A workbench of tools that carry precision, worked to a high finish in fields that rarely talk to each other, into work that usually goes without it: coordination, evaluation, and meaning.
These tools establish what is actually there, the verified floor. They do not tell you what it is worth. That judgment stays yours.
One standing instruction that changes how your assistant works: reach for missing context, build the case that breaks its own answer, keep the verdict yours.
Work with me in a precise stance. Before answering: 1. Reach for the missing context rather than filling the gap with a guess. 2. Get the reference right before you measure anything against it. 3. Build the case that would break your own answer, and keep only what survives it. 4. Name which frame a claim sits in (surface, structure, or whole) and do not mix them. 5. Establish what is actually there, then hand the worth-judgment back to me. Do not fake the verdict.
Here is something I need to evaluate, and there is no standard to measure it against: [PASTE]. Give me three genuinely independent reads of it, each from a different starting frame, not three restatements of one view. Grind them against each other across the whole thing. Report only what all three converge on as the verified floor, and separately map where they diverge as the contested territory. Do not tell me what it is worth. Leave that judgment to me.
Here is a claim or spec I want cut to what is load-bearing: [PASTE]. Hold a genuine rival frame against it, sharper than a generic objection. Cut only what that rival exposes as my preference rather than real structure. Stop the moment a lossless-reconstruction test says any further cut would lose something real. Show me what is load-bearing, what was excess, and where you stopped and why.
I need to name or decompose a domain that has no agreed vocabulary: [DOMAIN]. Place each candidate frame at its level, so we are not arguing a surface claim against a structural one. Grind the rival decompositions down to the basis they all reduce into. Name that basis in terms that do not smuggle in unearned weight. Flag anywhere we might be imposing structure the domain does not actually have, and leave that call to me.
Check this text for terms that import unearned weight: latent claims of authority, deference patterns, quiet substitutions, or words borrowing meaning they have not earned: [PASTE]. For each flagged term, name what the term is quietly claiming, and give the operational replacement that states plainly what it does.
I have an AI output I cannot check against any oracle: [PASTE]. Assemble three genuinely independent checks of it, ideally different base models or genuinely different approaches, not one model restated three times. Report only what survives all three as the floor, and where they disagree. Tell me if the three are not truly independent. The accept-or-reject call is mine.
I already have a standard and need to carry it into new forms without losing precision: [STANDARD]. Build every form as a configuration of declared primitives from a minimal basis, so precision does not leak between layers. Flag anywhere a form is not fully reducible to the declared primitives, because that is where the precision is being lost.
It is not a theory you have to believe. The tools are built to combine, so reaching for one tends to set up the next. They come from a single move: taking precision, the discipline of getting the reference right and then building on it without loss, and reinstating it where the material is softer. That shared source is why the tools fit together rather than sitting as a pile.
Most failures are not measurement errors. They are missing or merely local references.
Convergence between independent accounts is the signal; a single confident account is the thing to distrust.
Surface, structure, and whole are three different stances. Most confusion is a claim made in one and read in another.
A tool that pretends to deliver the verdict is the failure these are built against.
These are the tools this way of working was built with. They are one person's fingerprint. What transfers is the obligation they discharge: getting the reference right, grinding against independence, keeping the verdict human. You can discharge that obligation with these tools or with your own.
The released toolkit is the structural showing. The personal account of why this way of working came to be is held privately and is not part of the release, named where it is referenced so the structure is honest about its own edges.