Learning path · Level 9 · Tutorial 4

Build and Evaluate a Human-Reviewed AI Workflow

Build and evaluate a repeatable workflow for turning an approved content brief into a channel-specific draft. Use an optional free-tier assistant or text-based response examples, retaining human approval and comparing the result with a manual baseline.

  • Level: Advanced
  • Duration: 24 min
  • 8 steps
Start tutorial Back to the path
Build and Evaluate a Human-Reviewed AI Workflow

What you will go through

  1. Define the Workflow Contract Define a bounded human-reviewed workflow contract for the Harbor Workshops LinkedIn drafting task. 2 min
  2. Map Human Decisions and Handoffs Build a workflow map with named responsibilities, approval gates, and explicit return paths. 3 min
  3. Make the Workflow Reusable and Recoverable Package a human-reviewed workflow into reusable specifications, prompts, versioned artifacts, run records, and bounded fallback rules. 3 min
  4. Set the Comparison and Capture a Manual Baseline Establish a fixed, reviewable comparison protocol and capture a traceable manual baseline before any AI-assisted result is considered. 4 min
  5. Run, Revise, and Approve the Draft 4 min
  6. Calculate the Full Cost of the Run Calculate recurring and first-use effort differences from complete run records, keeping active work, elapsed duration, setup, and timing evidence distinct. 3 min
  7. Decide What the Evidence Supports Use quality, risk, complete effort, and evidence limits to make a proportionate workflow decision. 3 min
  8. Finalize Your Workflow and Decision Note Consolidate a reusable workflow package, verify its traceability, and write a bounded decision note grounded in the supplied comparison. 3 min

What you will learn

  • Document a bounded workflow with reusable prompts, input requirements, review responsibilities, approval gates, and fallback rules.
  • Walk a fictional marketing brief through the workflow and record revisions and approval decisions.
  • Compare manual and AI-assisted work using the same quality criteria and total effort, distinguishing observed timings from estimates.
  • Decide whether to adopt, revise, or abandon the workflow and explain the limits of the comparison.

Before you start

  • Choose Marketing Tasks for AI Assistance
  • Write Source-Grounded Marketing Prompts
  • Review AI-Generated Marketing Outputs

Step 1 of 8

Define the Workflow Contract

Define a bounded human-reviewed workflow contract for the Harbor Workshops LinkedIn drafting task.

Set the boundary first

A bounded workflow

A human-reviewed workflow is a repeatable, bounded sequence: eligible inputs enter, a Drafter may use limited AI assistance, humans make the decisions, and the workflow ends with an approved draft. It does not include publishing.

The contract makes that boundary explicit before anyone drafts. This prevents an assistant response, or an incomplete brief, from being treated as ready for use.

Where the workflow ends

Approved documents enter a workflow boundary; a checked draft exits it, while a social posting icon remains outside the boundary.

Inputs enter the workflow; the endpoint is an approved draft. Publication is outside the workflow.

Write the input and output contract

Define inputs and outputs

The input contract says what must be present before work begins. For this practice, the Workflow owner is the input owner: they confirm that the approved brief and source are complete for the requested channel.

Both the manual baseline and the AI-assisted run must create the same output bundle. Changing the deliverable between methods would make a later comparison unfair.

Harbor Workshops workflow contract

Use this contract for Harbor Workshops.

Required inputs and the common output bundle
PartRequirement
Approved sourcesource v1.0
Approved facts[F1] Plan Your First Content Calendar; [F2] free, 45-minute online workshop; [F3] solo business owners; [F4] topic selection and weekly scheduling; [F5] registration required; no replay
AudienceSolo business owners
ChannelLinkedIn
Input ownerWorkflow owner
Post requirementOne 80–120-word LinkedIn post in a warm, practical tone
Required CTARegister for the free workshop.
Also deliverA concise factual-claim-to-source checklist and any unresolved questions

Example

Keep evidence separate from copy

The finished post uses only reader-facing language; it does not show labels such as [F2]. The checklist may use those references to show where each factual claim came from.

At this stage, the contract defines the work. It does not yet define detailed review routing or prompt versions.

Check readiness before work begins

Apply entry conditions

A received brief is not automatically eligible. Before entry, confirm that the source is approved and current, the audience and channel are stated, the required output bundle is known, and the Workflow owner can identify the inputs.

For this fictional task, source v1.0 is approved and supplies the needed facts. The audience is solo business owners, the channel is LinkedIn, and the required bundle is specified. It is ready to enter the workflow.

Tip

Eligibility is a check, not a guess

If a condition is missing, pause before drafting and return for clarification or an approved source update. Do not fill a gap with an assumption or let an AI assistant invent it.

The endpoint remains an approved draft, not a live post, account action, or publication.

Practice: classify and admit

Classify the contract

Match each proposed requirement to its workflow classification.

Tap an item and then its matching pair.

Confirm entry readiness

Is the Harbor Workshops brief described in this step ready to enter the workflow?

Step 2 of 8

Map Human Decisions and Handoffs

Build a workflow map with named responsibilities, approval gates, and explicit return paths.

1. The workflow path

Map the route, not just the draft

A workflow map makes human judgment visible. It shows the normal route, the gates that stop unsuitable work, and the return paths when a decision is not positive.

For Harbor Workshops, the route begins only after the brief and approved source are ready. The workflow ends with an approved draft—not publication. Text may be copied manually between a document and an optional assistant; no integration, account access, or automatic publishing is involved.

Workflow map

A workflow diagram moves from input validation through drafting, checking, revision, rechecking, and final approval. Gates return incomplete inputs or route reviewed drafts to revision, rejection, or escalation. Publication sits outside the workflow boundary.

The map separates a reviewed, approved draft from any later publication activity.

2. People, artifacts, and decisions

Make each handoff actionable

At every handoff, record three things: who acts next, which artifact they receive, and which decision they must record. A handoff is incomplete if the next person has a draft but no stated decision to make.

The Reviewer and Approver are separate passes. A Reviewer may mark a draft Accept, Edit, Reject, or Escalate. Only the Approver records the final result: Approved or Not approved.

Responsibilities and recorded decisions

Use these roles consistently for the Harbor Workshops workflow.

Each role receives a defined artifact and leaves a traceable decision.
RoleReceivesActs and records
Workflow ownerBrief and approved sourceValidates readiness; returns incomplete input or resolves an escalation path
DrafterReady brief, source, and requirementsCreates or revises the output bundle; hands it to review
ReviewerDraft plus approved sourceChecks the draft; records Accept, Edit, Reject, or Escalate
ApproverReviewed draft and review decisionMakes the final approval pass: Approved or Not approved

Tip

Solo workflow discipline

In solo practice, perform all four roles as separate passes. This is self-review, not independent review: pause between passes and read the artifact from the next role's decision criteria.

3. Gates and return paths

Use explicit return paths

The input-readiness gate comes before any AI use. The Workflow owner confirms that the approved source and required output requirements are present. If a needed fact is missing, return the incomplete brief rather than asking the Drafter or an assistant to infer it.

After checking, a repairable problem follows the Edit path: the Reviewer identifies the issue, the Drafter revises, and the Reviewer rechecks. An unusable draft is marked Reject and cannot proceed to approval. An unresolved question is marked Escalate and goes to the Workflow owner for a decision.

Disposition routes

These paths protect the same endpoint even when the draft is produced manually or with optional AI assistance.

A disposition determines the next action; it is not merely a comment.
SituationRecorded dispositionNext owner and action
Required source fact is absentEscalateWorkflow owner returns the incomplete brief for completion before drafting
Supported draft has a fixable errorEditDrafter corrects the draft; Reviewer rechecks it
Draft is unusable for the taskRejectWorkflow owner receives the rejection record; the draft stops before approval
Question cannot be resolved from approved materialEscalateWorkflow owner decides how to resolve the issue; no unsupported claim is added
Draft meets review criteriaAcceptApprover performs the separate final approval pass

4. Map and route the workflow

Part 1: Arrange the repairable route

Arrange the stages for the repairable-draft route. Begin after the input has passed its readiness gate and end at the workflow endpoint.

Sequence check

Put these stages in the correct order for a repairable draft.

  1. Approver records Approved or Not approved
  2. Reviewer checks the draft and records Edit
  3. Drafter creates the draft
  4. Reviewer rechecks the revision
  5. Drafter revises the draft

Route the decision

Match each situation to its next owner and action.

Tap an item and then its matching pair.

Step 3 of 8

Make the Workflow Reusable and Recoverable

Package a human-reviewed workflow into reusable specifications, prompts, versioned artifacts, run records, and bounded fallback rules.

A specification is the reusable backbone

Package the workflow

A reusable workflow is a small set of plain-text artifacts, not a one-off chat. The workflow specification tells each role what may enter, what must leave, who decides, and what happens when the normal route fails. It connects the input contract, reusable prompt, output bundle, review ownership, gates, and fallback rules. Its endpoint is an approved draft; publication remains outside the workflow.

The reusable workflow boundary

Diagram of approved inputs moving through a workflow specification, prompt, drafting, and human review toward an approved draft, with separate recovery paths for incomplete inputs, unavailable tools, and failed revisions.

The workflow specification links normal work and recovery paths without extending to publication.

workflow v1.0

Keep this specification as reusable plain text. It defines the operating rules; a run record later captures what actually happened in one run.

markdown
# workflow v1.0
Purpose: Turn an eligible approved brief into an approved draft for a named channel. The workflow ends at approval, not publication.

Input contract
- Workflow owner confirms an approved source version, audience, channel, and required output bundle.
- Return incomplete or unapproved inputs to the Workflow owner before drafting.

Expected output bundle
- One 80–120-word LinkedIn post in a warm, practical tone.
- Exact CTA: "Register for the free workshop."
- A concise factual-claim-to-source checklist and any unresolved questions.

Roles and gates
- Drafter creates or revises the bundle from the approved source and prompt.
- Reviewer records Accept, Edit, Reject, or Escalate.
- Approver records Approved or Not approved only after every required review check passes.

Fallback rules
- If an assistant is unavailable or limited, complete the bundle manually.
- Allow one focused AI revision for a repairable output.
- If it still cannot meet acceptance conditions, complete manually or Escalate; do not use unlimited retries.
- Preserve source, prompt, output revisions, review reasons, and approval records.

Separate fixed instructions from approved inputs

prompt v1.0

Double braces mark replaceable approved inputs. They are plain-text placeholders, not executable code. Keep the headings and fixed constraints unchanged when reusing the prompt.

markdown
Objective
Draft one {{channel}} post that invites the stated audience to Harbor Workshops' workshop.

Audience
{{audience}}

Source material
{{approved_source}}

Constraints
- Treat the source material as the only authority for factual claims.
- Do not invent benefits, conditions, testimonials, links, or replay availability.
- Write 80–120 words in a warm, practical tone.
- Include this CTA exactly: Register for the free workshop.
- Do not put [F#] references in the post.

Output criteria
Return exactly these three sections:
1. Post
2. Factual-claim-to-source checklist, using [F#] references
3. Unresolved questions, or "None" if there are no unresolved questions.

Version what changed

Keep versions separate because they answer different questions. source v1.0 records approved facts. prompt v1.0 records the reusable instruction. A draft revision records what the run produced. If approved facts change, create a new source version. If fixed instructions change, create a new prompt version. If only the produced draft changes, create the next output revision, such as A1-r2 after A1-r1. Do not overwrite an earlier artifact.

Tip

Provided example versus manual fallback

A provided example is a learning route: you can inspect a supplied response without using an external assistant. In real work, an unavailable assistant or free-tier limit triggers the manual fallback. The human still uses the same approved source, required output bundle, review gate, and approval rule.

Make decisions traceable and recoverable

Record the run, not just the draft

A run record is short but sufficient to reconstruct a decision. Start it with the source and prompt versions used for that run, then record the draft revision, review reason, and final approval separately. Use Not disclosed when a model identifier is unavailable. Use Not used only when a field genuinely does not apply, such as Tool/model in a manual run.

Run-record template

Use this field set for each actual run. Fill it with the values that occurred; the template is not a pre-approved result.

A concise run record preserves the decision trail without changing prior artifacts.
FieldValue to record
Run{{run_id}}
ModeManual, Live assistant, or Provided example
Source version{{source_version}}
Prompt version{{prompt_version}} or Not used
Tool/model{{tool_or_model}}; Not disclosed; or Not used
Draft revision{{draft_revision}}
Review dispositionAccept, Edit, Reject, or Escalate
Reason and changesEvidence, requested change, or decision
ApprovalApproved or Not approved
ApproverApprover or Not used
Timing basisObserved or Scenario estimate

Use bounded recovery

Use the fallback that matches the failure. A repairable output receives one focused AI revision attempt, not an open-ended cycle. If that attempt still cannot meet acceptance conditions, complete the bundle manually or record Escalate for a decision. In either case, keep approval as Not approved until all required checks pass.

Fallback rules at a glance

A fallback protects the human gate rather than lowering it.

Choose and record the response that fits the situation.
SituationRequired action and record
Input is incomplete or unapprovedReturn it to the Workflow owner. Do not begin a draft run.
Assistant or free tier is unavailableUse the manual fallback; record Mode as Manual and Tool/model as Not used.
First output is repairableRecord Edit, state the defect, and make one focused AI revision.
Focused revision still fails conditionsComplete manually or record Escalate; keep Approval as Not approved until conditions are met.
Output is unusable or raises an unresolved concernRecord Reject or Escalate; do not send it forward as an approved draft.

Check the artifact and recovery path

Apply versioning and recovery rules

Match each situation to the required version or fallback and run-record entry.

Tap an item and then its matching pair.

Step 4 of 8

Set the Comparison and Capture a Manual Baseline

Establish a fixed, reviewable comparison protocol and capture a traceable manual baseline before any AI-assisted result is considered.

Define a fair comparison

Fix the comparison before viewing results

Design the comparison before inspecting either result. Keep four elements equivalent:

  1. Input: both routes use the same approved source v1.0 for Harbor Workshops.
  2. Deliverables: both produce the same output bundle: one 80–120-word LinkedIn post in a warm, practical tone, a concise factual-claim-to-source checklist, and unresolved questions.
  3. Quality rule: both use the same four-row rubric.
  4. Completion: a bundle is complete only when every rubric row is Pass and the Approver has recorded Approved.

Compare approved endpoints, but do not erase the route taken to reach them. Keep initial defects, corrections, review dispositions, and any rejected attempts in the run record. A quick first draft in one route is not comparable with an approved draft in the other.

Equal inputs, equal endpoint

A diagram shows source v1.0 entering matched manual and AI-assisted lanes that share an output bundle, a review gate, and an approved-draft endpoint.

Compare equivalent approved work, not an early draft in one lane with a completed result in the other.

Tip

Baseline order matters

Create the manual baseline before opening or reading an AI response. Otherwise, the manual work can be influenced by wording, structure, or ideas already seen in that response. This order reduces imitation and practice effects; it does not eliminate every source of variation.

Apply the common rubric

Apply one shared acceptance rule

Use the established review rubric as a gate, not a score to average. A polished tone cannot compensate for unsupported facts, and a strong factual record cannot compensate for a safety or rights concern.

For each completed bundle, record Pass, Revise, or Block for every row. A Revise result identifies a repairable issue and sends the work back through correction and recheck. A Block prevents progress unless the issue is resolved through the workflow. Final approval is a separate decision: it can be Approved only after all four rows pass.

Approval requires every row to pass

Apply these rows to both the manual baseline and the AI-assisted run.

Do not average ratings: one Revise or Block means the bundle is not yet ready for approval.
Rubric rowReviewer checksApproval condition
Factual supportClaims trace to approved source materialPass
Brief and channel fitRequired bundle, LinkedIn format, length, tone, and CTAPass
Brand and readabilityWarm, practical, clear wordingPass
Safety and rightsNo unsuitable, unsupported, or rights-related concernPass

Tip

Keep the correction history

Retain the path to approval. If an initial draft needs a correction, preserve its Edit disposition and reason. Then record the corrected revision, its recheck, its Accept disposition, and the final approval decision separately. This history makes quality risk and correction effort visible in the later comparison.

Capture M1 and its timing basis

Example

Inspect the completed manual baseline

Completed supplied manual baseline — M1-r2

LinkedIn post

Looking for a practical way to plan content? Harbor Workshops is hosting Plan Your First Content Calendar, a free, 45-minute online workshop for solo business owners. The workshop covers topic selection and weekly scheduling. Bring the questions you have about choosing topics and setting a weekly schedule, then use the session to focus on those two planning tasks. Registration is required, and no replay is available. Set aside 45 minutes to focus on your content planning. Register for the free workshop.

Unresolved questions: None.

This is the completed manual bundle, not the initial draft. Its post is 81 words and contains no source-reference labels.

Trace M1-r2 factual claims

The finished post stays reader-facing. The checklist keeps the evidence trail outside the post.

References such as [F2] belong in the checklist and review record, not in the LinkedIn post.
Factual claimSource
Workshop title[F1]
Free, 45-minute, online[F2]
Solo business owners[F3]
Topic selection and weekly scheduling[F4]
Registration required; no replay[F5]
Required CTAOutput requirement

Example

Preserve the M1 correction and recheck

Initial manual record

  • Run: M1
  • Mode: Manual
  • Source version: source v1.0
  • Prompt version: Not used
  • Tool/model: Not used
  • Draft revision: M1-r1
  • Review disposition: Edit
  • Reason and changes: “Will help you plan a content calendar” claimed an outcome beyond source v1.0. Remove the unsupported outcome wording and recheck the complete bundle.
  • Approval: Not approved
  • Approver: Not used
  • Timing basis: Scenario estimate

Post-correction record

  • Run: M1
  • Mode: Manual
  • Source version: source v1.0
  • Prompt version: Not used
  • Tool/model: Not used
  • Draft revision: M1-r2
  • Review disposition: Accept
  • Reason and changes: Unsupported outcome wording removed. Rechecked the complete bundle; all four rubric rows Pass.
  • Approval: Approved
  • Approver: Approver
  • Timing basis: Scenario estimate

Separate active work from elapsed duration

Supplied M1 timing by stage. These are scenario estimates for the complete manual route, including its correction and final approval.

Tool waiting is zero in this manual run. The recurring total excludes One-time setup.
StageActive human workElapsed duration
Preparation2.0 person-min2.0 min
Drafting8.0 person-min8.0 min
Tool waiting0.0 person-min0.0 min
Verification3.0 person-min3.0 min
Correction1.0 person-min1.0 min
Approval2.0 person-min2.0 min
M1 recurring total16.0 person-min16.0 min

Choose and label the timing route

Record reusable work separately. For the supplied route, workflow v1.0 setup is 6.0 person-min and 6.0 min, both labeled Scenario estimate. It is not part of M1’s recurring total.

Choose one route for this learning activity:

  • Learner-timed route: create your own manual baseline before using an assistant; label personally measured stages Observed.
  • Supplied-example route: use M1 as shown and keep supplied timings labeled Scenario estimate.

The supplied-example route takes approximately 24 minutes. Optional original drafting or live-tool use may take longer.

Practice: complete the comparison record

Document a fair baseline

Write four labeled bullets to complete the comparison protocol and baseline record.

  1. Protocol: state the fixed input and common output bundle.
  2. Acceptance rule: state what must be true before a bundle counts as complete.
  3. Baseline status: identify M1’s timing basis, its final completion status, and the correction history that remains in the record.
  4. Fairness condition: name at least one condition that could make the comparison unfair, including how you would control it.

Use the supplied M1 record. Do not introduce AI-assisted results or calculate savings yet.

Write at least 240 characters (0/240).

Step 5 of 8

Run, Revise, and Approve the Draft

Run A1 and identify the defect

Start A1 with the approved artifacts

Choose one bounded route: transfer the approved material to an optional free-tier assistant, or inspect the complete supplied route on these screens. Neither route publishes anything.

For the supplied route, create the starting record before inspecting the response: Run: A1; Mode: Provided example; Source version: source v1.0; Prompt version: prompt v1.0; Tool/model: Not disclosed; Timing basis: Scenario estimate.

Keep the approved source and starting prompt distinct from later revisions. References such as [F5] belong in review notes and the factual-claim-to-source checklist, never in the reader-facing LinkedIn post. Reading a supplied example is learning time; it is not production timing.

One run, one focused revision

A workflow diagram showing an approved source and first prompt producing an initial draft, followed by a revised prompt, corrected draft, review, and separate approval.

The supplied chain preserves the source while the reusable instruction changes: source v1.0 plus prompt v1.0 produces A1-r1; prompt v1.1 and the same source produce A1-r2.

Example

Inspect the first supplied response

Supplied initial output — A1-r1

LinkedIn post

Looking for a practical way to plan content? Harbor Workshops is hosting Plan Your First Content Calendar, a free, 45-minute online workshop for solo business owners. The workshop covers topic selection and weekly scheduling. Bring the questions you have about choosing topics and setting a weekly schedule, then use the session to focus on those two planning tasks. Registration is required, and a replay will be available afterward. Set aside 45 minutes to focus on your content planning. Register for the free workshop.

Factual-claim-to-source checklist

  • Workshop title: [F1]
  • Free, 45-minute, online workshop: [F2]
  • Solo business owners: [F3]
  • Topic selection and weekly scheduling: [F4]
  • Registration required: [F5]
  • “A replay will be available afterward”: contradicts [F5]
  • Required CTA: Output requirement

Unresolved questions: None.

Correct, recheck, and approve A1-r2

Revise the reusable instruction once

A1-r1 is an 83-word post, includes the required CTA, and contains the complete output bundle. Its only defect is the replay statement, which directly contradicts [F5]. The Reviewer records Edit, not silent acceptance, and approval remains Not approved.

This failure calls for a specific control in the reusable instruction. Create prompt v1.1; preserve prompt v1.0 as the artifact used for A1-r1. Make one focused revision request: “Using source v1.0 and prompt v1.1, revise A1-r1 to correct the replay condition. Return the complete output bundle for recheck.” The resulting output revision is A1-r2.

prompt v1.1

The complete current reusable prompt adds an explicit no-replay control. It does not replace source v1.0.

markdown
Objective
Draft one {{channel}} post that invites the stated audience to Harbor Workshops' workshop.

Audience
{{audience}}

Source material
{{approved_source}}

Constraints
- Treat the source material as the only authority for factual claims.
- Do not invent benefits, conditions, testimonials, links, or replay availability.
- State that registration is required and that no replay is available.
- Do not state or imply that a replay will be available.
- Write 80–120 words in a warm, practical tone.
- Include this CTA exactly: Register for the free workshop.
- Do not put [F#] references in the post.

Output criteria
Return exactly these three sections:
1. Post
2. Factual-claim-to-source checklist, using [F#] references
3. Unresolved questions, or "None" if there are no unresolved questions.

Example

Return the complete corrected bundle

Supplied revised output — A1-r2

LinkedIn post

Looking for a practical way to plan content? Harbor Workshops is hosting Plan Your First Content Calendar, a free, 45-minute online workshop for solo business owners. The workshop covers topic selection and weekly scheduling. Bring the questions you have about choosing topics and setting a weekly schedule, then use the session to focus on those two planning tasks. Registration is required, and no replay is available. Set aside 45 minutes to focus on your content planning. Register for the free workshop.

Factual-claim-to-source checklist

  • Workshop title: [F1]
  • Free, 45-minute, online workshop: [F2]
  • Solo business owners: [F3]
  • Topic selection and weekly scheduling: [F4]
  • Registration required; no replay: [F5]
  • Required CTA: Output requirement

Unresolved questions: None.

Recheck evidence

Recheck the full bundle, not only the sentence that changed. The Reviewer applies the same shared rubric used for M1; the Approver acts only after the review result is Accept and every row passes.

A1-r2 is an 81-word LinkedIn post and can proceed to the separate approval pass only after every rubric row is Pass.
Rubric rowA1-r1A1-r2Review evidence
Factual supportRevisePassThe replay claim is corrected to match [F5]; remaining claims trace to [F1]–[F5].
Brief and channel fitPassPass81-word LinkedIn post, exact CTA, checklist, and unresolved questions.
Brand and readabilityPassPassWarm, practical, readable post without padded repetition.
Safety and rightsPassPassNo concern in the supplied bundle.

Preserve the trace and fallback path

Example

Keep review and approval as separate evidence

Initial review record

  • Run: A1
  • Mode: Provided example
  • Source version: source v1.0
  • Prompt version: prompt v1.0
  • Tool/model: Not disclosed
  • Draft revision: A1-r1
  • Review disposition: Edit
  • Reason and changes: “A replay will be available afterward” contradicts [F5]. Create a focused reusable prompt revision and recheck the complete bundle.
  • Approval: Not approved
  • Approver: Not used
  • Timing basis: Scenario estimate

Post-correction record

  • Run: A1
  • Mode: Provided example
  • Source version: source v1.0
  • Prompt version: prompt v1.1
  • Tool/model: Not disclosed
  • Draft revision: A1-r2
  • Review disposition: Accept
  • Reason and changes: prompt v1.1 explicitly requires no replay; A1-r2 states the approved condition. Complete bundle rechecked; all four rubric rows Pass.
  • Approval: Approved
  • Approver: Approver
  • Timing basis: Scenario estimate

Warning

Use the fallback without weakening the gate

If an optional assistant is unavailable, use the manual fallback rather than inventing an AI result: record Mode: Manual and Tool/model: Not used. If the one focused AI revision still fails the acceptance conditions, complete the bundle manually or record Escalate. In either case, keep approval as Not approved until a complete bundle passes every rubric row and the Approver makes a separate decision.

For supplied examples, keep all generation and production timing labeled Scenario estimate. Do not treat the time spent reading these screens as a measured run.

Practice: audit the A1 approval decision

Complete the A1 run record and endpoint verdict

Act as the record auditor. Reconstruct the two supplied A1 entries: the initial review of A1-r1 and the post-correction recheck of A1-r2. For each entry, include every run-record field: Run, Mode, Source version, Prompt version, Tool/model, Draft revision, Review disposition, Reason and changes, Approval, Approver, and Timing basis.

Then write a one- or two-sentence endpoint verdict. State whether the final artifact may leave the workflow as an approved draft and justify your answer from the revision path and human gates.

Write at least 500 characters (0/500).

Step 6 of 8

Calculate the Full Cost of the Run

Calculate recurring and first-use effort differences from complete run records, keeping active work, elapsed duration, setup, and timing evidence distinct.

Build a complete timing ledger

Count the whole run

Use the completed M1 manual baseline and A1 AI-assisted run as a stage-by-stage ledger. Both completed output bundles passed the common rubric, so their effort can be compared.

Record active human work in person-min and elapsed duration in min separately. Include Preparation, Drafting, Tool waiting, Verification, Correction, and Approval. Correction includes work caused by an initial defect or unsuccessful attempt.

Sequential stages can be summed. If durations overlap, do not count the same elapsed interval twice. Unattended tool waiting adds to elapsed duration but not to active human work. These supplied values are Scenario estimate, not Observed timings.

Canonical M1 and A1 timing by stage

Scenario-estimate timing ledger for the same approved output bundle.

Tool waiting is elapsed time only; it does not require active human work. The M1 stages match the completed manual-baseline record.
StageM1 person-minM1 minA1 person-minA1 min
Preparation2.02.02.02.0
Drafting8.08.02.02.0
Tool waiting0.00.00.01.0
Verification3.03.04.04.0
Correction1.01.02.02.0
Approval2.02.01.01.0
Recurring total16.016.011.012.0

Compare totals without hiding setup

Why draft speed is not total speed

A comparison of complete manual and AI-assisted timing bars, with a separate setup block before the AI-assisted route.

A faster drafting stage can coexist with waiting, verification, correction, and setup work. Compare complete approved runs, not drafting alone.

Calculate recurring and first-use differences

For a recurring comparison, exclude One-time setup and calculate each measure as manual baseline − AI-assisted run.

  • Active-work difference: 16.0 − 11.0 = 5.0 person-min
  • Elapsed-time difference: 16.0 − 12.0 = 4.0 min

For the first use, add the separate One-time setup of 6.0 person-min and 6.0 min to A1. Its first-use total is therefore 17.0 person-min and 18.0 min. Against M1 at 16.0 person-min and 16.0 min, manual − AI-assisted is −1.0 person-min and −2.0 min. In this supplied scenario, A1 takes more total effort on its first use.

Do not silently spread setup across future runs. First state a reuse assumption, then calculate any allocation separately.

Example

Worked calculation

Recurring calculation
M1 (16.0 person-min; 16.0 min) minus A1 (11.0 person-min; 12.0 min) gives 5.0 person-min and 4.0 min.

First-use calculation
A1 plus One-time setup is 11.0 + 6.0 = 17.0 person-min and 12.0 + 6.0 = 18.0 min. M1 minus this first-use total is negative: −1.0 person-min and −2.0 min. Setup remains visible because this comparison covers one use, not an assumed future series of runs.

Tip

Evidence rule

Keep Observed records separate from Scenario estimate records. A missing stage, mixed timing basis, or unknown overlap means the comparison cannot support a confident savings claim.

Apply the full-cost calculation

Audit the full-cost claim

Using the supplied Scenario estimate ledger, write a short calculation note.

  1. Calculate the recurring active-work difference and recurring elapsed-time difference using manual baseline − AI-assisted run.
  2. Add One-time setup to A1. State the first-use active-work and elapsed-time differences using the same subtraction order.
  3. A misleading comparison uses only Drafting. Identify one omitted stage that must be included through approval.
  4. Explain why the active-work difference and elapsed-time difference are not the same measure.

Write at least 180 characters (0/180).

Step 7 of 8

Decide What the Evidence Supports

Use quality, risk, complete effort, and evidence limits to make a proportionate workflow decision.

Gate quality before measuring savings

Use the right decision sequence

A workflow decision starts with quality and risk—not with a fast-looking draft.

  1. Confirm that every common rubric row is Pass and the draft received human approval.
  2. Ask whether the review gate, correction path, and fallback remain practical for this bounded use.
  3. Compare total recurring effort, then first-use effort including One-time setup.
  4. Check the strength of the evidence: Observed timings support a local claim; Scenario estimate supports only a scenario-based hypothesis.

Output volume or a quick first draft is not decision evidence by itself.

A visual sequence showing quality and risk checks before an effort comparison, followed by three possible workflow decisions.

First establish acceptable quality and manageable risk; only then evaluate effort and choose a path.

Turn comparison results into bounded evidence

What each decision means

Use the decision that fits the evidence, while keeping the approval gate intact.

A decision changes the workflow’s use, not the requirement for human approval.
DecisionUse whenNext action
AdoptQuality passes, risk is manageable, and the bounded use shows a useful net benefit.Define the permitted use and continue recording runs.
ReviseA specific failure, bottleneck, or uncertainty can be addressed without weakening safeguards.Change the relevant control and retest comparable briefs.
AbandonQuality, safeguards, or net benefit remain inadequate after a reasonable attempt.Stop this workflow for the task and use the manual path or escalate.

Read the supplied evidence carefully

For the supplied scenario, both M1 and A1 passed the common rubric. The recurring comparison is 5.0 person-min and 4.0 min in favor of A1 (manual − AI-assisted). But its evidence is Scenario estimate, not Observed.

First use tells a different story: M1 is 16.0 person-min and 16.0 min; A1 plus One-time setup is 17.0 person-min and 18.0 min. The first A1 draft also violated the no-replay condition before human correction and approval.

Treat this as evidence about a bounded scenario—not proof of general productivity. Approval effort, correction work, fallback frequency, free-tier availability, and unresolved issues can outweigh a faster drafting stage.

Tip

State limits and plan a proportionate next check

A small local comparison has important limits:

  • Different briefs may require different checking and correction effort.
  • Doing M1 before A1 can create practice or order effects.
  • Reviewer judgments can vary, even with a shared rubric.
  • A model response and free-tier limits can change between runs.

A targeted next check is stronger than a broad claim: test several comparable approved briefs, retain the same output bundle and rubric, label timings Observed, and record corrections, approvals, and any fallback use.

Practice a proportional decision

Choose the evidence-based response

In the supplied scenario, M1 and A1 both pass the common rubric. A1 has estimated recurring savings, but first use costs more after setup, and its first response required correction for a no-replay error. What is the most defensible next decision?

Make and justify your workflow decision

Write the decision note

Write a 90–150 word decision note for workflow v1.0 using the supplied scenario.

Choose Adopt, Revise, or Abandon. Cite the common quality result and both recurring and first-use effort evidence. State at least two limitations of the comparison, and name one targeted next check needed before broader use. Keep the timing basis clear and do not claim statistical certainty.

Write at least 450 characters (0/450).

Step 8 of 8

Finalize Your Workflow and Decision Note

Consolidate a reusable workflow package, verify its traceability, and write a bounded decision note grounded in the supplied comparison.

1. Assemble the reusable package

Compact workflow specification

A reusable package is a short operational reference, not a collection of disconnected notes. Consolidate the artifacts already built into workflow v1.0:

  • Entry: The Workflow owner confirms an eligible approved source, audience, channel, and required output bundle.
  • Method: The Drafter uses the current reusable prompt v1.1 with approved inputs. Preserve prompt v1.0 because it documents the earlier A1 attempt.
  • Output: One 80–120-word LinkedIn post in a warm, practical tone; the exact CTA “Register for the free workshop.”; a factual-claim-to-source checklist; and unresolved questions.
  • Human gates: The Reviewer applies all four rubric rows. The Approver records Approved only after every row is Pass.
  • Fallback: Return incomplete inputs; allow one focused AI revision for a repairable output; then complete manually or Escalate if conditions still fail.

Keep source, prompt, and output versions distinct. The endpoint is an approved draft. Publication remains outside the workflow.

Workflow boundary and human gates

A diagram of approved inputs moving through drafting, review, revision, and human approval before an approved draft leaves a bounded workflow; publication is outside the boundary.

The reusable package preserves a human approval gate and ends with an approved draft, not publication.

Summary

Package checklist

Use this check before reusing the workflow.

  • workflow v1.0 states eligible inputs, the output bundle, roles, gates, and fallback rules.
  • prompt v1.1 is the current reusable instruction; prompt v1.0 remains part of the earlier A1 record.
  • A run record preserves Run, Mode, Source version, Prompt version, Tool/model, Draft revision, Review disposition, Reason and changes, Approval, Approver, and Timing basis.
  • Traceability must connect the approved source to each output revision and to a separately recorded final approval.
  • The workflow creates an approved draft only. It does not automatically publish content.

2. Verify traceability and evidence

Trace the corrected A1 history

Traceability is a chain of evidence, not merely a final file. The canonical supplied history for the AI-assisted run is:

source v1.0 + prompt v1.0 → A1-r1 → Edit → prompt v1.1 + source v1.0 → A1-r2 → Accept → Approved

A1 is the run identifier. A1-r1 is the first draft revision. The Reviewer records Edit because its replay statement contradicts [F5]. The reusable instruction is then changed to prompt v1.1; that changed prompt and source v1.0 produce A1-r2.

The Reviewer rechecks the complete A1-r2 bundle. Factual support, Brief and channel fit, Brand and readability, and Safety and rights must all be Pass. Only then does the Approver separately record Approved. Keep the defect, change reason, and earlier Not approved status in the record; do not overwrite history.

A1 run-record evidence

The two records below preserve the initial defect and final human decision.

A compact record of the supplied AI-assisted run.
RecordSource and promptDraft and dispositionReason and changesApproval and approverTiming basis
A1; Provided example; Tool/model: Not disclosedsource v1.0; prompt v1.0A1-r1; EditReplay statement contradicts [F5]; revise the reusable instruction.Not approved; Not usedScenario estimate
A1; Provided example; Tool/model: Not disclosedsource v1.0; prompt v1.1A1-r2; AcceptNo-replay requirement addressed; complete bundle rechecked; all rubric rows Pass.Approved; ApproverScenario estimate

Preserve the fair comparison

A comparison is fair only because the completed M1 manual baseline and A1 AI-assisted run pass the same rubric and deliver the same bundle. In the supplied Scenario estimate, recurring effort is M1 minus A1: 5.0 person-min and 4.0 min in favor of A1.

For first use, retain One-time setup: A1 is 17.0 person-min and 18.0 min after adding setup, versus M1 at 16.0 person-min and 16.0 min. Therefore, manual minus AI-assisted is −1.0 person-min and −2.0 min for first use. This is local estimated evidence, not an observed productivity claim or a guarantee for another brief, reviewer, model response, or free-tier condition.

Tip

Keep the gates after revision

The corrected prompt is not a reason to weaken review. prompt v1.1 is a control to retest with comparable approved briefs, while the Reviewer and Approver still apply the same gates to every new draft.

3. Finalize and defend your decision

Final application: defend the workflow

Write a concise final decision note with these five labeled parts. Use only the supplied records.

  1. Workflow package: identify workflow v1.0, the current prompt v1.1, the required output bundle, roles, gates, and one fallback rule.
  2. Traceability: show the canonical A1 chain from source v1.0 and prompt v1.0 through A1-r1, the [F5] review issue, prompt v1.1, A1-r2, recheck, and approval.
  3. Comparison: state the shared quality result, recurring difference, first-use difference with setup, and timing basis.
  4. Recommendation: choose Adopt, Revise, or Abandon for a bounded future use. State a condition and at least one limitation of this local comparison. Acknowledge that prompt v1.1 already addresses the replay failure; do not propose that same change again.
  5. Disruption: In a hypothetical rerun, source v1.0 is superseded by approved source v1.1 immediately before A1-r2 would receive approval. Apply the existing readiness, review, and fallback rules. Explain how you preserve records and prevent approval from being bypassed.

Write at least 700 characters (0/700).

Tutorial complete

Congratulations! You completed: Build and Evaluate a Human-Reviewed AI Workflow

Approval remains human. Your comparison supports only the tested task and its stated timing basis; reuse the workflow with its records, gates, and fallback rules intact.

Download the App now to have access to + 5000 free courses, exercises, certificates and lots of content without paying anything!

  • 100% free online courses from start to finish

    Thousands of online courses in video, ebooks and audiobooks.

  • More than 60 thousand free exercises

    To test your knowledge during online courses

  • Valid free Digital Certificate with QR Code

    Generated directly from your cell phone's photo gallery and sent to your email

Cursa app on the ebook screen, the video course screen and the course exercises screen, plus the course completion certificate