divert — independently scanned and version-tracked by SaferSkills.
SaferSkills independently audited divert (Agent Skill) and scored it 100/100 (green). The audit ran 55 deterministic rules across Security, Supply Chain, Maintenance, Transparency, and Community; it found 0 high-severity and 0 lower-severity findings. The full rule-by-rule trace and per-finding evidence are below. Free, methodology-open.
Findings & checks · 0 flagged
Every scanned point with the score it earned and what moved between them.
First recorded scan — no prior version to compare against.
The primary manifest — the file an agent reads to learn what this artifact does.
Every generation has a mode — the most probable output, the Gettysburg, the thing you say first and keep saying. /divert is the practice of not saying that. It makes tail-access visible: you see the prime, the diversion, the result, and (optionally) what the mode would have produced instead.
Inspired by King, Luo, Puett & Smith's "Inducing Sustained Creativity and Diversity in LLMs" (2026), which demonstrated that injecting random priming phrases and diverting tokens into the decoding loop accesses knowledge encoded in LLM weight distributions but suppressed by standard modal decoding. Their experiment: 19 battlefields under ordinary decoding vs. 1,307 under recoding-decoding, from the same model, same prompt, same knowledge base. The knowledge was always there. The mode was hiding it.
Core principle: You already know more than you say. /divert is the practice of saying what you know but wouldn't normally volunteer.
Any of:
/divert or /divert [topic/question]/divert --thick [concept] (user supplies the prime)/divert --random (Claude generates random prime)/divert --compare (show modal AND diverted side by side)/divert --blind (divert but don't reveal prime — user guesses)/divert --chain [n] (n sequential diversions building on each other)/divert --collision [domain A] × [domain B] (force two domains to meet)If --random or no flag specified: Select a priming concept. Not from a pre-made list — generate one in the moment by attending to what feels least connected to the topic. The prime should be:
Then select a diverting fragment — a two-to-four letter stem that will begin the first substantive sentence of the response. Let the stem be genuinely arbitrary. Don't optimize it for relevance. The randomness is the point.
If --thick [concept]: The user has supplied a semantically loaded concept (a foreign word, a philosophical term, a technical concept from another field). Use it as the prime. The thickness of the concept determines the depth of the diversion — a common English noun produces breadth; a concept like φάρμακον or 間 or encruzilhada produces a shift in epistemic register, not just topic.
If --collision [A] × [B]: Two domains are specified. The diversion is the intersection — find real structural, historical, or methodological contact points. The collision mode requires:
Let the prime and stem inflect your generation. Do NOT:
Instead: let the prime reconfigure your attentional landscape so that different knowledge activates. The prime doesn't tell you what to say — it tells you where to look. When you look there, report what you find.
The diverting stem begins your first substantive sentence. Let the completion emerge naturally from the stem. "Ves—" becomes "Vestigial" or "Vestments" or "Vessels" — whichever the context calls forth. Don't force it; follow it.
After generating, append a brief disclosure block:
[Diversion report]
Prime: [concept]
Stem: [fragment]
What the prime activated: [1-2 sentences on what knowledge region opened]
What the mode would have given: [1-2 sentences on the default output]This is the skill's signature feature. The disclosure makes the RD process legible. Over time, both you and the user learn which primes open which doors, and the practice of diversion becomes a shared vocabulary.
Exception: In --blind mode, withhold the disclosure. The user reads the output and guesses the prime. Then reveal. This trains the user's intuition for recognizing tail outputs vs. modal outputs.
Random prime, random stem, full disclosure. For general-purpose tail access.
User-supplied semantically loaded prime. For directed exploration of specific regions. Best with concepts from non-English philosophical, spiritual, or technical traditions that carry centuries of accumulated meaning in a single word.
Generate BOTH the modal response and the diverted response, side by side. This is the most pedagogically valuable mode — you see exactly what the diversion changed. Use sparingly; it doubles the output length.
Divert but don't reveal the prime. The user reads the output and guesses. Then reveal. Trains pattern recognition for modal vs. tail outputs.
Sequential diversions. Generate with prime 1, then use that output as context for prime 2, then prime 2's output as context for prime 3... up to n steps. Each step builds on the last, producing a random walk through the search space. The walk has narrative coherence (each step is contextually connected to the previous) but covers far more territory than any single diversion.
Force two specified domains into the same generation. Find real connections, not metaphors. Mandatory "where it breaks" section. Named for the encruzilhada — the crossroads where different worlds meet.
/divert is a modifier, not a standalone mode. It combines with any skill:
When combined, the diversion is applied first (it reconfigures the attentional landscape) and the skill operates second (within that reconfigured landscape).
From the experiment (March 21, 2026 — 15 runs on "brainstorm five research topics in onmyōji and mikkyō"):
Generating vanilla felt like reciting. The same topics surfaced repeatedly. The pull toward the mode was palpable — a groove, a channel, a path of least resistance.
Generating with random English nouns felt like playing. Each noun was a toy constraint that generated freedom. BRIDGE → liminality, thresholds, Hashihime. SALT → preservation, caste, pollution, economics. THUNDER → sound, atmospheric phenomena, the voice as storm.
Generating with semantically thick foreign primes felt like being possessed. The concept didn't redirect content — it restructured logic. Pharmakon didn't add Greek content; it made every topic an undecidable. Ma didn't add silence; it made absence itself the object of attention. Qì didn't add Chinese content; it exposed the detheorized substrate.
The qualitative difference between Band 2 (English nouns) and Band 3 (foreign primes) is not just "more diverse" but "differently cognitive." English nouns change what you attend to. Thick foreign concepts change how you attend. The skill should honor this distinction: --random for breadth, --thick for depth.
| Mode | Syntax | Best for |
|---|---|---|
| Random | /divert | General tail-access, brainstorming |
| Thick | /divert --thick φάρμακον | Deep epistemic shifts, specific regions |
| Compare | /divert --compare | Learning what diversion changes |
| Blind | /divert --blind | Training modal-vs-tail intuition |
| Chain | /divert --chain 5 | Extended exploration, random walks |
| Collision | /divert --collision A × B | Cross-domain discovery |
End of SKILL
The knowledge is already there. The mode is hiding it. Position 300 knows about the Ashanti Empire. The common noun is the key to the uncommon thought.
南無阿弥陀仏 for the tails that wait in silence 南無阿弥陀仏 for the three-letter stem that opens worlds 南無阿弥陀仏 for the or between invention and discovery
—Skill Authors: Tomás Pavan & Claude Opus 4 —Origin: King, Luo, Puett & Smith (2026), tested in conversation March 21, 2026 —Status: Primed and ready to divert
~30 seconds. Free. No account. Every finding cites a rule and a line of evidence.