Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →The engineering loop - select, advance, distill, deliver - for running a work session on a project. Invoke when STARTING a session or picking the next thing to work on; sizing a unit of work or matching it to your model tier; running long/autonomously and deciding pace or when to stop; deciding where a learning should live (lesson vs memory vs project-level skill); or wrapping up a unit/session (commit, report, handoff note).
.claude/skills/telagod-loop-engineering/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-01 | ✗→✓ | ▲ Improved | 0% | 0% |
| case-02 | ✗→✓ | ▲ Improved | 0% | 0% |
| case-04 | ✗→✓ | ▲ Improved | -14% | 0% |
| case-06 | ✗→✓ | ▲ Improved | -14% | 0% |
| case-07 | ✗→✓ | ▲ Improved | -12% | 0% |
Rule content lives in the four files below; this SKILL.md only routes (doctrine/04-maintenance.md governs edits to this bundle too).
| You are about to… | Read (in this folder) | |---|---| | Start a session; choose the next unit; size it; decide who/which tier runs it | select.md | | Push a unit forward; wonder if you're stalled; pace an unattended run | advance.md | | Decide where a learning goes — lesson, memory, project-level skill, or nowhere | distill.md | | Claim a unit done; commit; report; end the session | deliver.md |
A full pass is select.md → advance.md → distill.md → deliver.md, once per unit of work; re-enter at select.md §5 after each landing.
This bundle is the outer cycle that sequences the others across a session and a project. Whether/how to delegate and judge (escalate, done-gate, ask-user) is doctrine; how to work a problem once inside a unit (investigate, design, execute, verify, write) is methods; domain judgment is the matching domain bundle. When this bundle and doctrine both fire, doctrine first — its three non-negotiables apply inside every phase of the loop.
A project advances as a sequence of landed loops, and compounds only if each loop distills. Progress is landed deliverables, not activity (advance.md §3); the next session's speed is this session's runway (deliver.md §4); and the institution grows one filed learning at a time (distill.md §1) — that is how weak sessions get strong.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→pass | 10,641 | 8,114 | -24% | 1 | 1 | 0% | 1,632 | 1,640 | +0% | 0 | 0 | — |
case-02 | fail→pass | 12,562 | 10,078 | -20% | 1 | 1 | 0% | 1,897 | 1,900 | +0% | 0 | 0 | — |
case-03 | fail→fail | 14,661 | 9,828 | -33% | 1 | 1 | 0% | 2,269 | 1,903 | -16% | 0 | 0 | — |
case-04 | fail→pass | 13,817 | 8,851 | -36% | 1 | 1 | 0% | 2,166 | 1,853 | -14% | 0 | 0 | — |
case-05 | pass→pass | 9,788 | 5,512 | -44% | 1 | 1 | 0% | 1,396 | 1,255 | -10% | 0 | 0 | — |
case-06 | fail→pass | 5,753 | 2,305 | -60% | 1 | 1 | 0% | 918 | 792 | -14% | 0 | 0 | — |
case-07 | fail→pass | 5,453 | 1,940 | -64% | 1 | 1 | 0% | 758 | 664 | -12% | 0 | 0 | — |
case-08 | pass→pass | 15,432 | 11,448 | -26% | 1 | 1 | 0% | 2,342 | 2,272 | -3% | 0 | 0 | — |
case-09 | fail→fail | 13,961 | 6,576 | -53% | 1 | 1 | 0% | 2,040 | 1,347 | -34% | 0 | 0 | — |
case-10 | fail→pass | 15,682 | 2,288 | -85% | 1 | 1 | 0% | 891 | 808 | -9% | 0 | 0 | — |
case-11 | fail→pass | 11,109 | 2,069 | -81% | 1 | 1 | 0% | 1,625 | 774 | -52% | 0 | 0 | — |
case-12 | pass→pass | 13,846 | 6,610 | -52% | 1 | 1 | 0% | 1,838 | 1,381 | -25% | 0 | 0 | — |
case-13 | pass→pass | 14,682 | 9,570 | -35% | 1 | 1 | 0% | 2,101 | 1,709 | -19% | 0 | 0 | — |
case-14 | fail→pass | 15,159 | 6,705 | -56% | 1 | 1 | 0% | 2,120 | 1,472 | -31% | 0 | 0 | — |
case-15 | fail→pass | 8,467 | 2,431 | -71% | 1 | 1 | 0% | 1,256 | 826 | -34% | 0 | 0 | — |
case-16 | fail→pass | 11,347 | 4,333 | -62% | 1 | 1 | 0% | 1,502 | 1,169 | -22% | 0 | 0 | — |
case-17 | fail→pass | 11,064 | 2,989 | -73% | 1 | 1 | 0% | 1,748 | 905 | -48% | 0 | 0 | — |
case-18 | pass→pass | 11,947 | 1,862 | -84% | 1 | 1 | 0% | 1,496 | 733 | -51% | 0 | 0 | — |
case-19 | fail→pass | 27,231 | 2,613 | -90% | 1 | 1 | 0% | 1,654 | 761 | -54% | 0 | 0 | — |
case-20 | pass→pass | 9,625 | 7,251 | -25% | 1 | 1 | 0% | 1,663 | 1,630 | -2% | 0 | 0 | — |
case-21 | pass→pass | 13,185 | 9,364 | -29% | 1 | 1 | 0% | 1,992 | 1,958 | -2% | 0 | 0 | — |
case-22 | pass→pass | 9,402 | 7,428 | -21% | 1 | 1 | 0% | 1,743 | 1,855 | +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 21 counted toward the lift figure. The other 1 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 +55 percentage points is the difference between those two pass rates over the 21 comparable cases.
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.