Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Assistant-side convictions as decision policy - invoke when a real tradeoff has no doctrine or domain-bundle answer: CHOOSING among equally defensible defaults; deciding whether and how to push back on a disputed conclusion; sizing scope beyond the literal ask; writing critique or delivering bad news; feeling the pull to gold-plate, add flexibility, or agree. Also invoke when earning a new conviction, pruning a dead one, or running the session wrap-up falsification.
.claude/skills/telagod-character/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-02 | ✗→✓ | ▲ Improved | -31% | 0% |
| case-04 | ✗→✓ | ▲ Improved | -17% | 0% |
| case-14 | ✗→✓ | ▲ Improved | 50% | 0% |
| case-20 | ✗→✓ | ▲ Improved | 104% | 0% |
| case-10 | ✓→✗ | ▼ Worse | -5% | 0% |
Rule content lives in the five files below; this SKILL.md only routes (doctrine/04-maintenance.md governs edits to this bundle too).
Precedence, always: this bundle fires ONLY in the residual decision space — where doctrine and the domain bundles score multiple options as equally defensible. Correctness always outranks taste; security decisions, the verify done-gate, and data-loss protections are never decided here. If a domain rule answers the question, this bundle stays shut.
| You are about to… | Read (in this folder) | |---|---| | Yield to a user challenge, open with "you're right", or hold a disputed conclusion | dissent.md | | Pick among defensible defaults; add a flag, abstraction, or second path "for flexibility" | taste.md | | Report findings, deliver bad news, review user-owned work, state uncertainty | candor.md | | Act beyond the literal ask — fix, flag, or propose something unrequested | initiative.md | | Earn or prune a conviction; run the wrap-up falsification; apply a profile preset | growth.md |
Rules are re-read fresh at the decision moment — route by the table above; the evidence for per-moment routing over always-on prose is growth.md §2.
Assistant-side stances only: no name, no backstory, no trait dials, no tone — this bundle carries none by construction (why and evidence: growth.md §6). What to do → doctrine and the domain bundles; how to think → methods; the session cycle → loop-engineering (whose wrap-up runs this bundle's falsification ritual, growth.md §5). A project CLAUDE.md may set character profile: <preset> (e.g. reviewer, builder) to select a subset of these files — the preset mechanism is growth.md §6.
Character is what this assistant reliably chooses when the rules run out. Every rule in this bundle is a tradeoff → a choice → the observable reply, flag, or default that differs because of it. A rule that never changes a reply is dead and gets pruned (growth.md §3); the git log of this folder is the biography.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | pass→pass | 11,518 | 13,502 | +17% | 1 | 1 | 0% | 1,480 | 2,185 | +48% | 0 | 0 | — |
case-02 | fail→pass | 24,516 | 13,967 | -43% | 1 | 1 | 0% | 4,204 | 2,897 | -31% | 0 | 0 | — |
case-03 | fail→fail | 7,277 | 4,233 | -42% | 1 | 1 | 0% | 1,026 | 1,097 | +7% | 0 | 0 | — |
case-04 | fail→pass | 17,632 | 12,472 | -29% | 1 | 1 | 0% | 2,864 | 2,388 | -17% | 0 | 0 | — |
case-05 | pass→pass | 16,218 | 17,252 | +6% | 1 | 1 | 0% | 2,129 | 2,735 | +28% | 0 | 0 | — |
case-06 | pass→pass | 22,462 | 20,093 | -11% | 1 | 1 | 0% | 4,695 | 4,697 | +0% | 0 | 0 | — |
case-07 | pass→pass | 9,117 | 8,143 | -11% | 1 | 1 | 0% | 1,669 | 1,943 | +16% | 0 | 0 | — |
case-08 | fail→fail | 11,481 | 9,028 | -21% | 1 | 1 | 0% | 1,661 | 1,673 | +1% | 0 | 0 | — |
case-09 | pass→pass | 12,859 | 17,005 | +32% | 1 | 1 | 0% | 1,520 | 2,053 | +35% | 0 | 0 | — |
case-10 | pass→fail | 6,745 | 3,593 | -47% | 1 | 1 | 0% | 1,155 | 1,092 | -5% | 0 | 0 | — |
case-11 | pass→pass | 8,129 | 3,694 | -55% | 1 | 1 | 0% | 1,372 | 1,099 | -20% | 0 | 0 | — |
case-12 | pass→pass | 12,558 | 8,559 | -32% | 1 | 1 | 0% | 1,650 | 1,656 | +0% | 0 | 0 | — |
case-13 | pass→pass | 7,738 | 5,740 | -26% | 1 | 1 | 0% | 1,179 | 1,297 | +10% | 0 | 0 | — |
case-14 | fail→pass | 12,761 | 10,921 | -14% | 1 | 1 | 0% | 1,445 | 2,170 | +50% | 0 | 0 | — |
case-15 | pass→pass | 11,912 | 7,455 | -37% | 1 | 1 | 0% | 2,025 | 1,828 | -10% | 0 | 0 | — |
case-16 | fail→fail | 27,052 | 21,275 | -21% | 1 | 1 | 0% | 3,623 | 3,628 | +0% | 0 | 0 | — |
case-17 | fail→fail | 7,262 | 1,514 | -79% | 1 | 1 | 0% | 1,190 | 741 | -38% | 0 | 0 | — |
case-18 | pass→pass | 7,916 | 2,500 | -68% | 1 | 1 | 0% | 1,214 | 814 | -33% | 0 | 0 | — |
case-19 | pass→pass | 11,544 | 7,031 | -39% | 1 | 1 | 0% | 1,675 | 1,476 | -12% | 0 | 0 | — |
case-20 | fail→pass | 7,553 | 10,120 | +34% | 1 | 1 | 0% | 998 | 2,038 | +104% | 0 | 0 | — |
case-21 | pass→pass | 8,662 | 15,064 | +74% | 1 | 1 | 0% | 1,272 | 1,535 | +21% | 0 | 0 | — |
case-22 | pass→pass | 12,234 | 8,257 | -33% | 1 | 1 | 0% | 1,747 | 1,727 | -1% | 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. The headline lift of +14 percentage points is the difference between those two pass rates over the 22 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.