skill-conductor

Skill Conductor

A skill that creates, evaluates, and improves other skills. Meta-level.

Architecture-first skill lifecycle: design → build → test → evaluate → package.

Most skill tools jump straight to "write SKILL.md." Conductor makes you choose the architecture first — because rewriting a wrong pattern costs more than writing it right.

Install

# skills.sh — installs into ~/.claude/skills
npx skills add smixs/skill-conductor

# Claude Code plugin
/plugin marketplace add smixs/skill-conductor
/plugin install skill-conductor@smixs

• Principle #10: Match the form to the failure — classify the baseline failure before writing a rule; prohibitions bulletproof discipline failures but measurably backfire on shaping failures (obra/superpowers wording tests + Guardrails polarity data). Plus: no nuance clauses, exemption clauses don't scope.
• Critique-before-verdict judges — all three eval agents (grader, comparator, bineval) now write the detailed evidence critique BEFORE committing to the 1/0 verdict, with a borderline few-shot example in each (Hamel Husain's judge methodology).
• Threshold-blind judging — the BinEval judge no longer computes the overall score or the GATE; the orchestrator aggregates. A judge that knows the bar is biased toward it.
• Automatic cross-family judge calibration — a second judge from a different model family answers the same bank; stable disagreement flags a badly worded question, not a dispute. Self-preference-bias guard on final acceptance.
• Variance discipline — improvements on non-critical questions count only when they reproduce in 2 consecutive runs; the 3-iteration cap now carries its STICK rationale.
• references/pressure-testing.md — micro-test protocol (no-guidance control, 5+ reps, variance as a metric) + pressure scenarios for discipline skills (7 pressure types, forced A/B/C choice, rationalization tables).
• Pushy description formula — [What] + Use when [4-5 phrasings] + "even if they don't explicitly say '<canonical term>'" + Do NOT use for [...], deduped to a single canonical home in Principle #2.
• Question bank v1.1 — 5 new questions: pushy triggers, nuance clauses, directive reference loading, time-rot language, redundant-content (E:A:R).
• Self-hosted proof — this release was produced by Conductor evaluating and improving itself: 3 gated iterations, dual-family judges (Claude + GPT via codex), all critical questions passing.

• Held-out gate for body edits — Mode 2 IMPROVE now splits evals into train/held-out (scripts/split_evals.py, deterministic, stratified by optional evals[].category). Lessons and edits come from TRAIN only; a candidate is accepted iff no held-out assertion regresses (flip-confirmation re-run for noise), train pass-rate strictly improves, and no new critical failure. Methodology borrowed from microsoft/SkillOpt.
• Edit budget — at most 3 atomic edits per iteration (one edit = one lesson, labeled); no wholesale rewrites, so gate rejections stay attributable.
• Case transitions — assertion-level diff parent→candidate (improved / regressed / persistent-fail / stable-success) recorded as an additive transitions block in benchmark.json.
• Refactor — split_eval_set generalized into utils.split_evals(stratify_key=...); run_loop.py (Mode 5 OPTIMIZE) delegates to it, split unchanged bit-for-bit.

• BinEval evaluation — replaces the old 5-axis 1-10 scoring with atomic binary yes/no questions across 5 dimensions (Discovery, Clarity, Structure, Robustness, Completeness). Each answer carries grounding evidence; the pass criterion is a gate on critical questions, not an opaque number. Adapted from "Ask, Don't Judge" (arXiv 2606.27226).
• Deterministic + LLM split — eval_skill.py --json emits structural checks as binary question records; an evaluator agent answers the judgment questions with evidence and a self-update loop feeds failing questions back into edits.
• 9 authoring principles — a universal canon (pre-flight, no-process-in-description, MOC, fresh-practitioner author, TWI "why", blind-agent test, inline checklists, one-term-per-concept, cut-the-fat) in references/sop-practices.md, applied to every skill.
• Dual-channel install — one repo, one source of truth, installable via skills.sh and the Claude Code plugin marketplace.

• references/sop-practices.md — 80 years of Standard Operating Procedure wisdom applied to skill authoring. Inline checklists at risk-points, pre-flight checks, programmatic validation, exception handling patterns. Use for procedural skills (client intake, onboarding, reporting, escalation)
• scripts/test_smoke.py — fast safety net for skill-conductor scripts themselves. Verifies critical scripts execute on known-good skills, fail on known-bad, produce expected output shapes. Run: uv run scripts/test_smoke.py
• Updated eval agents (grader, comparator, analyzer) with refined rubrics
• Improved package_skill.py, eval_skill.py, and schema validation
• Updated patterns.md and schemas.md with tighter definitions

Anthropic updated their skill-creator with serious eval infrastructure. We took the best of it:

From Anthropic's skill-creator:

• 3 specialized agents: grader (assertion checking + claim extraction), comparator (blind A/B testing), analyzer (post-hoc root cause analysis)
• Parallel eval execution with isolated contexts (no cross-contamination)
• Automated description optimization with train/test split (60/40)
• Benchmark tracking: pass rate, tokens, time with variance analysis
• HTML eval viewer with qualitative + quantitative tabs

What Conductor adds on top:

• Architecture before code. 5 patterns (Sequential, Iterative, Context-Aware, Domain Intelligence, Multi-MCP) with selection criteria. Pick wrong = rewrite everything later
• Degrees of freedom. Low (deterministic scripts) → Medium (pseudocode) → High (free text). Match freedom to risk tolerance
• TDD RED before writing. Verify the agent fails WITHOUT the skill first. If it already handles the task — you don't need a skill
• Quality scoring with a gate (now BinEval — see v1.0.0). Numbers and evidence, not a "vibe check"
• Skill categorization. Capability uplift (teaching something new) vs Encoded preference (sequencing known abilities). Different skills need different testing strategies

Synthesized from

1. Anthropic Skill Creator — eval infrastructure, grader/comparator/analyzer agents, benchmark pipeline
2. The Complete Guide to Building Skills for Claude — architecture patterns, success metrics
3. Superpowers / writing-skills by Jesse Vincent — TDD approach, the "description trap" discovery, match-the-form-to-the-failure, the micro-test protocol, pressure scenarios and rationalization tables
4. Skills Best Practices by Minko Gechev — three-stage LLM validation, eval methodology
5. hamelsmu/evals-skills by Hamel Husain — critique-before-verdict judge outputs, borderline few-shot examples, judge calibration discipline
6. grafana/skills — skill-authoring — the pushy description pattern in production, judge score-variance discipline ("three consecutive local passes before shipping")
7. softaworks/agent-toolkit — skill-judge — the Expert/Activation/Redundant knowledge-delta taxonomy, directive loading triggers, the freedom-consequence test
8. neolabhq/context-engineering-kit — threshold-blind judges (never tell the judge the bar)
9. trailofbits/skills — skill-improver — the stop-hook pattern for unattended improvement loops (referenced, not implemented)

Methodology foundations

• Standard Operating Procedures: A Writing Guide — Richard Stup, Penn State Extension. Format selection, hierarchical vs. flowchart procedures.
• Procedure Writing: Principles and Practices — Wieringa, Moore & Barnes (Battelle Press, 1998). Imperative steps, removing modal weasel-words.
• Toyota TWI (Training Within Industry) — the "Job Instruction" method: step → key point → why; the 5 Whys root-cause practice (Job Methods manual, The Roots of Lean).
• McDonald's Operations Manual — the canonical 600+ page SOP system; checklists at the point of use. The manual itself is proprietary; it is documented in John F. Love's McDonald's: Behind the Arches.
• Ask, Don't Judge: Binary Questions for Interpretable LLM Evaluation and Self-Improvement — the BinEval method behind Conductor's evaluation.
• Guardrails Beat Guidance (5000+ Claude Code runs on SWE-bench) — rule polarity: helpful rules are negative constraints, harmful ones positive directives → Principle #10.
• TICK: Generated Checklists Improve LLM Evaluation and Generation — checklist as spec + eval + feedback; refinement plateaus and degrades past 3–4 iterations → the 3-iteration cap.
• CheckEval and Prosa — binary decomposition makes judges reproducible across model families → cross-family judge calibration.
• Self-Preference Bias in Rubric-Based Evaluation — judges favor their own family even on binary rubrics → the out-of-family acceptance rule.
• LLMs Cannot Self-Correct Reasoning Yet and BIG-Bench Mistake — self-correction needs an external gate; models fix errors well only when an external checker locates them → the gated self-update loop.