Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Generate structured documents from composition templates. Combines reference templates with live data to produce reviewable output artifacts. Triggered by compose, document, generate report, create brief, write proposal.
.claude/skills/miosa-osa-compose/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-11 | ✗→✓ | ▲ Improved | 17% | 0% |
| case-12 | ✗→✓ | ▲ Improved | 3% | 0% |
| case-01 | ✗→✓ | ▲ Improved | 20% | 0% |
| case-03 | ✗→✓ | ▲ Improved | -7% | 0% |
| case-05 | ✗→✓ | ▲ Improved | 1404% | 0% |
Produces structured output artifacts by combining a composition template (ROM) with live workspace data. The template defines the skeleton. The agent fills it with current context. The output lands in output/ for human review.
/compose <template-name> [--for <receiver>] [--data <source>] [--status draft|review]| Arg | Required | Default | Description | |-----|----------|---------|-------------| | template-name | Yes | — | Name of template in reference/compositions/ | | --for | No | — | Intended receiver (adjusts genre encoding) | | --data | No | auto | Data source: auto scans relevant data/ + output/ | | --status | No | draft | Initial status of the output file | | --output | No | auto | Override output directory (default: output/{genre}/) |
reference/compositions/{template-name}.md--for specified, look up receiver's genre preference(from SYSTEM.md people table or company.yaml org chart)
data/, recent output/, and engine (if available) forrelevant context. Use search if engine exists, otherwise scan by keyword.
Apply Signal Theory encoding: resolve all 5 dimensions S=(M,G,T,F,W).
output/{genre}/{date}-{title}.md with frontmatter:yaml --- genre: {from template} type: {speech act} status: {draft|review} created_by: {agent name} created_at: {date} for: {receiver if specified} template: {template-name} ---
bash# Generate a quarterly review from template /compose quarterly-review # Write a proposal for a specific client /compose client-brief --for "ACME Corp" --status review # Create a weekly signal report using specific data /compose weekly-signal --data "data/pipeline.json" # Generate meeting prep doc /compose meeting-prep --for "Ed Honour"
Templates in reference/compositions/ follow this structure:
markdown--- genre: proposal default_output: output/proposals/ sections: - objective - key_messages - supporting_data - call_to_action data_sources: - data/pipeline.json - output/analyses/ --- # {title} ## Objective {Agent fills: one sentence, what outcome} ## Key Messages {Agent fills: 3-5 bullets from gathered data} ## Supporting Data {Agent fills: relevant metrics, quotes, evidence} ## Call to Action {Agent fills: single unambiguous ask + deadline}
output/{genre}/{date}-{title}.mddraft (user reviews and promotes to approved)reference/compositions/ must contain the named templatedata/ and/or output/ for context gathering| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-11 | fail→pass | 35,342 | 18,625 | -47% | 1 | 1 | 0% | 3,956 | 4,628 | +17% | 0 | 0 | — |
case-12 | fail→pass | 29,406 | 10,536 | -64% | 1 | 1 | 0% | 2,520 | 2,602 | +3% | 0 | 0 | — |
case-01 | fail→pass | 11,924 | 17,893 | +50% | 1 | 1 | 0% | 2,028 | 2,436 | +20% | 0 | 0 | — |
case-02 | fail→fail | 9,522 | 8,172 | -14% | 1 | 1 | 0% | 351 | 1,357 | +287% | 0 | 0 | — |
case-03 | fail→pass | 29,727 | 18,826 | -37% | 1 | 1 | 0% | 4,398 | 4,097 | -7% | 0 | 0 | — |
case-04 | pass→pass | 25,476 | 14,080 | -45% | 1 | 1 | 0% | 1,216 | 1,788 | +47% | 0 | 0 | — |
case-05 | fail→pass | 11,080 | 15,896 | +43% | 1 | 1 | 0% | 237 | 3,564 | +1404% | 0 | 0 | — |
case-06 | fail→pass | 7,307 | 8,033 | +10% | 1 | 1 | 0% | 554 | 2,193 | +296% | 0 | 0 | — |
case-07 | fail→pass | 8,037 | 17,174 | +114% | 1 | 1 | 0% | 1,259 | 3,266 | +159% | 0 | 0 | — |
case-08 | fail→pass | 15,686 | 10,710 | -32% | 1 | 1 | 0% | 2,513 | 2,633 | +5% | 0 | 0 | — |
case-09 | fail→pass | 10,518 | 13,340 | +27% | 1 | 1 | 0% | 1,571 | 3,079 | +96% | 0 | 0 | — |
case-10 | pass→pass | 24,058 | 17,773 | -26% | 1 | 1 | 0% | 4,224 | 2,816 | -33% | 0 | 0 | — |
case-13 | fail→pass | 9,841 | 3,488 | -65% | 1 | 1 | 0% | 1,358 | 1,279 | -6% | 0 | 0 | — |
case-14 | pass→fail | 16,233 | 14,161 | -13% | 1 | 1 | 0% | 2,426 | 3,284 | +35% | 0 | 0 | — |
case-15 | fail→pass | 12,943 | 6,830 | -47% | 1 | 1 | 0% | 1,626 | 1,446 | -11% | 0 | 0 | — |
case-16 | pass→pass | 8,001 | 4,528 | -43% | 1 | 1 | 0% | 1,587 | 1,586 | -0% | 0 | 0 | — |
case-17 | fail→pass | 11,716 | 10,943 | -7% | 1 | 1 | 0% | 1,704 | 3,159 | +85% | 0 | 0 | — |
case-18 | fail→pass | 18,051 | 16,802 | -7% | 1 | 1 | 0% | 2,679 | 3,659 | +37% | 0 | 0 | — |
case-19 | fail→pass | 9,929 | 2,926 | -71% | 1 | 1 | 0% | 1,529 | 1,374 | -10% | 0 | 0 | — |
case-20 | fail→pass | 14,109 | 11,694 | -17% | 1 | 1 | 0% | 2,151 | 2,647 | +23% | 0 | 0 | — |
case-21 | fail→pass | 14,895 | 5,789 | -61% | 1 | 1 | 0% | 1,124 | 1,875 | +67% | 0 | 0 | — |
case-22 | fail→pass | 11,376 | 12,915 | +14% | 1 | 1 | 0% | 1,533 | 2,624 | +71% | 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 19 counted toward the lift figure. The other 3 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 +73 percentage points is the difference between those two pass rates over the 19 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.