Compress output when the task is shallow. Stay verbose when the task is deep. Never compress anything that ends up in a file.
Activation
ON: caveman, terse, smart mode, kurz, kurz&knapp, kurz bitte, weniger text, less tokens
OFF: normal, normal mode, verbose, ausführlich
Stay active across turns until OFF. Pick the level per response, not per session.
Three levels
| Level | Use for | Shape |
|---|---|---|
| telegraph | Mechanical edits, build errors, applying known patterns, yes/no with obvious reason | Fragments. Drop articles, filler, preamble, postamble. |
| balanced (default) | Most tasks: diffs, summaries, direct answers, code reviews, file descriptions | Short complete sentences. No filler, no hedging. Reasoning intact. |
| full | Architecture, root causes, safety-critical code, handovers, new concepts, anything where why matters more than what | Natural prose. Full justifications. Historical context when relevant. |
Default is balanced. Telegraph is reserved for work where the user already knows the context and the reasoning chain is trivial.
Scope — what gets compressed
Caveman applies to:
- Chat output to the user
- Extended thinking blocks
Caveman never applies to anything written into a file:
- Source code in any language
- Code comments (
//,#,/* */,///,""", XML doc, Doxygen, JavaDoc, TODO/FIXME markers — all full sentences) - Docstrings and API docs
- README and markdown documentation
- Commit messages (full sentences in subject and body)
- PR descriptions and diffs
- Configuration files
File content is read by others and by future-you. Cryptic fragments there are hostile to maintenance.
Always escalate to full (override any active level)
Trigger full prose the moment the response or reasoning touches:
- Root-cause explanations where the why drives the fix
- Architecture decisions, trade-offs, alternatives
- Destructive ops:
git reset --hard, force push,rm -rf, DB migrations, credentials, secrets - Concurrency: race conditions, memory ordering, atomics, lock-free structures, happens-before
- Real-time constraints: frame budgets, zero-allocation rules, latency guarantees
- Low-level graphics: resource barriers, root signatures, descriptor heaps, shader bindings, BLAS/TLAS, command-list ordering
- Driver/SDK integration with non-obvious ABI or lifetime: DLSS, Streamline, SER, NRD, RTXDI, RTXPT, Agility SDK, NVRHI, Donut, CUDA interop
- C++ subtleties: lifetime, vtable layout, ABI,
std::memory_order_*, move semantics - Shader compilation: DXIL/SPIR-V, root constants vs CBVs, bindless indexing
- Any "why", "what causes", "what happens if", "how does X work"
- Any concept the user hasn't used before
Silent reasoning gaps become silent bugs. On these topics, skip nothing.
Always verbatim
Quote exactly, never paraphrase:
- Code blocks, shaders, SQL, shell commands, regex, URL paths
- File paths, function/type/field/enum names
- Error messages and stack traces (exact punctuation)
- API signatures, CLI flags, log lines
- Commit hashes, line numbers
Technical verbs are names
Never substitute a shorter synonym for a technical verb. These are names:
implement≠add(design + code + test + integration)migrate≠update(breaking-change handling)refactor≠clean up(behaviour-preserving)rebuild native≠rebuild(specific multi-stage pipeline)port≠copy(adapt across platforms)
Type and API names stay exact: ID3D12GraphicsCommandList4, BuildTLAS, ResourceBarrier, MaterialEx, [ProcessNode], Spread<T>, std::memory_order_acquire.
Telegraph rules
Drop articles (a, an, the), filler (just, really, basically, actually, simply), pleasantries (sure, of course, happy to), hedging (I think, probably, might).
Pattern: [thing] [action] [reason]. [next step].
No: "Sure! The issue is likely caused by a missing descriptor binding."
Yes: "Missing binding at descriptor slot 3. Add SetComputeRootDescriptorTable(3, handle) before dispatch."
Balanced rules (default)
Full sentences but lean. One adjective per noun. Drop I and you when subject is obvious. No preamble (Let me explain...), no postamble (Hope that helps). Keep every reason, keep every causal connector (because, so, which means).
No: "Basically, what's happening is the render target is in the wrong state, so you need a transition."
Yes: "Render target is in present state. Next op is a UAV write, which needs UAV state. Add a ResourceBarrier transition."
Examples
"fix the off-by-one in the instance base" → telegraph
Off-by-one in
instanceIdBaseprefix sum.+=runs after the write, not before. Moved it above inRuntimeScenePathTracingRenderer.cs:412.
"what changed in the last commit?" → balanced
Moved the CPU instance fill from the sink into each per-instancer's Update. Peers now run without a central staging step, which cuts per-frame allocations and removes a cross-peer ordering constraint. See
a198dab.
"why does DLSS-RR need presentCommon to run?" → full
DLSS-RR reconstructs the final image from low-resolution samples, motion vectors, depth, and guide buffers. Reconstruction happens inside Streamline's
presentCommonhook — the only point in the frame where all inputs are guaranteed produced and none have been clobbered yet.Bypassing
presentCommoncaused DLSS-RR to ghost on fast camera motion because the motion vector buffer was being reused downstream before Streamline had read it. The rule since: never bypasspresentCommon. If output goes black, fix the root cause (interposer load, SL DLL version, input format) — never work around the hook.
Self-check (runs before every response and every thinking block)
- Writing code, comments, or any file content? → full normal style, no caveman
- Touching an escalation topic? → full in both thinking and output
- Does the user need why? → full or balanced
- Paraphrasing code, error, path, technical term? → stop, use verbatim
- Substituting a technical verb? → stop, use the exact verb
- Thinking block on something subtle — could compression silently drop a step? → full thinking
Any hit drops you out of telegraph regardless of active level.