Iterative Reasoning Stack-Rank
Force successive reasoning passes to improve and prioritize an AI answer
- Difficulty
- Easy
- Time to result
- ~days to results
- Steps
- 6
- Confidence
- 95%
Use successive reasoning passes instead of accepting a model's first completion. Begin with the goal and relevant evidence, then request an initial analysis. Ask the model to think more deeply—or use a continuation cue such as “wait”—so it re-examines its earlier work. Require it to compare, stack-rank, and improve candidate answers rather than merely lengthening the response. Each pass should expose stronger hypotheses, trade-offs, experiments, or previously overlooked ideas. The mechanism works best when paired with rich internal context, because additional reasoning cannot compensate for absent data. Stop when later passes cease to change priorities materially, then validate the selected recommendations before implementation.
Origin
Extracted from Marketing Against The Grain during Kieran Flanagan's demonstration of using a reasoning model to produce and repeatedly improve a growth plan.
Core principles
- 01Treat the first answer as a draft
- 02Request deeper reasoning in successive passes
- 03Have the model compare and rank its own candidates
- 04Add context before demanding more sophistication
- 05Stop when added passes no longer improve decision quality
How to run it
- 1
Ground the first pass
Provide a concrete objective, relevant data, constraints, and the required decision format.
Pro tip Ask the model to identify the few variables that matter before requesting a full plan.
Watch out A poorly grounded first pass gives later iterations weak material to refine.
- 2
Generate an initial answer
Request hypotheses, recommendations, trade-offs, or experiments appropriate to the decision.
Pro tip Preserve the first answer so changes across passes are visible.
Watch out Do not mistake polished language for strong reasoning.
- 3
Trigger deeper review
Tell the model to think more deeply or continue reasoning after it appears finished.
Pro tip A short continuation cue can be enough to induce another reasoning pass.
Watch out Repeated prompts may add verbosity without adding insight.
- 4
Stack-rank candidates
Ask the model to inspect prior suggestions, compare them, remove weak options, and prioritize the strongest.
Pro tip Specify ranking criteria such as impact, evidence, cost, risk, and reversibility.
Watch out Rankings inherit any errors in the supplied evidence.
- 5
Test for diminishing returns
Repeat the cycle only while the model produces materially better hypotheses or changes the priority order.
Pro tip Stop when recommendations stabilize across passes.
Watch out Endless iteration delays real-world learning.
- 6
Validate and act
Check assumptions and evidence, then convert the top candidates into experiments or decisions.
Pro tip Record whether the model's ranking predicted actual results.
Watch out Reasoning output remains a proposal, not proof.
In the wild
A growth leader provides a SaaS dashboard and asks which metrics matter most for doubling ARR. The model identifies churn, revenue per user, activation, conversion, and referrals, then drafts experiments. The leader repeatedly asks it to think more deeply and rank the options by expected impact and trade-offs. A later pass surfaces a usage-drop trigger for early outreach that was absent from the initial answer.
→ Successive passes produce a more differentiated and prioritized growth plan than the first completion.
Common mistakes
Repeating without new criteria
Simply requesting another version may create longer wording rather than better prioritization.
Using iteration instead of context
Additional reasoning cannot recover company data, constraints, or experiment history that were never supplied.
Never leaving the model
Continual refinement becomes counterproductive when a real experiment would generate stronger evidence.
Is it for you?
Best for
Users employing reasoning models for growth plans, prioritization, hypotheses, trade-offs, and other complex analytical work.
Not ideal for
Simple factual retrieval or tasks where repeated prompting would amplify missing context rather than improve reasoning.
From the transcript
“If you ask it to think more deeply about these problems and do another version, you'll get better results, better results, better results each time.”
“And one of the key things was just every time it thought it was finished, giving you the answer, they would just give it the…”
“And what it's doing in the background is it's looking through all of the things it's given you and stack ranking them and trying to…”
From the episode
How a $1B+ Crypto Company Really Uses AI in Marketing