Human-plus-Model Feedback Format
Pair your judgment with model critique and disclose where they differ.
- Difficulty
- Easy
- Time to result
- ~days to results
- Steps
- 5
- Confidence
- 99%
The Human-plus-Model Feedback Format creates two explicitly attributed layers of review. The reviewer first examines the work independently and records personal feedback before consulting AI, preventing the model from anchoring the human assessment. Next, the reviewer chooses a model appropriate to the problem and requests a separate critique. The two reviews are compared rather than silently blended. The final response contains the reviewer's own judgment, the model's additional observations, and a clear statement of where the reviewer agrees or disagrees with the model. This format increases coverage and gives the recipient useful second-opinion evidence while keeping responsibility with the human. It also makes the AI contribution visible, preventing model-generated criticism from being presented as wholly personal expertise.
Origin
Extracted from Marketing Against The Grain through Kieran Flanagan's repeatable format for responding to feedback requests.
Core principles
- 01Human judgment and model analysis provide different signals.
- 02Feedback recipients should know which conclusions belong to the human and which came from AI.
- 03Agreement and disagreement are more informative than an undifferentiated blended answer.
- 04The human remains accountable for the feedback delivered.
How to run it
- 1
Review independently
Read the material and draft your own feedback before involving a model.
Pro tip Capture your initial priorities and rationale in writing.
Watch out Consulting AI first can anchor your judgment.
- 2
Select an appropriate model
Choose a model whose strengths match the problem, such as strategy, writing, coding, or research.
Pro tip Use prior task-based evaluations to guide the choice.
Watch out Do not assume one model is best for every kind of review.
- 3
Generate a separate critique
Give the model the material, objective, audience, and constraints, then request its assessment.
Pro tip Ask it to prioritize findings by impact.
Watch out Do not include private material unless the model and deployment are approved for it.
- 4
Compare judgments
Identify overlap, unique observations, contradictions, and questionable model claims.
Pro tip Verify factual model criticisms against the source material.
Watch out Similarity between two opinions does not automatically make them correct.
- 5
Deliver attributed feedback
Present your feedback, the model's feedback, and concise notes explaining agreement or disagreement.
Pro tip Lead with the points most useful to the recipient rather than the novelty of using AI.
Watch out Never conceal model involvement when its conclusions materially shape the review.
In the wild
A manager independently reviews a proposed launch strategy, then asks a strategy-capable reasoning model to analyze the same document. The manager sends the author personal feedback, adds the model's distinct concerns, and explains why one model recommendation is rejected due to an operational constraint.
→ The author receives broader feedback without losing clarity about who endorses each point.
Kieran said that whenever someone asks for feedback, he provides his own assessment, adds feedback from the model he considers best suited to the problem, and explains where he agrees or disagrees with the model.
→ Feedback becomes a transparent human judgment supplemented by a visible AI second opinion.
Common mistakes
Letting AI write the first opinion
Starting with model output can suppress independent human observations and turn the final feedback into lightly edited AI text.
Blending sources invisibly
Recipients cannot evaluate accountability or disagreement if human and model opinions are presented as one voice.
Forwarding unverified criticism
Model feedback may contain incorrect assumptions and should be checked before it is passed to another person.
Is it for you?
Best for
Managers, collaborators, editors, and specialists who routinely review other people's work.
Not ideal for
Sensitive feedback that cannot be shared with an external model or situations requiring confidential human-only judgment.
From the transcript
“If anybody asks me for feedback on something, I have a format now.”
“I give them my feedback and I tell them I also give them the feedback from the model I think is most appropriate for that…”
“And I tell them what I agree with or disagree with that the model says”
From the episode
GROK 3 vs GPT-4: The AI War Just Got Real [First Look]