Instruction-file hygiene for LLM lexical tics: linting, mirroring, hook filtering
Parent: LLM Context Engineering · Published reference · snapshot 2026-09-23 · skill ai-mcp-sdk-prompting/references/instruction-file-tic-hygiene.md
↓ Facts as markdownall context files
Source-anchored research for Instruction-file hygiene for LLM lexical tics: linting, mirroring, hook filtering. Run status: COMPLETED.
These notes link each claim to its source. Some citations identify archived research reports that are no longer hosted on this site. A published reference means the content is available; it does not certify independent review or accuracy.Read the editorial policy and follow the sources before relying on a claim.
Definitions
- Overused-word lists are the norm in public 'humanize' skills, e.g. a Tier 1 banned list of 18 words including delve, tapestry, pivotal, testament; such files embed the seeded tokens in the instruction file itself. (low confidence) (verified as of 2026-09-23) [source]
How it works
- Anthropic states Claude Opus 4.5 and 4.6 are more responsive to the system prompt, so prompts written to reduce undertriggering may now overtrigger; the fix is to dial back aggressive language, e.g. replace 'CRITICAL: You MUST use this tool when...' with 'Use this tool when...'. (low confidence) (verified as of 2026-09-23) [source]
- Anthropic also advises removing over-prompting such as 'If in doubt, use [tool]', which causes overtriggering, and says Claude 4.6 models may overtrigger on anti-laziness instructions that earlier models needed. (low confidence) (verified as of 2026-09-23) [source]
- Language models show structural priming: they reuse recently seen syntactic structures, including when prime and target share no lexical overlap, indicating abstract structural representations. (high confidence) [source]
Problems, failure modes and limitations
- Hooks cannot directly modify assistant message text. Stop can only prevent stopping and continue the conversation, and last_assistant_message on Stop/SubagentStop is read-only input for analysis. (medium confidence) (verified as of 2026-09-23) [source]
- A mechanistic study of 40,000 samples reports that explicitly naming a forbidden word can raise its generation probability; priming failure accounts for 87.5% of negative-constraint violations and override failure for 12.5%. (medium confidence) [source]
- The same study finds violation probability tracks the model's intrinsic 'semantic pressure' toward the token (p = sigmoid(-2.40 + 2.27*P0)), so the risk is highest for high-prior tics such as delve, which are exactly the ones people list. (low confidence) [source]
Comparisons and alternatives
- Independent evidence that emotional or emphatic framing improves accuracy is mixed: a 2026 arXiv study tests emotional framing across GPT, Llama, Qwen3 and DeepSeek on GSM8K, BBH, MedQA and others, and secondary reporting cites contradictory politeness studies (e.g. a 2025 GPT-4o result where impolite prompts scored 84.8% vs 80.8% for very polite). Effect sizes were not verified from the paper text. (low confidence) [source]
- Anthropic's prompting docs tell authors to say what to do instead of what not to do, and note that the formatting style of the prompt can influence response style, so instruction files should be written in the target voice. (high confidence) (verified as of 2026-09-23) [source]
- A workable lint design: scan instruction files for tokens on a tic list and classify each hit as ban-list entry, example text or incidental prose; flag examples and prose hits first, since they show the word as a model to copy rather than as a prohibition. This is inferred from the sources, not documented anywhere. (low confidence) [source]
- Practical mitigation for instruction files: write the instruction file in the plain style wanted in the output (prose, little markdown) and state the desired format positively, rather than only listing prohibitions. (low confidence) [source]
- EQ-Bench's Slop Score is a weighted composite (60% slop words, 25% not-X-but-Y patterns, 15% slop trigrams) computed against human baselines (wordfreq for words, a human essay/creative corpus for trigrams); it is not an AI-vs-human classifier and may underreport outside creative writing and essays. (medium confidence) [source]
- Arena's baseline counts on Opus 5 (510 vs 158 words per response against Opus 4.5, em dashes 2.3x, 'load-bearing' about 2x, sentences 58% longer) show per-response rates that can serve as a before figure, but the guide's fix was written for Fable 5.1 and nothing shows it works on Opus 5. (low confidence) (verified as of 2026-08-12) [source]
Facts and statements
- OpenAI's GPT-4.1 guide says all-caps or incentives like bribes or tips are generally unnecessary, recommends starting without them, and warns that existing prompts using them may make GPT-4.1 attend to them too strictly. It attributes this to GPT-4.1 following instructions more literally than predecessors. (medium confidence) [source]
- Two vendors independently converge on the same hygiene rule: with current literal, steerable models, drop emphatic markers by default and add them back only if a specific prompt needs them. (medium confidence) (verified as of 2026-09-23) [source]
- Guidance is model-scoped: the Anthropic advice targets Opus 4.5/4.6 and later; the same emphatic wording may still be needed on weaker or older models, so a lint rule should be per-model, not universal. (low confidence) (verified as of 2026-09-23) [source]
- A Stop hook can act as a post-generation filter: it inspects the final response and blocks with a reason, forcing the model to revise before the user sees the result. (medium confidence) [source]
- A regex-only Stop hook produces false positives; combining the regex with a second condition (e.g. prior tool-call activity) removes most of them. (low confidence) [source]
- A Stop hook that re-blocks without checking stop_hook_active can loop indefinitely; one reported case ran about 50 minutes and consumed the session quota. A nag-once pattern (block the first violation, let an identical repeat pass) avoids it. (medium confidence) (verified as of 2026-05-03) [source]
- PostToolUse can replace a tool's output before the model sees it (updatedToolOutput), so filtering applies to tool results but not to the model's own prose. (medium confidence) (verified as of 2026-09-23) [source]
- Lexical overrepresentation (delve, intricate, underscore; 21 focal words in scientific abstracts) is not explained by architecture, algorithm or training data; RLHF is a suspect but evidence is mixed. The lint word list must therefore be empirical and per-model, not a fixed canon. (medium confidence) [source]
- Anthropic's prompting guide says the formatting style of a prompt may influence Claude's response style, and that removing markdown from the prompt can reduce the volume of markdown in the output. (medium confidence) (verified as of 2026-09-23) [source]
- Prompt format alone can change measured performance a lot: up to 76 accuracy points on LLaMA-2-13B in few-shot settings, persisting with larger models, more shots, or instruction tuning; format performance correlates only weakly across models. (high confidence) [source]
- Format sensitivity falls with model strength: GPT-4 variants varied less across plain text, markdown, JSON and YAML than GPT-3.5-turbo, and no single format is best across models. (medium confidence) [source]
- A 2026 study (9,649 experiments, 11 models, YAML/markdown/JSON/TOON context) found no significant aggregate accuracy effect of context format (p=0.484), though individual models, mostly open-source, showed format-specific differences. (low confidence) (verified as of 2026-02-01) [source]
- The Verbal Tic Index (VTI) combines TicRate (share of responses containing a tic), MATTR lexical diversity, a sycophancy score and phrase repetition rate with weights 0.3/0.2/0.3/0.2; the authors say other weights may reorder models, so it is a comparative summary, not an absolute measure. Use response-level tic share as a simple before/after rate. (medium confidence) [source]
- The VTI study is observational (160,000 responses, 8 models); it reports no before/after intervention, so it gives baselines but no evidence of how much a prompt fix reduces tics. (medium confidence) [source]
- Anthropic's Fable 5.1 guide supplies a 'mannered prose' definition paragraph (or the short 'Please remove all mannered prose.') to add to a user message or system prompt, and tells users to test against their own evals, but publishes no measured before/after rate for it. (medium confidence) (verified as of 2026-09-23) [source]
- Automatic lexical metrics only partly capture slop: in a copy-editor annotation study, linear models on available automatic metrics reached AUPRC 0.52-0.55 on news and MS MARCO, and zero-shot GPT-4 judging scored kappa near 0, so a tic-rate drop does not prove the text is better. (medium confidence) [source]
Children
- Anti-laziness prompt legacy (frontier)
- Classifying hits: ban list vs example vs prose (frontier)
- Emphatic language in prompts and overtriggering
- Few-shot example style leakage (frontier)
- Fixed prompt-set regression harness for style (frontier)
- Hook-based filtering of agent output
- Linting instruction files for seeded vocabulary
- Negative instructions vs positive framing (frontier)
- Per-model prompt migration (frontier)
- Positive framing of format instructions (frontier)
- PostToolUse updatedToolOutput (frontier)
- Prompt-format mirroring and structural priming in output
- Stop hook nag-once pattern (frontier)
- Synonym substitution after word bans (frontier)
- Tic-word list curation per model (frontier)
- Verifying a vocabulary fix with before-and-after tic rates
Frontier under this node: Anti-laziness prompt legacy, Classifying hits: ban list vs example vs prose, Few-shot example style leakage, Fixed prompt-set regression harness for style, Negative instructions vs positive framing, Per-model prompt migration, Positive framing of format instructions, PostToolUse updatedToolOutput, Stop hook nag-once pattern, Synonym substitution after word bans, Tic-word list curation per model