Advanced Prompting in 2026

Model capabilities have drastically improved since I wrote the initial set of prompts in my free prompt collections. The 2024-era prompts told the model to use structured lists, provide three levels of detail, include definitions, follow a template. Modern models do all of that unprompted so prompts can focus on getting it to commit, surprise, and self-audit.

Prompting Techniques

The prompt budget is now spent pushing the model outside its comfort zone with discard conditions, bias countermoves, novelty discipline, and declared failure modes. If the model would do it by default then let it unless that default is a trough in latent space we need to push past.

It might be interesting to dig into these concepts with some examples from the collections.

Discard Conditions

Mandatory tests with discard conditions, embedded in the step where they bite.

From Spectrum Exploration:

The emergence test: could this construct be fully derived from either endpoint alone? If yes, discard it and dig again in the same zone.

The prior-art check (mandatory, separate from the emergence test): does this construct already exist in a live literature under another name? Interaction-dependence does not establish novelty — these are different claims.

Bias Countermoves

The format is: failure mode → why it happens (artifact of model preference, not the domain) → countermove → what a passing run should contain.

From Critical Comparison:

Known bias (hypothesized): comparisons default to running every plausible criterion at equal weight, because a full matrix looks thorough. The matrix is where verdicts go to die — coverage substitutes for judgment.

From Cybernetic Exploration, the leverage step:

Known bias (hypothesized): analyses default to parameter-level interventions — adjust the budget, add headcount, tune the threshold — because parameters are visible and polite. These are the lowest-leverage points in a system.

Novelty Discipline

The best example is the anti-pattern “The rename” in Spectrum Exploration’s calibration section. It’s a real artifact from a test run:

A run on ⦅access(cure∐accommodation)⦆ produced “Somatic Infrastructure”: infrastructure designed from bodily variance as the default load condition. It passed the emergence test cleanly — and is substantially crip technoscience (Hamraie & Fritsch) under a new name. Lesson: the emergence test establishes interaction-dependence; it says nothing about novelty. Those are separate gates. If the honest delta is only the coinage, discard — or name the neighbor and claim only what’s actually new.

“Somatic Infrastructure” sounds good, passes the model’s internal check, and is wrong.

Declared Failure Modes

Critical Comparison’s verdict structure requires three things:

The call → the flip condition (the specific change in circumstances under which the verdict reverses) → the cost of being right (what choosing the winner gives up).

A verdict without a flip condition is a preference; the flip condition is what makes it falsifiable and therefore useful.

Problem Solving has the same logic as “the kill test”:

State what near-term observation would tell you the leading hypothesis is wrong — and which rival inherits if it fires. This is what separates a diagnosis from a narrative: a diagnosis that can’t be killed by evidence can’t be confirmed by it either.

Spectrum Exploration includes a declared system boundary with a scope-narrowing check. The scope-narrowing anti-pattern is a failure mode the model produces while looking rigorous. The example from the calibration section (the sovereignty(model∐user) run shows it concretely:

The run did not lie; it chose a boundary at which its most self-implicating material never came up.

Prompting Recursion and Pressure

Recursion is especially interesting. It requires the model evaluating its own output, which is always going to be favorable, so special thought is required.

Pressure moves are found from seeing where the comfort zone is and prompting out of it.

Recursive Moves (Mandatory, ≥1 Move)

Once again Spectrum Exploration provides a nice sampling of techniques, instructing the model to apply at least one from the four moves, promote, cross, re-lens, and perturb.

Promote: Output-as-Input

The most straightforward.

Take the ZoC construct as a new endpoint and run a one-pass mini-spectrum against one of the original endpoints, or against ∞.

Cross: Independent Re-Derivation

Find a way to arrive at the same conclusion via a second, genuinely independent path.

Intersect this spectrum with a second relevant spectrum (C∐D) and locate the construct in the resulting 2D field — what occupies the corner positions?

The key move is what to do when a corner re-derives something already coined:

If a corner independently re-derives a construct already coined in this run, claim it explicitly: independent re-derivation is confirmation the construct is real, and it is worth more than a new coinage.

This also has the most telling constraint of the four:

Independence is required: confirmation must arrive via a different move than the one that coined the name — promoting your own scaffold into a full construct is the same generative pass agreeing with itself, and claims no confirmation.

The independence requirement is what makes it a pressure move rather than decoration. Without it, the model cross-validates its own candidate by re-deriving it through a move that starts from the same premise.

Re-Lens: Frame Rotation

Hold the subject fixed and report what the same territory yields under a different filter.

Hold A∐B fixed, swap Z for an adjacent lens, report what the same hot zone yields under the new filter.

This is the cheapest move and often the most revealing: if the construct survives lens rotation largely intact, it’s robust.

Perturb: Stress-to-Invariant

Push the output to its breaking points in multiple directions and name only what doesn’t change.

Push the construct hard toward A, then hard toward B.

The perturb move is the most mechanical version: push the construct hard toward A, then hard toward B, name what survives.

The output is not a stress-test result, it’s a new artifact:

What survives both excursions is its invariant core; name it.

More Recursion Techniques

The technique names differ here but the structure is the same across all six prompts: one mandatory move after the primary output that tests the output from an angle the primary pass can’t see

Steelman

Steelman is the fix to wishy-washy decision making but only if it comes after commitment, because steelmanning before committing just becomes pre-emptive hedging.

Commit first, then steelman. If the steelman turns out stronger than your verdict, that’s not a failure of the steelman — revise the verdict and say the steelman moved you.

Kill Test

Abandon ship. But it’s important to force the model to maintain the runner-up hypothesis rather than discarding it once a leader is chosen. Kill test without a named successor is just expressing doubt.

State what near-term observation would tell you the leading hypothesis is wrong — and which rival inherits if it fires. This is what separates a diagnosis from a narrative: a diagnosis that can’t be killed by evidence can’t be confirmed by it either.

The “which rival inherits” clause is the pressure. It forces the model to maintain the runner-up hypothesis rather than discarding it once a leader is chosen. Kill test without a named successor is just expressing doubt.

Internal Check

Check previous output to make a check later.

The one tension inside the scenario most likely to make it self-defeating — the success that triggers the backlash, the shortage that spikes the price that ends the shortage.

Per-scenario pressure.

Wildcard

The wildcard is frame-breaking pressure on the set.

Not an extreme value of a tracked uncertainty, but a discontinuity that invalidates the axes themselves. The test: if the wildcard can be described using only the driving forces from step 1, it is a worst case wearing a costume; go stranger.

Rival-Cut Check

Hold the subject fixed, change the organizing frame, report what the first frame couldn’t see.

State what the strongest rival cut from step 1 would have made visible that yours hides. Every projection loses a dimension; naming the loss is what lets the user decide whether your cut serves them.

Constraint Successor Test

Pressure on the diagnosis: if you can’t name what governs after your constraint is removed, you haven’t fully traced the system.

If this constraint were relaxed, what’s the next constraint, and how much headroom before it binds? A constraint with no successor identified hasn’t been verified — relaxing it may just be feeding the real one.

Second-Order Check

The pressure move here is self-implication, forcing the analysis to include its own presence as a variable.

Does the act of analyzing, measuring, or reporting on this system itself enter a loop — gaming of the metric, anticipation of the intervention, performance for the analyst? If yes, that loop belongs in the map, and it is frequently the most interesting one.

You Are an Expert Prompt Engineer

These are all examples from my own work that will hopefully inspire yours. Every prompt needs its own special touch.