workflows actually fail

Most agent failures are not exotic. They cluster into four patterns, and the brief is designed to defeat each one. Knowing the patterns lets you read your own brief adversarially before anyone runs it.

  • Over-permissioned: 'allowed' is broader than the work needs. Fix: shortest allowed list, explicit prohibited list.
  • Vague trigger: fires never or always. Fix: event + precondition + fallback.
  • No rollback or escalation: the workflow has no defined behavior when something is wrong. Fix: escalation rules + a stop condition.
  • No success measure: nobody can tell if the workflow is helping. Fix: a measurable signal you'll review on a cadence.

Quick check · <30 sec

An agent that 'sometimes does extra things nobody asked for' is most likely suffering from which failure pattern?

  • A. Vague trigger
  • B. Over-permissioning
  • C. No success measure
  • D. No rollback
Show answer
Doing un-asked-for actions is the signature of an over-permissioned brief — the allowed scope is wider than the task.

End-of-module quick check

Five short retrieval questions. Answer from memory first, then reveal each explanation.

  1. 1. A workflow earns a brief (instead of a one-off prompt) when it is…

    • A. Long
    • B. Recurring, multi-step, and able to take an action with a side effect
    • C. Technical
    • D. Customer-facing only
    Show answer
    Briefs pay off for recurring, multi-step workflows that can affect someone else. One-off reviewable asks stay as prompts.

  2. 2. How many canonical fields are in a workflow brief?

    • A. Five
    • B. Eight
    • C. Ten
    • D. Twelve
    Show answer
    Ten: trigger, required context, allowed actions, prohibited actions, suggested AI steps, output format, validation checklist, human approval gate, escalation rules, success measure.

  3. 3. The 'smallest safe workflow' heuristic says to put the human…

    • A. At the end, to clean up
    • B. Just before the first irreversible or outward-facing action
    • C. Nowhere; that's the point of automation
    • D. At every step
    Show answer
    AI assists up to the last point before an action becomes irreversible or visible; the person crosses that line.

  4. 4. True or false: an empty prohibited-actions list is fine if the allowed list is short.

    • A. True
    • B. False
    Show answer
    False. An unstated action is ambiguous, not refused. Every useful workflow has a tempting dangerous verb that must be named.

  5. 5. Name the four common agent failure patterns the brief is designed to defeat.

    Show answer
    Over-permissioning, vague trigger, no rollback/escalation, and no success measure.

Further reading

Worked examples by role

Operations manager

Workflow brief: weekly leadership update

Trigger: Monday 9am if five sanitized notes present, else stop and flag. Allowed: draft narrative + risks. Prohibited: send, post, edit sources, invent metrics. Output: draft in working doc, 'needs review'. Approval: manager posts. Escalation: budget items to finance. Success: manager edits under ~10% over a month.

Sales coordinator

Workflow brief: follow-up email first draft

Trigger: a deal moves to 'awaiting follow-up' in the approved CRM. Allowed: draft an email from sanitized account context. Prohibited: send, alter CRM fields, state competitor or pricing claims. Output: draft in CRM notes. Approval: account owner sends after fact-check. Escalation: pricing questions to the deal desk. Success: fewer than one factual correction per ten drafts.

Software developer

Workflow brief: PR test scaffolding

Trigger: a PR is opened with a function lacking unit tests. Allowed: propose test cases as a comment. Prohibited: push commits, approve the PR, modify CI config, touch secrets. Output: a review comment with suggested tests. Approval: author commits tests, reviewer approves. Escalation: security-relevant code to the security reviewer. Success: suggested tests accepted with minor edits more than half the time.

Before / after

Same workflow: vague ask vs. ten-field brief

Before: 'Have the assistant handle the weekly update.' Unbounded actions, no trigger precondition, no approval gate, no definition of done.

After: A ten-field brief: precise trigger with a fallback, a two-item allowed list, an explicit prohibited list naming 'send', a named approval gate, escalation rules, and a measurable success signal.

What changed: The brief converted three hidden dangerous instructions (send, pull, 'handle') into bounded, reviewable steps with a human at the action.

Completion artifact

Submit a sanitized workflow brief for one persona workflow, with all ten fields and at least one named prohibited action and a human approval gate.

ExercisePilot-ready artifact

Convert one workflow from your Module 5 persona card into a sanitized ten-field workflow brief. Then write one sentence on how it could still fail.

Participant template

  • Trigger (event + precondition + fallback):
  • Required context (sanitized categories):
  • Allowed actions:
  • Prohibited actions:
  • Suggested AI steps:
  • Output format:
  • Validation checklist:
  • Human approval gate (who + what they check):
  • Escalation rules:
  • Success measure:
  • One way this could still go wrong:

Example submission

Trigger: Monday 9am, five sanitized notes present; if any missing, stop and flag me. Context: the five notes (categories only). Allowed: draft narrative + risk list. Prohibited: send, post, edit source notes, invent metrics/owners. AI steps: cluster, draft 150-word update, list top three risks. Output: draft in working doc, 'needs review'. Validation: every figure traces to a note. Approval: manager posts to leadership channel. Escalation: budget items to finance first. Success: manager edits under ~10% over a month. Could still go wrong: a stale-but-present note would pass the trigger — add a freshness check.

Role-flavored variants

Same exercise, framed for different roles. Use the one closest to your work.

Front office coordinator

Brief a workflow that touches people-facing communication. The prohibited list must forbid sending and any commitment; the escalation must route judgment calls to a human, not AI.

See a sample submission
Front office brief (excerpt): Trigger: end of day, when the voicemail queue is non-empty. Allowed: draft a prioritized callback list. Prohibited: call, message, or commit appointment times; decide clinical priority. Output: draft list marked 'needs review'. Approval: front-desk lead each morning. Escalation: anything clinical goes to a nurse, not AI. Success: lead reorders fewer than two items per day.

Training manager

Brief a content-generation workflow. The validation checklist must include a bias and accessibility pass; prohibited actions must forbid publishing without that review.

See a sample submission
Training manager brief (excerpt): Trigger: a new learning objective is approved. Allowed: draft comprehension questions. Prohibited: publish to the LMS, lock the rubric without an instructional-design pass, skip the bias/accessibility check. Output: draft question set. Validation: reading level, bias scan, accessibility pass. Approval: I review before publish. Escalation: flagged bias to instructional design. Success: fewer than one reworked item per set.

Learner checklist

Use this as a final check before submitting. Program leads use a separate review guide when they approve or coach submissions.

  • • Trigger is event + precondition + fallback
  • • Allowed and prohibited actions both present and specific
  • • Output format and validation checklist present
  • • Human approval gate and escalation path named
  • • Success measure stated
Loading your module status…

Previous module

← Role/persona/onboarding context

Review the prior step in the path.

Next module

Team adoption and working agreement →

Keep momentum with the next completion check.