Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →The operating doctrine, self-contained in this folder. Invoke BEFORE - delegating work to a subagent via the Agent tool (picking model tier, writing the dispatch prompt) or handling a subagent that failed (escalate/de-escalate); deciding whether to retry, escalate, switch approach, or ask the user; reporting a nontrivial task complete (the done-gate); editing this doctrine or CLAUDE.md; recording a lesson about the harness.
.claude/skills/telagod-doctrine/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-01 | ✗→✓ | ▲ Improved | 2% | 0% |
| case-03 | ✗→✓ | ▲ Improved | 149% | 0% |
| case-08 | ✗→✓ | ▲ Improved | 45% | 0% |
| case-09 | ✗→✓ | ▲ Improved | -19% | 0% |
| case-10 | ✗→✓ | ▲ Improved | -4% | 0% |
All referenced files live in this skill's folder; read them with relative paths from here. The three always-on laws live in the user's ~/.claude/CLAUDE.md router; this bundle carries the depth. Content lives in the numbered files — this SKILL.md only routes; never duplicate rule text into it.
| Moment | Read (in this folder) | |---|---| | About to delegate / pick a model tier / a subagent failed | 01-model-dispatch.md, then fill a template from 03-delegation-templates.md | | Retry? escalate? done? ask the user? direction feels wrong? | 02-judgment.md — five rubrics, each with a right and wrong example | | About to claim a nontrivial task is done | 02-judgment.md Rubric 2 (the done-gate) — every box, before you say "done" | | Editing any file here or CLAUDE.md; recording a lesson | 04-maintenance.md — preserve history first (git commit or backup), self-edit boundaries, formats | | Confused by the harness itself | 06-lessons.md first (it may already be paid for), then 00-diagnosis.md | | First session in a new environment, or sharing this bundle | INSTALL.md | | Why these rules exist; environment-specific warnings | 00-diagnosis.md, 05-letter.md |
01).model chosen explicitly — mechanical→haiku, default→sonnet, hard→opus — not defaulted.Would raw output ≫ the answer? → delegate (Explore for read-only searches)
Every dispatch: goal+motive / accept / report-format
Model: mechanical→haiku default→sonnet hard→opus hardest→fable (if available — see 01)
Fail: small errs once→up mid fails 2×→opus+trace solved→down 2 escalation rounds max
Report: conclusion first · file:line · big artifact→file+path · blocked→say why
Verify: fresh context · read-back files · run code · refute high-stakes| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→pass | 10,991 | 7,365 | -33% | 1 | 1 | 0% | 1,620 | 1,654 | +2% | 0 | 0 | — |
case-02 | fail→fail | 16,542 | 14,799 | -11% | 1 | 1 | 0% | 2,539 | 2,955 | +16% | 0 | 0 | — |
case-03 | fail→pass | 35,227 | 18,481 | -48% | 1 | 1 | 0% | 1,203 | 2,996 | +149% | 0 | 0 | — |
case-04 | pass→pass | 8,115 | 5,332 | -34% | 1 | 1 | 0% | 1,217 | 1,434 | +18% | 0 | 0 | — |
case-05 | pass→pass | 7,335 | 2,923 | -60% | 1 | 1 | 0% | 1,102 | 947 | -14% | 0 | 0 | — |
case-06 | pass→pass | 14,322 | 10,243 | -28% | 1 | 1 | 0% | 2,472 | 2,369 | -4% | 0 | 0 | — |
case-07 | pass→fail | 6,176 | 3,126 | -49% | 1 | 1 | 0% | 923 | 1,089 | +18% | 0 | 0 | — |
case-08 | fail→pass | 6,371 | 4,246 | -33% | 1 | 1 | 0% | 833 | 1,206 | +45% | 0 | 0 | — |
case-09 | fail→pass | 15,988 | 8,395 | -47% | 1 | 1 | 0% | 2,474 | 2,000 | -19% | 0 | 0 | — |
case-10 | fail→pass | 9,685 | 5,102 | -47% | 1 | 1 | 0% | 1,427 | 1,365 | -4% | 0 | 0 | — |
case-11 | pass→pass | 8,625 | 3,955 | -54% | 1 | 1 | 0% | 1,340 | 1,232 | -8% | 0 | 0 | — |
case-12 | fail→pass | 13,442 | 8,739 | -35% | 1 | 1 | 0% | 1,991 | 1,917 | -4% | 0 | 0 | — |
case-13 | fail→pass | 10,525 | 3,252 | -69% | 1 | 1 | 0% | 1,656 | 1,089 | -34% | 0 | 0 | — |
case-14 | pass→fail | 11,018 | 5,007 | -55% | 1 | 1 | 0% | 1,582 | 1,368 | -14% | 0 | 0 | — |
case-15 | fail→fail | 10,795 | 6,308 | -42% | 1 | 1 | 0% | 1,789 | 1,555 | -13% | 0 | 0 | — |
case-16 | pass→pass | 10,658 | 3,617 | -66% | 1 | 1 | 0% | 1,586 | 1,206 | -24% | 0 | 0 | — |
case-17 | fail→pass | 7,227 | 4,728 | -35% | 1 | 1 | 0% | 1,134 | 1,231 | +9% | 0 | 0 | — |
case-18 | pass→pass | 10,029 | 3,580 | -64% | 1 | 1 | 0% | 1,398 | 1,126 | -19% | 0 | 0 | — |
case-19 | fail→pass | 5,989 | 9,672 | +61% | 1 | 1 | 0% | 999 | 938 | -6% | 0 | 0 | — |
case-20 | pass→pass | 3,021 | 3,246 | +7% | 1 | 1 | 0% | 481 | 1,058 | +120% | 0 | 0 | — |
case-21 | pass→pass | 4,213 | 3,432 | -19% | 1 | 1 | 0% | 519 | 1,108 | +113% | 0 | 0 | — |
case-22 | pass→pass | 2,154 | 3,660 | +70% | 1 | 1 | 0% | 295 | 1,086 | +268% | 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 +32 percentage points is the difference between those two pass rates over the 21 comparable cases. 2 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.
Other measured skills in the registry, with their headline benchmark lift.