How Do You Get Honest Trade-offs From Claude?
Name your criteria in order of importance, then ask for a table with a concrete judgment in every cell and a label saying where each judgment came from. You can then see which cells rest on your code, which on a current source, and which on an assumption.
We're choosing between [option A], [option B] and [option C] for [purpose]. Read [paths] so your judgments fit this code. My criteria, most important first: [1], [2], [3]. Hard limits: [e.g. no new paid services]. Build a table with options as rows and my criteria as columns. Each cell needs a concrete judgment with a number or a named consequence, never "good" or "scalable". End each cell with REPO if it comes from this code, DOCS plus a link if you checked a current source, or ASSUMED if neither. Search the web for versions, limits and prices, since they may have changed since your training. Then recommend one, name the criterion that decided it, and name the one fact about our scale or team that would flip the recommendation.
- REPOjudged from this code
- DOCSchecked a current source, with a link
- ASSUMEDneither, so verify before you commit
Try it now
- 1Run
/effort highin Claude Code, since this is a decision worth more thinking, then send the prompt. - 2Open every link marked DOCS, and treat ASSUMED cells as questions to settle.
- 3Run
/branch, swap your top two criteria, and ask for the table again./resumereturns to the original conversation. - 4It worked if you can point to the cell that decided it, and to what would have to change for the other option to win.
Use it when
Choices that reasonable engineers would argue about and that you'll have to defend: build or buy, which library, which database, a schema you'll live with. It works in the Claude app and Claude Code. Use Claude Code when the answer depends on what's in your repo.
Skip it when
When one option wins on every criterion, or a hard limit rules out the rest. A table adds nothing without real tension. If you haven't got options yet, generate them first, then compare them here.
Why it works
A bare recommendation hides its weighting, and Anthropic's docs point out that Claude can't read your preference between two reasonable approaches from the code. So you supply the ranking. The source labels handle a second problem. Anthropic lists a reliable knowledge cutoff of June 2026 for its current models, so anything about a library's version, limits or price may have changed since. Its Sonnet 5.5 guide says the model sometimes answers from training when a search would catch details that have changed, and suggests telling it to check. Claude Code has WebSearch and WebFetch tools for that. A column of labels shows you which parts to verify before you commit.
Example
Say your Node app runs on Postgres with no Redis, and you need background jobs. You compare BullMQ, pg-boss and a hosted queue on operational simplicity, then throughput at 200 jobs a second, then observability. A useful cell for BullMQ reads "needs a Redis instance we don't run today, REPO", not "more infrastructure". The flip fact might be that you already plan to add Redis for caching.
Going further
Get a second read that isn't anchored on the first. Anthropic's docs call a second session the easiest independent reviewer, because it never watched the reasoning happen. Paste only the table and your criteria into a new session and ask what it would challenge. If effort isn't enough, the docs suggest Fable 5.1 for demanding reasoning, selected with /model fable. Check the bill first: Anthropic's model docs say Fable can bill to usage credits instead of your plan's included limits, depending on your plan. You don't need to request the full chain of reasoning, either. Its Opus 5.5 guide says pushing the model to reproduce internal reasoning in the response might get declined, while a short explanation is fine.
Common mistakes
- Letting Claude choose the criteria. The table then optimizes for something you never asked for.
- Accepting cells like "good" or "scalable". If two cells can't be compared, the table hasn't told you anything.
- Trusting library facts from memory. Anything about versions or prices that isn't marked DOCS is unverified.
- Treating the table as objective. Your criteria and their order drive the result, and re-running with them swapped shows how much.
Sources
- Best practices for Claude Code
- Models overview, Claude Platform Docs
- Prompting Claude Sonnet 5.5, Claude Platform Docs
- Prompting Claude Opus 5.5, Claude Platform Docs
- Tools reference, Claude Code Docs
- Commands, Claude Code Docs
- Model configuration, Claude Code Docs
Checked 6 November 2026. Products change often, so confirm details with the vendor.