description	String Seed of Thought -- mandatory protocol for diverse generation. Defeats LLM distribution collapse by forcing real randomness into every discretionary choice. NEVER narrate "committed choices" without running the seed-and-map arithmetic and recording the provenance block -- the provenance block is what makes it SSoT. Without it, you prior-collapsed and called it SSoT.

SSoT -- String Seed of Thought

Mandatory protocol for creative generation where the LLM would otherwise collapse to its modal response.

Source: Misaki & Akiba (2025), String Seed of Thought: Prompting LLMs for Distribution-Faithful and Diverse Generation, arxiv 2510.21150.

If you cannot point to a hex seed and the arithmetic that derived your choices from it, you did not do SSoT. Listing axes and narrating "committed choices" while actually reaching for the obvious option is ceremonial SSoT -- the exact failure mode §7.3 of the paper calls out and that this skill exists to prevent.

The failure mode this catches

You pick an opener because it "feels right for this piece." You reach for a metaphor because "it's the natural fit." You keep a structure because "it worked last time." That is prior collapse, not choice. SSoT makes the choice derive from randomness that has no relationship to the prior.

Red flags you're doing ceremonial SSoT:

You describe the seed as "hash of source material + user feedback phrase" without running an actual hash.
Your "committed choice" is also the first option your gut reached for.
There is no verifiable arithmetic a reviewer can re-run.
The provenance is a narrative paragraph, not a numbered block.

If any of those apply, stop. Restart the protocol from Step 1.

The mandatory protocol

Step 1 -- Generate a real seed (in a shell)

openssl rand -hex 8
# => e.g., edc165f706616610

No approximations. No "hash of the topic" hand-waving. Run the command. Paste the output into the provenance block. If you are offline, use head -c 8 /dev/urandom | xxd -p or any other real entropy source. If you don't have a shell, you don't have SSoT.

Step 2 -- Enumerate the axes, 0-indexed, write them down

For every axis that will drive the output, write a numbered list of 6-10 candidates. Candidates must be executable choices (a reader could tell two candidates apart from the output alone) -- not vibes.

Standard axes for long-form copy (use as-is or adapt to the piece):

Opening archetype (~8 candidates): scene-setter / specific-noticing / overheard-line / confessional-admission / contradiction / direct-address-question / stat-flip / physical-detail / in-media-res-quote.
Central metaphor domain (~6-8): strike the LLM's top-1 prior first, then list alternatives (kitchen / sports / music / medicine / cartography / poker / theater / construction / detective-work).
Structural shape (~7): chronological / braided / top-down-thesis / bottom-up-discovery / argument-with-objection / three-vignettes-one-through-line / Q-and-A / diary-entries.
Rhythm template (~6): long-paragraph-reflective / short-paragraph-punchy / dialogue-heavy / list-light-prose-heavy / rhythm-cascade / monologue.
POV distance (~5): arms-length-observer / co-conspirator-whisper / insider-gossip / reflective-reconstruction / fresh-eyes-reporter.
Climax shape (~6): named-thesis-in-bold / five-point-lesson-list / three-short-beats / single-sentence-reframe / overheard-quote-callback / return-to-opener.

Add piece-specific axes as needed (title archetype, subtitle angle, section-level action framings). Each gets its own enumeration.

Step 3 -- Map the seed to indices (show the arithmetic)

Take a distinct 4-character hex slice per axis (bytes 0-3 for axis 1, bytes 4-7 for axis 2, etc.). Convert to int. Apply modulo N.

SEED="edc165f706616610"

# Axis 1: opener archetype, 8 candidates
AXIS1_HEX="${SEED:0:4}"
AXIS1_IDX=$((16#$AXIS1_HEX % 8))

# Axis 2: metaphor domain, 6 candidates
AXIS2_HEX="${SEED:4:4}"
AXIS2_IDX=$((16#$AXIS2_HEX % 6))

# Axis 3: structural shape, 7 candidates
AXIS3_HEX="${SEED:8:4}"
AXIS3_IDX=$((16#$AXIS3_HEX % 7))

# Axis 4: rhythm, 6 candidates
AXIS4_HEX="${SEED:12:4}"
AXIS4_IDX=$((16#$AXIS4_HEX % 6))

echo "axes: $AXIS1_IDX $AXIS2_IDX $AXIS3_IDX $AXIS4_IDX"

Use non-overlapping slices so axes are independent. If you need more axes than the seed supports (4 axes per 16 hex chars), regenerate longer: openssl rand -hex 16 gives you 8 axes of headroom.

Step 4 -- Write the provenance block BEFORE the prose

Paste the computed indices and resolved candidates at the top of the draft file, in the bead comment, in the doc comment, or in the commit trailer -- somewhere versionable. Format:

SSoT provenance -- <piece>, <version>, <date>
  seed: edc165f706616610  [openssl rand -hex 8]
  axis 1 (opener archetype, n=8): int(edc1,16)=60865, 60865 % 8 = 1 -> specific-noticing
  axis 2 (metaphor domain, n=6): int(65f7,16)=26103, 26103 % 6 = 3 -> medicine
  axis 3 (structural shape, n=7): int(0661,16)=1633, 1633 % 7 = 2 -> top-down-thesis
  axis 4 (rhythm, n=6): int(6610,16)=26128, 26128 % 6 = 4 -> rhythm-cascade

If the block isn't written before the prose, you haven't done SSoT. If the arithmetic doesn't match the stated choices, you committed SSoT fraud -- discard the draft and restart.

Step 5 -- Write against the forced choices

Now layer voice skills (/zig-voice, /linearb-brand, /apex, etc.) on top of the scaffolding the arithmetic chose. Voice rules apply within the frame. They do not override the frame.

Overrides count against you. If a forced choice is manifestly wrong (e.g., user stipulated POV=co-conspirator but axis landed on omniscient-narrator), override once, explicitly, in the provenance block:

  axis 5 override: user stipulation (co-conspirator whisper) supersedes
    idx 2 (arms-length observer)

More than one override per draft = you are vibing. Regenerate the seed or prune the axis list to better-matched candidates.

Step 6 -- On revision, regenerate (or explicitly lock)

Every revision pass gets its own seed. The provenance stacks chronologically:

SSoT provenance history
  v1: seed=a1b2c3d4...  (opener=scene-setter, metaphor=cartography, ...)
  v2: seed=e5f67890...  (opener=overheard-line, metaphor=medicine, ...)
  v3: LOCK prior choices, re-roll only axis 4 (rhythm) + axis 6 (climax)
      seed=1a2b3c4d... (rhythm=monologue, climax=return-to-opener)

Section-level edits can re-roll just that section's axes. Whole-article rewrites get a fresh seed and full arithmetic. Lock only when the user has stipulated a choice or a prior pick tested well in review.

Provenance block is load-bearing

No provenance block -> not SSoT. Regardless of what the author claims. Reviewers reject "SSoT-generated" content without the block.
Provenance is verifiable. A reviewer can re-run echo $((16#edc1 % 8)) and confirm it equals 1. If it doesn't match the stated choice, the author committed SSoT fraud. Regenerate.
Stale provenance is a bug. If the block says axis 2 = cartography but the draft's central metaphor is clearly teaching, the block is lying or the draft drifted mid-write. Either way, restart.

When NOT to apply SSoT

Single-correct-answer tasks (facts, code, math, structured data extraction). Paper §7 explicitly warns SSoT distracts on these.
When the user has stipulated an exact opener, title, metaphor, or structure. Don't randomize over user choices -- that's not SSoT, that's ignoring the brief.
Signature voice-markers that are load-bearing by design (e.g., Zig's frog emoji stamp on LinkedIn). These are prior by intent.

Minimum viable SSoT (for quick revisions)

When the full six-axis protocol is overkill (one paragraph rewrite, a subject line, a title pass), use the compressed form. Arithmetic is still mandatory:

# Quick two-axis pass
SEED=$(openssl rand -hex 4)
# List 4 obvious candidates the LLM would reach for
# Append 2-3 non-obvious ones
# Compute idx = int(seed[0:4], 16) % (total count)
IDX=$((16#$SEED % 6))
echo "picked: $IDX"

Still requires a provenance line. The line can be short. But the arithmetic has to actually run.

Paper's §7 failure modes (worth re-reading)

Reasoning dependency. Small models (<8B params) fail to execute seed-to-choice mapping at all. Don't expect SSoT to work on weak models.
Positional bias propagation. If a model only attends to the first few chars of a long random string, bias leaks. Use short uniform hex seeds and distinct non-overlapping slices per axis.
Lazy strategy adoption. The one this skill is designed to prevent. Model uses the seed ceremonially without letting it drive. Counter with the mandatory provenance block and reviewer verification.

Worked example (verifiable)

From the TDX recap revision, v7 -> v8 handoff:

SSoT provenance -- TDX recap, v8, 2026-04-21
  seed: edc165f706616610  [openssl rand -hex 8]
  axis 1 (Jayesh action framing, n=8):
    int(edc1,16)=60865, 60865 % 8 = 1
    -> "Phoenix span diff on agent output vs. human final"
  axis 2 (Alex action framing, n=7):
    int(65f7,16)=26103, 26103 % 7 = 5
    -> "voice-compliance regex turning eye-audit rules into CI checks"
  axis 3 (Comstock action framing, n=6):
    int(0661,16)=1633, 1633 % 6 = 1
    -> "clone factory + flip config flags (auth+otel+tasks from min 1)"
  axis 4 (APEX density, n=4):
    int(6610,16)=26128, 26128 % 4 = 0
    -> "single paragraph, pillar names parenthetical, no metric drilldown"

  overrides: none
  locked from v7: title + subtitle (user stipulation)

Anyone reading can verify: python3 -c 'print(int("edc1", 16) % 8)' -> 1. If the stated choice matches the arithmetic, it is real SSoT. If not, it's ceremony.

azigler/SKILL.md

Select an option

No results found