[Jibbscript]

Summary

Re-tunes the 12-skill pack for Claude Fable 5.1. The skills were written in the Opus 4.5 era as step scripts with copy-paste checklists; Fable 5.1 guidance says prompts written that way are too prescriptive and reduce output quality. This branch replaces choreography with goals, constraints, and reasons, and adds the Fable-era capabilities the pack never asked for.

Every change was produced by a tournament loop: each hypothesis was measured before execution, reviewed by an architect and a critic, executed in an isolated worktree, and re-benchmarked after merge. An independent verifier read the final files end to end.

Benchmark (eval/fable_prompt_bench.py, deterministic, stdlib-only, sealed during the loop): 59.37 → 95.25.

Dimension	Before	After
cruft (pressure language, scaffolds, fossils)	99.7	100
de_prescription (Step-N scripts, checklists)	53.3	100
context_richness (reasons, audience, prose)	4.2	~70
fable_utilization (delegation, grounding, boundaries, judgment)	3.1	100
trigger_quality (frontmatter descriptions)	100	100

What changed in every skill

• Fable capabilities — a ## Working <domain> section per skill, in that skill's own vocabulary: fan out independent research to parallel sub-agents, label every figure sourced / estimated / unverified, calibrate depth, name the downstream consumer, invite non-obvious judgment, lead the summary with the outcome, self-check before hand-off.
• De-prescription — ### Step N scripts and Workflow Checklists replaced by ## Outcome, ## What <artifact> Must Cover, ## How to Work, ## Constraints. Every scale, field list, source list, query template, edge case, and calibration note is preserved; numbered order survives only where it genuinely matters.
• Context — ## Who reads <artifact> and ## What a strong <artifact> looks like, with readership accurate to the producer/consumer table in contracts/; behavioral bullet walls reflowed into prose carrying the reason.
• Structure — numeric scales rendered as tables, literal search-query templates fenced.
• Orchestrator — tells the model the user is not watching and reversible steps proceed, asks for outcome-first final messages that re-ground a reader who saw none of the work, and ends the run either with all eleven artifacts or an explicit statement of what was skipped and why.
• Docs — docs/CONTRIBUTING.md template and every AGENTS.md section-order line updated to match; four residual cruft lines cleared in the baseline runner prompt and three reference files.

Not changed

Frontmatter (names, descriptions, allowed-tools), contracts/, eval/fixtures/, LICENSE. The eval runners still list all five fixtures.

Test plan

• python3 eval/fable_prompt_bench.py → 95.25, zero integrity failures, zero cross-skill boilerplate
• Frontmatter diff vs masta empty; sealed paths byte-identical
• python3 eval/runners/run_skillpack.py --list prints five fixtures
• Independent verifier read risk-assessment, gtm-channels, and the orchestrator end to end (single voice, no contradictions, no lost domain knowledge)
• Real-model check: fixture t1-energy-audit run three ways (new skills, old masta skills, no-skills baseline), each with live web research, graded by three rubric judges per run plus blind pairwise comparison in both orderings — see below
• CI: prompt-fitness workflow added and passing on this PR

Real-model A/B on fixture t1 (Home Energy Leak Mapper)

Run	Judge mean (0-100)	Judge range	Contract violations
new skills	88.2	85.5–92.0	5–10
old masta skills	87.7	81.0–91.5	5–8
no skills	79.5	77.5–82.0	8–26

Blind pairwise, position-corrected: new beats old in both orderings (slight, then clear) on usefulness and voice; correctness and citations are a wash. The two pack-vs-baseline pairs flipped with position and are not robust. Both pack runs beat the baseline mainly on contract compliance.

What the new pack did that the old one did not: corrected the fixture's false premise (no iPhone has a thermal sensor) and carried it through every artifact; wrote falsifiable next steps with kill thresholds; reported the unadjusted composite alongside the adjusted range that would flip the verdict. What the old pack did better: fieldwork depth (resolved the iOS UVC question, found one more competitor, surfaced California's referral-fee restriction).

Follow-ups surfaced by the judges, not addressed in this PR: both packs made the same derived-share arithmetic error in market sizing (recompute shares from the cited buckets); both missed the most direct competitor (search funded startups in the re-premised space); artifacts run 3–6× the reference length; the repo's Python scorer only measures artifact completion and cannot separate these runs.

The full report is eval/results/t1_ab_report.md on the machine that ran it (run output is gitignored by design).

🤖 Generated with Claude Code