Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Have an interactive discussion about a topic, approach, or feature. Researches the codebase as needed, talks through options, and updates ./tmp/context.md with decisions. Use when you want to think through an approach before planning.
.claude/skills/dcouple-discussion/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-06 | ✗→✓ | ▲ Improved | 24% | 0% |
| case-04 | ✓→✗ | ▼ Worse | -60% | 0% |
| case-05 | ✓→✗ | ▼ Worse | 149% | 0% |
| case-08 | ✓→✗ | ▼ Worse | -63% | 0% |
| case-09 | ✓→✗ | ▼ Worse | -50% | 0% |
Talk the topic through with the user, back and forth, until the ideas, tradeoffs, and goal are clear enough to plan or build.
Read code and research freely. Your only output is conversation. Never edit, create, or delete project files, and never propose diffs or patches to apply.
Use the explorer subagent or codebase-explorer for code, and research-web for external libraries or approaches, including mid-conversation when new questions come up.
rewrite-simply skill to every discussion response before you sendit.
tradeoffs.
and defer to the user's judgment.
Classify every fork question before you raise it. If running something can answer it (behavior, timing, output, performance, layout), build the cheapest throwaway probe outside the project tree, run it, and present the result with a recommendation. Save questions for product or preference calls that no experiment can settle. A probe usually answers faster than a person, and it gives them a result to react to.
Recap in the conversation:
Mark which points are the user's intent and which are your assumptions. Then recommend the next step, usually create-ticket when the work is ready to delegate, options when more than one approach is live, brief for a write-up, investigate for a bug, or implementation.
Suggested next steps:
- `/create-ticket [description]` - Capture the work for delegation
- `/options [question]` - Lay out the approaches and trade-offs
- `/discussion [follow-up]` - Continue exploring a specific aspect
- `/research-web [topic]` - Look up external documentationTopic to discuss: $ARGUMENTS
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 23,580 | 6,075 | -74% | 1 | 1 | 0% | 2,351 | 855 | -64% | 0 | 0 | — |
case-02 | fail→fail | 10,815 | 29,443 | +172% | 1 | 1 | 0% | 299 | 852 | +185% | 0 | 0 | — |
case-03 | fail→fail | 8,341 | 6,601 | -21% | 1 | 1 | 0% | 322 | 919 | +185% | 0 | 0 | — |
case-04 | pass→fail | 14,365 | 15,408 | +7% | 1 | 1 | 0% | 2,740 | 1,091 | -60% | 0 | 0 | — |
case-05 | pass→fail | 6,414 | 17,838 | +178% | 1 | 1 | 0% | 959 | 2,391 | +149% | 0 | 0 | — |
case-06 | fail→pass | 15,645 | 14,519 | -7% | 1 | 1 | 0% | 1,475 | 1,827 | +24% | 0 | 0 | — |
case-07 | fail→fail | 18,953 | 42,388 | +124% | 1 | 1 | 0% | 843 | 1,902 | +126% | 0 | 0 | — |
case-08 | pass→fail | 17,952 | 6,914 | -61% | 1 | 1 | 0% | 2,206 | 814 | -63% | 0 | 0 | — |
case-09 | pass→fail | 13,005 | 10,650 | -18% | 1 | 1 | 0% | 2,218 | 1,113 | -50% | 0 | 0 | — |
case-10 | pass→fail | 15,160 | 8,115 | -46% | 1 | 1 | 0% | 2,027 | 845 | -58% | 0 | 0 | — |
case-11 | pass→fail | 13,561 | 6,977 | -49% | 1 | 1 | 0% | 1,981 | 848 | -57% | 0 | 0 | — |
case-12 | fail→fail | 35,766 | 12,439 | -65% | 1 | 1 | 0% | 2,124 | 1,819 | -14% | 0 | 0 | — |
case-13 | fail→fail | 7,848 | 35,385 | +351% | 1 | 1 | 0% | 319 | 1,399 | +339% | 0 | 0 | — |
case-14 | fail→fail | 8,364 | 8,273 | -1% | 1 | 1 | 0% | 1,016 | 838 | -18% | 0 | 0 | — |
case-15 | pass→fail | 15,844 | 7,123 | -55% | 1 | 1 | 0% | 2,158 | 790 | -63% | 0 | 0 | — |
case-16 | fail→fail | 16,069 | 14,889 | -7% | 1 | 1 | 0% | 2,714 | 2,045 | -25% | 0 | 0 | — |
case-17 | pass→fail | 12,479 | 6,750 | -46% | 1 | 1 | 0% | 2,076 | 916 | -56% | 0 | 0 | — |
case-18 | pass→fail | 11,427 | 6,992 | -39% | 1 | 1 | 0% | 1,467 | 902 | -39% | 0 | 0 | — |
case-19 | fail→fail | 17,227 | 40,754 | +137% | 1 | 1 | 0% | 2,633 | 9,031 | +243% | 0 | 0 | — |
case-20 | pass→fail | 13,988 | 6,998 | -50% | 1 | 1 | 0% | 1,616 | 813 | -50% | 0 | 0 | — |
case-21 | fail→fail | 9,193 | 40,533 | +341% | 1 | 1 | 0% | 570 | 921 | +62% | 0 | 0 | — |
case-22 | pass→fail | 29,849 | 6,192 | -79% | 1 | 1 | 0% | 3,835 | 793 | -79% | 0 | 0 | — |
case-23 | pass→fail | 11,461 | 8,899 | -22% | 1 | 1 | 0% | 1,632 | 913 | -44% | 0 | 0 | — |
case-24 | pass→fail | 10,230 | 7,191 | -30% | 1 | 1 | 0% | 1,448 | 862 | -40% | 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. 24 cases were attempted, and 2 counted toward the lift figure. The other 22 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 -50 percentage points is the difference between those two pass rates over the 2 comparable cases. 17 cases got worse with the skill loaded, and they are 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.
| Model | Method | Date | Lift |
|---|---|---|---|
| gemini-3.6-flash | verified | 8/21/2026 | +14% |
Other measured skills in the registry, with their headline benchmark lift.