The Screening Log

24 candidate Power Modes, screened against the catalogue. 10 turned out to be modes we already had.

Every candidate that reaches this project gets screened before it gets tested. Most do not survive the screen, and the reason is almost never that the idea was bad.

What "screened" means here

A candidate arrives as a claim about what an instruction makes the AI do. Before anything is tested, it is scored on five dimensions against a written rubric, and two of those act as hard gates: how original the mechanism is, and how distinct it is from the modes already in the catalogue.

Those two are gates rather than weighted inputs for one reason. A behaviour can be genuinely useful and still not be a new mode, because the useful thing is already available under an existing name. Rating it highly for usefulness would let it in anyway, which is the exact failure the screen exists to prevent.

Screening is a desk check. It can judge whether a claimed mechanism would be distinct. It cannot judge whether the mechanism actually operates — only a test settles that, and this project has run few enough tests that the difference matters.

a desk check can evaluate whether a claimed mechanism would be distinct, but not whether the mechanism actually operates

Recorded in the project’s own test notes, after a screen scored a candidate higher than its test did.

Where they landed

10 Already a mode
10 Not a mode
2 Tested and kept
1 Open — needs narrowing
1 Out of scope

24 screened. 3 of them went on to a test; the rest were settled at the screen.

Every candidate

The whole record, so the claim above is checkable rather than asserted.

Candidate	Verdict	What we found
LEARN FASTER WITH 80/20	Already a mode	Find the few ideas that carry a subject, then explain those first. Both halves already exist: one mode picks what matters most, another explains it plainly. Useful together, not a third thing.
SIMPLIFY COMPLEX TOPICS	Already a mode	Explaining from zero prior knowledge is what plain-language explanation already does. Starting lower on the ladder is a setting you can ask for, not a different behaviour.
TEST MY KNOWLEDGE	Already a mode	Questions that find the gaps rather than explanations that fill them. That is adaptive quizzing, which the wider taxonomy already covers under its own name.
MEMORY ENCODING	Open — needs narrowing	Turning ideas into memory hooks. Tested, and the promised trade-off never appeared — it produced more, not less. The behaviour is real; the explanation of it needs rewriting before it can be tested again.
IMPROVE MY WRITING	Not a mode	Cleaning up prose without losing the author''s voice. Three existing modes already cover this between them, split by the actual problem: too long, too robotic, or wrong voice.
BEAT WRITER'S BLOCK	Already a mode	Several ways into a stuck topic. That is idea generation, which the wider taxonomy already covers under its own name.
CREATE ANY TYPE OF CONTENT	Not a mode	One template for any content request. A mode has to be able to say when not to use it; a template that fits everything constrains nothing, so there is nothing for it to change.
WRITE IN MY STYLE	Already a mode	Writing in someone's voice on a topic they have not covered. Already a mode — supply the samples and it matches cadence, formality and vocabulary. The two-stage profiling step is worth folding in.
ACRONYM SLOT SCAFFOLD	Not a mode	Naming the fields a good prompt needs so none get missed. That is a form, and the Builder already asks for those fields directly. Useful structure, not a change in how the AI reasons.
DOMAIN CHECKLIST MODES	Not a mode	Prebuilt field lists for particular jobs. Genuinely handy as reference material, but a checklist tells you what to supply rather than changing what the AI does with it.
SLASH COMMAND NAMESPACE	Not a mode	A short token standing in for a long prompt. That is a shortcut, and shortcuts already exist — two modes answer to one. It saves typing without changing the behaviour underneath.
IDENTITY-PRESERVING EDIT CONSTRAINTS	Out of scope	Change one attribute of a picture and guarantee the rest survives. A real problem, and out of scope rather than weak — Power Modes work on text. Parked in case the manual takes on images.
CRITIQUE BEFORE REBUILD	Not a mode	Attack the idea before improving it. The attacking half already has a mode; what is distinctive is insisting on the order. Narrow it to that claim and it can be tested.
BUYER OBJECTION LENS	Tested and kept	Occupy the person who has to say yes, and have them refuse. Tested against both a plain review and a general critique, and it produced reasons neither did. In the product.
BACKCAST	Tested and kept	Fix the finished state and work backward to what must already be true at each step. Tested against forward planning and its nearest relative; it produced conditions, not activities. In the product.
10X EDITOR	Not a mode	Rewriting so text reads as human rather than generated. That is exactly what an existing mode does, down to the instruction not to invent personality.
INVISIBLE TEAM	Not a mode	Turning a one-off task into a standing set of specialists. That is workflow design — real and valuable, and a different thing from a mode, which changes one response.
SECOND BRAIN	Already a mode	That's Socratic questioning, which already has a mode. Hold the advice, ask one question at a time, and let the reader get there first.
EMPIRE PROMPT	Not a mode	One word producing a full year's plan. The underlying idea — work backward from the goal — did survive, as its own mode. What did not survive is the promise that a single word carries it.
ABSOLUTE MODE	Already a mode	Strip out filler, flattery and follow-up questions. Closest to an existing mode, and arguably the precise definition that mode was flagged as needing. Worth reworking as a refinement rather than a rival.
CHALLENGE ME	Already a mode	Ask for the argument against instead of agreement. That is broad adversarial review, which already has a mode — and the phrasing now routes to it.
EVIDENCE-TIERED RESEARCH	Already a mode	Sort research into what is verified, what is supported and what is still open. Valuable, but so far it describes the shape of the answer rather than how the material gets weighed. Narrow it to that.
EDIT WITH CHANGE LEDGER	Not a mode	Edit, then list every change and why. Nothing else reports on its own edits, which is genuinely new — but it changes what comes back rather than how the editing is done. Narrow it to the guarantee.
INTERROGATE BEFORE BUILDING	Already a mode	Have the AI ask what it needs before it starts. That already has a mode, which asks only the questions that would change the result. Batching them up front is a choice of format, not a new way of working.

10 modes we already had

The interesting group. Every one of these describes a real behaviour — and every one of them already had a name. Where the covering mode is built, its instruction is shown as well — taken from the Builder’s own data rather than described, so you can judge the overlap yourself.

LEARN FASTER WITH 80/20

Surface the highest-leverage ideas in a subject, then teach those first.

Two behaviours, and the catalogue has both. One mode finds the few items that carry most of the outcome; another explains things in plain language for a reader who lacks the background. Run in sequence they do exactly what this asked for. The proposal was a good instinct about how to learn a subject — it just describes a pairing rather than a mode.

Already covered by IMPACT PRIORITIZATION

What IMPACT PRIORITIZATION actually tells the AI to do

Identify the small number of risks, causes, actions, or factors most likely to drive most of the practical impact. Explain why they deserve attention first, but do not omit mandatory requirements or lower-ranked items that could combine into material risk.

SIMPLIFY COMPLEX TOPICS

Explain a topic from zero prior knowledge upward.

That is plain-language explanation, which already has a mode. Where you start on the ladder is something you tell it — "assume I know nothing about statistics" — rather than a different behaviour. The distinction is worth holding onto: a parameter you can vary inside a mode is not evidence of a new one.

Already covered by ELI10

What ELI10 actually tells the AI to do

Explain this in simple language for someone without technical background. Keep the important concept accurate. Replace jargon with plain language, but do not remove the facts, uncertainty, caveats, or risks needed to understand the issue.

TEST MY KNOWLEDGE

Find the gaps in what you know by being questioned, rather than by being explained to.

Adaptive quizzing, which the wider taxonomy already covers. This one came closest of the seven to being a second mode, because guiding someone with questions is also a mode — and the difference is what the questions are for. Quizzing tests what you already know. Guiding walks you toward an answer you have not reached yet. Both are real, both are covered, and this is the first one.

Already covered by QUIZ ME (covered in the taxonomy, no mode page)

QUIZ ME is covered in the wider taxonomy but has no built mode on this site yet, so there is no instruction to show. You can see where it sits in the taxonomy: QUIZ ME in the taxonomy

BEAT WRITER'S BLOCK

Generate several entry points into a topic you are stuck on.

Idea generation under a different name. The wider taxonomy already covers it, and the same merge has happened before — an earlier candidate proposing open-ended ideation went the same way. Being stuck is a real situation; needing a mode named after the feeling is not.

Already covered by BRAINSTORM (covered in the taxonomy, no mode page)

BRAINSTORM is covered in the wider taxonomy but has no built mode on this site yet, so there is no instruction to show. You can see where it sits in the taxonomy: BRAINSTORM in the taxonomy

WRITE IN MY STYLE

Write in an author’s voice on a topic they have never written about.

Already a mode. Supply samples and it matches cadence, formality, sentence length and vocabulary while holding the meaning fixed. This proposal added a two-stage step — profile the voice first, then write — which is a genuinely good idea and worth folding into the existing mode’s instruction rather than standing up a second one beside it.

Already covered by VOICE MATCH

What VOICE MATCH actually tells the AI to do

Rewrite this so it matches the voice, cadence, level of formality, sentence length, and vocabulary clearly supported by the supplied examples. Preserve the meaning. Match only stable patterns; do not exaggerate quirks or invent personality traits. If the samples are insufficient, say so.

SECOND BRAIN

Coach the reader to their own answer instead of supplying one.

That is Socratic questioning, which already has a mode: hold the advice, ask one question at a time, let the reader get there first. What makes this the best example on the page is that the person proposing it already suspected it duplicated something — they proposed merging it with another candidate. They were right twice over. Both were duplicates. They just were not duplicates of each other: this one is Socratic questioning, and the other turned out to be adaptive quizzing.

Already covered by SOCRATIC MODE

What SOCRATIC MODE actually tells the AI to do

Do not answer yet. Ask me the few most important questions needed to clarify the decision. Ask one at a time, use my answers to decide what to ask next, ask only questions that could materially change the recommendation, and stop when you have enough information to give useful guidance.

ABSOLUTE MODE

Strip the conversational packaging out of every answer.

Half of this was already the job of the mode that takes the machine out of writing — the filler, the hype, the motivational line that states nothing — and that half is now written into it as a checkable list, which is what the taxonomy had asked for and never had. The other half is about the shape of a turn rather than the shape of prose: no mirroring your mood, no offer of further help, no closing question invented to keep the exchange going. Useful, and not a mode. Ask when it should fire and the honest answer is always, which makes it a setting you turn on rather than one you choose between.

Already covered by NATURAL VOICE

What NATURAL VOICE actually tells the AI to do

Rewrite this so it sounds natural, direct, and human rather than polished by a template. Preserve the meaning, facts, and important caveats. Remove robotic phrasing, unnecessary transitions, repetitive setup, and generic AI cadence. Do not invent personality.

CHALLENGE ME

Ask for the argument against something instead of agreement.

Broad adversarial review, which already has a mode — and the proposal named the same three things that mode names: counterarguments, hidden assumptions, failure points. The one part that was not already covered, asking for fixes alongside the criticism, is something you request of the result rather than a different way of thinking. The phrasing now routes to the existing mode.

Already covered by REDTEAM

What REDTEAM actually tells the AI to do

Actively search this work for material weaknesses, hidden assumptions, failure points, overlooked dependencies, evidence gaps, and credible counterarguments. Prioritize only issues that could materially change the decision, outcome, risk, or feasibility.

EVIDENCE-TIERED RESEARCH

Sort research into what is verified, what is supported, and what is still open.

Something real does survive the narrowing: deciding how well each claim is evidenced happens before any of it is laid out, so this is not only a format request. But that judgement is source checking, which the wider taxonomy already names — identify the claims needing external verification, and weigh the quality and agreement of what is behind them. The four buckets are one way of presenting it. Admitting the behaviour under a new name would let it in through a side door the taxonomy has already ruled on.

Already covered by ASSUMPTION CHECK

What ASSUMPTION CHECK actually tells the AI to do

Identify the assumptions that materially support this conclusion. Separate stronger assumptions from weak or unverified ones, and rank them by how much changing them could change the result. Ignore assumptions unlikely to change the recommendation, plan, estimate, or decision.

INTERROGATE BEFORE BUILDING

Make the AI ask what it needs to know before it produces anything.

Good instinct, and already a mode: hold the answer, ask only the questions that would change the result, stop once there is enough to go on. The proposal asked for all the questions in one batch and a finished draft at the end, where the existing mode asks one at a time and ends in advice. Both are things you can ask of it — "put all your questions up front", "then write it" — rather than a different way of working. Asking for questions this way already routes to the existing mode.

Already covered by ASK QUESTIONS FIRST (covered in the taxonomy, no mode page)

ASK QUESTIONS FIRST is covered in the wider taxonomy but has no built mode on this site yet, so there is no instruction to show. You can see where it sits in the taxonomy: ASK QUESTIONS FIRST in the taxonomy

The one that was tested, and failed

One candidate in this batch got further than a screen. It claimed a trade: convert ideas into memory hooks so they stick, at some cost to completeness. That is a falsifiable claim, and it was worth the cost of testing because the mechanism was specific enough to be wrong.

It was wrong.

The claimed mechanism did not operate -- no accuracy trade-off occurred in any task; arm B was more complete, not less.

Arm B is the run that used the candidate. The promised trade never appeared: the output was not less complete, it was more.

What this test does and does not establish

Three conditions belong next to that finding, because they limit how far it generalises.

Model not recorded. The capture does not record which model produced the outputs. The notes presume one and mark it unconfirmed.
One run per arm. Twelve arms across three tasks, each run once. A single run cannot separate a real effect from variation between runs.
A confound in one task. Interactive chart widgets rendered in three of the four arms of the first task but not the fourth, inflating one comparison. The deciding pair both had them, so the finding itself survives.

None of that undoes the result. The trade-off did not appear in any of the three tasks, and that is what the candidate staked its case on. What the conditions limit is how confidently it can be generalised beyond this run — which is why the candidate was returned for a rewrite rather than rejected.

It is not a rejection. The behaviour it actually produces is real — scaffolding that makes material easier to retrieve — and it is close enough to an existing mode that the difference needs restating before it can be tested again. It sits open.

Still open

Not every candidate that fails to enter the catalogue is a rejection. Some describe a real behaviour in terms too broad to test, and the honest answer is to narrow the claim and look again rather than to decide either way.

These are open. Listing them is the point — a log that showed only the ones that were settled would be a different and less useful document.

MEMORY ENCODING Turning ideas into memory hooks. Tested, and the promised trade-off never appeared — it produced more, not less. The behaviour is real; the explanation of it needs rewriting before it can be tested again.

How to tell a new technique from an old one renamed

These are the questions the screen actually asks. They work on anything you come across, and most of the work is in the second one.

State the mechanism without using its name.If you cannot describe what the AI must do differently without saying the label, there is nothing underneath the label yet.
Name the closest thing you already use. Then say what disappears if you drop this and use that instead.Answers that do not count: it is worded better, it is more thorough, it sounds more professional. An answer counts when it names a behaviour the closest alternative does not produce.
Say when you would not use it.A technique that applies to everything constrains nothing. If you cannot name a situation where it is the wrong choice, it is not doing specific work.
Ask whether it changes the thinking or the output.Wanting the answer in a particular shape is a formatting request. Wanting the AI to reason differently is a mode. Both are useful; only one is a new technique.

The modes that did make it

Fifteen of them, each with what it does and when it is the wrong answer.

Read the Power Modes →