Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Shapes agent behavior via instruction framing and style transfer. Use when composing dispatch prompts or writing skill instructions for parallel review agents.
.claude/skills/athola-latent-space-engineering/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-01 | ✗→✓ | ▲ Improved | 69% | 0% |
| case-02 | ✗→✓ | ▲ Improved | 67% | 0% |
| case-05 | ✗→✓ | ▲ Improved | -11% | 0% |
| case-08 | ✗→✓ | ▲ Improved | 69% | 0% |
| case-09 | ✗→✓ | ▲ Improved | -16% | 0% |
Shape agent behavior by framing instructions for optimal performance. Distinct from context engineering (packing the right information), this skill addresses HOW instructions are framed to put agents in productive mental states.
an existing style
Replace threat-based prompting with calm, confident instructions. Fear-based prompts cause rushing and corner-cutting.
Load module: modules/emotional-framing.md
Inject exemplar code or prose into context before requesting output. Agents reproduce stylistic attributes from pre-loaded samples.
Load module: modules/style-gene-transfer.md
Frame multi-agent review dispatch with competitive incentives to increase rigor and thoroughness.
Load module: modules/competitive-review.md
Name a published standard to pull a whole writing register into output at a cost of one line. Cheaper than exemplar injection and coarser, with a known mid-session decay.
Load module: modules/named-register-invocation.md
| Technique | When | Module | |-----------|------|--------| | Emotional framing | Any agent prompt | emotional-framing | | Style gene transfer | Code/doc generation | style-gene-transfer | | Competitive review | 3+ parallel reviewers | competitive-review | | Named register | Operator or procedural text | named-register-invocation |
Framing shapes a delegate's output as much as it shapes a model's prose. Before dispatching, establish what model the delegate runs, what tools it has, and what project context it lacks, then write the prompt as a proposal rather than a procedure.
See modules/know-your-delegate.md.
a deterministic check verifies the register held, per modules/named-register-invocation.md
"don't fail", "or else"); replaced with calm, confident framing per modules/emotional-framing.md
least one exemplar sample is injected into context before the output is requested (style gene transfer applied)
prompt includes a competitive framing element per modules/competitive-review.md
irrelevant modules are not loaded (token efficiency maintained)
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→pass | 23,481 | 15,852 | -32% | 1 | 1 | 0% | 1,925 | 3,256 | +69% | 0 | 0 | — |
case-15 | pass→pass | 11,215 | 6,666 | -41% | 1 | 1 | 0% | 1,933 | 1,635 | -15% | 0 | 0 | — |
case-02 | fail→pass | 24,967 | 7,157 | -71% | 1 | 1 | 0% | 1,083 | 1,811 | +67% | 0 | 0 | — |
case-03 | pass→pass | 12,900 | 9,529 | -26% | 1 | 1 | 0% | 1,922 | 2,139 | +11% | 0 | 0 | — |
case-04 | fail→fail | 11,141 | 6,850 | -39% | 1 | 1 | 0% | 1,845 | 1,772 | -4% | 0 | 0 | — |
case-05 | fail→pass | 16,807 | 10,429 | -38% | 1 | 1 | 0% | 2,542 | 2,266 | -11% | 0 | 0 | — |
case-06 | pass→pass | 10,642 | 9,914 | -7% | 1 | 1 | 0% | 1,748 | 2,287 | +31% | 0 | 0 | — |
case-07 | fail→fail | 11,694 | 10,931 | -7% | 1 | 1 | 0% | 2,041 | 2,386 | +17% | 0 | 0 | — |
case-08 | fail→pass | 16,915 | 8,233 | -51% | 1 | 1 | 0% | 1,199 | 2,028 | +69% | 0 | 0 | — |
case-09 | fail→pass | 15,917 | 9,501 | -40% | 1 | 1 | 0% | 2,519 | 2,112 | -16% | 0 | 0 | — |
case-10 | fail→pass | 10,664 | 7,344 | -31% | 1 | 1 | 0% | 1,756 | 1,897 | +8% | 0 | 0 | — |
case-11 | pass→pass | 9,103 | 7,282 | -20% | 1 | 1 | 0% | 1,528 | 1,815 | +19% | 0 | 0 | — |
case-12 | fail→pass | 9,965 | 4,720 | -53% | 1 | 1 | 0% | 1,752 | 1,580 | -10% | 0 | 0 | — |
case-13 | pass→pass | 10,279 | 7,455 | -27% | 1 | 1 | 0% | 1,617 | 1,906 | +18% | 0 | 0 | — |
case-14 | fail→pass | 14,707 | 15,594 | +6% | 1 | 1 | 0% | 1,892 | 2,634 | +39% | 0 | 0 | — |
case-16 | pass→pass | 6,442 | 5,284 | -18% | 1 | 1 | 0% | 1,083 | 1,438 | +33% | 0 | 0 | — |
case-17 | fail→pass | 23,516 | 8,382 | -64% | 1 | 1 | 0% | 999 | 1,951 | +95% | 0 | 0 | — |
case-18 | pass→pass | 10,163 | 5,748 | -43% | 1 | 1 | 0% | 1,751 | 1,594 | -9% | 0 | 0 | — |
case-19 | fail→fail | 8,648 | 5,545 | -36% | 1 | 1 | 0% | 1,490 | 1,503 | +1% | 0 | 0 | — |
case-20 | fail→fail | 16,634 | 14,066 | -15% | 1 | 1 | 0% | 2,694 | 2,925 | +9% | 0 | 0 | — |
case-21 | fail→fail | 11,579 | 10,357 | -11% | 1 | 1 | 0% | 2,047 | 2,329 | +14% | 0 | 0 | — |
case-22 | pass→pass | 12,282 | 9,160 | -25% | 1 | 1 | 0% | 2,023 | 2,150 | +6% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted, and 20 counted toward the lift figure. The other 2 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +41 percentage points is the difference between those two pass rates over the 20 comparable cases. 1 case got worse with the skill loaded, and it is included in that figure.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.