feat(skills): add capability-tuning skill + refresh Claude/Codex capability facts#1
Merged
Merged
Conversation
…bility facts Research-driven pass over the kit's prompt surfaces (3 deep-research agents: Karpathy's method, Claude Code mid-2026 config, Codex mid-2026 config): - new productivity/capability-tuning skill: max-capability settings.json / config.toml templates for both runtimes, instruction-file budgets, cross-model phrasing rules, Karpathy context-engineering layer (16/16) - skills/AGENTS.md: Codex catalog truncation corrected (even truncation, 2%-of-window-in-tokens budget; 8k chars is only the fallback) + new goals-over-step-lists rule for Fable 5 / Opus 4.8 - codex-agents-builder.md: effort set gains 'none', Claude max->xhigh mapping, sonnet maps to mainline gpt-5.4 (codex line absorbed at 5.4); codex-facts.sh guard now pins 'none' - code-review: reject for missing evidence, never for low confidence — recall protection for literal-instruction Opus models - capability-tuning listed in plugin/marketplace descriptions + READMEs https://claude.ai/code/session_01HQ2Qevpwxq4ECfutPuSkyX
…discovery Codex discovers .claude-plugin/plugin.json natively (DISCOVERABLE_PLUGIN_MANIFEST_PATHS, re-verified 2026-06-10 against openai/codex HEAD) — the mirror's value is the marketplace catalog, Codex-tailored interface + degradation surfacing, and version lockstep, not basic discovery. https://claude.ai/code/session_01HQ2Qevpwxq4ECfutPuSkyX
Evidence-gated fixes from a full engineering-category audit: - code-review + fix-workflow: retire stale '/security 3 parallel scanners' / '/refactor command' architecture refs (kit is sequential single-context); approval gates rephrased to the enforceable turn-ending form - dep-vuln-workflow: drop fabricated tokio 1->2 migration row (no tokio 2.x exists), fix axum 0.8 + reqwest 0.12 break details, replace nonexistent 'uv pip audit' with uv export | pip-audit; tag bare fences; add Not-for exclusion to description - design-tokenization: Tailwind v4 DOES auto-detect sources (gitignored / out-of-root paths are what @source covers); kill dead v4-beta docs link - react-component-patterns: useEffectEvent is stable since React 19.2 (was 'experimental, do not use'); trim description 502->445 chars - lint-no-suppressions: add BAD/GOOD fence, bare-suppression-scope gotchas, Rust #[expect] (1.81+); ESLint flat-config row (13 -> 16 score) - solid refs: '# type: ignore' removed from a GOOD example, assert_never is Python 3.11+ not PEP 661, dyn-compatibility note for native async fn traits, once_cell::Lazy -> std LazyLock - type-safety: #private is class-only syntax (factory closures are what's private), was claimed to work on plain objects - security: remove constraint-contradicting MAY-run-concurrently clause - human-docs-workflow: description 552 -> 488 chars All engineering skills now 16/16 except vendored make-interfaces (10, body frozen by policy). https://claude.ai/code/session_01HQ2Qevpwxq4ECfutPuSkyX
…I gate Productivity skills (evidence-gated, real content only): - plan-init + skill-agent-pipeline: descriptions trimmed under the 500-char CSO tier (15 -> 16 each) - write-skill: third constraint promoting the updated+content-hash bookkeeping rule (14 -> 16) - zoom-out: 7-module cap promoted to a constraint, situation -> output-form table, BAD/GOOD map example, fences tagged (9 -> 16) - caveman: persistence rule promoted to constraint, drop/keep rules table, BAD/GOOD labels (8 -> 12; stays sub-16 by design — it's a brevity skill) - plan-review agent: missing ## Output Format section (14 -> 15; both shipped agents now 15/15) Validator hardening (no floor loosened): - shellcheck -S warning is now a CI gate: ci.sh §3b (self-skips when not installed locally) + ci.yml guard job + scripts/AGENTS.md validator row - 6 shellcheck findings fixed: cd||exit in ci.sh + idempotency test, unused loop var in release.sh, xargs -> sed path derivation in tree/guard.sh, 2 documented SC2043 disables - slop scorer strips fenced blocks + code spans first — quoting a banned word (ban lists, BAD examples) is not prose slop - BSD date fallback for the freshness point (macOS parity) - UTF-8 locale forced so description tiers count chars, not bytes - dead extract_yaml_value removed from skills scorer - agents scorer credits explicit full claude-* model IDs Category totals: engineering 212 -> 218, productivity 189 -> 204, agents 29 -> 30. ci.sh fully green including the new gate. https://claude.ai/code/session_01HQ2Qevpwxq4ECfutPuSkyX
…ersionless examples The Codex marketplace policy values are documented now (installation: AVAILABLE/NOT_AVAILABLE/INSTALLED_BY_DEFAULT; authentication: ON_INSTALL/ON_USE) — drop the 'not documented yet' hedge. Worked examples use X.Y.Z placeholders instead of a stale hardcoded 0.3.0 (no version numbers in prose). https://claude.ai/code/session_01HQ2Qevpwxq4ECfutPuSkyX
…ng form Loop tick 1/12: deep-read skill-maintenance (clean) + multi-tool-bridge — its Step-3 split gate used the bypassable 'wait for confirmation' phrasing the kit itself documents as ignored by literal-instruction models. https://claude.ai/code/session_01HQ2Qevpwxq4ECfutPuSkyX
…o-author-scripts pattern Loop tick 2/12: context-tree's shipped body + 2 references named the docks author script tree/guard.sh in the bare form (no scripts/ prefix) the no-author-scripts guard couldn't see — broken the moment the skill runs in a consumer repo. Replaced with the self-contained pair check; the guard pattern now also catches the distinctive bare names (tree/guard.sh, content-hash.sh, transform-guard.sh, no-author-scripts.sh, codex-facts.sh, guard-spec.sh). plan-manager + plan-sidecar re-read: clean. https://claude.ai/code/session_01HQ2Qevpwxq4ECfutPuSkyX
Loop tick 3/12: deep-read scaffold + plan-init + all 9 skill-agent-pipeline phase references — all era-accurate; the only fix is scaffold's gate, which used the bypassable 'wait for confirmation' phrasing. Full shipped surface (every SKILL.md + reference + agent body) has now been read this session. https://claude.ai/code/session_01HQ2Qevpwxq4ECfutPuSkyX
Loop tick 4/12: read the final un-read references (write-skill data-preservation, lint pre-commit-hook, plan-init template + scaffold spec-schema spot-checks) — all clean except a legacy .eslintrc.* mention now framed as eslint.config.js-first. Hook payload shapes re-verified. https://claude.ai/code/session_01HQ2Qevpwxq4ECfutPuSkyX
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Research-driven pass over the kit's prompt surfaces (3 deep-research agents:
Karpathy's method, Claude Code mid-2026 config, Codex mid-2026 config):
config.toml templates for both runtimes, instruction-file budgets,
cross-model phrasing rules, Karpathy context-engineering layer (16/16)
2%-of-window-in-tokens budget; 8k chars is only the fallback) + new
goals-over-step-lists rule for Fable 5 / Opus 4.8
mapping, sonnet maps to mainline gpt-5.4 (codex line absorbed at 5.4);
codex-facts.sh guard now pins 'none'
recall protection for literal-instruction Opus models
https://claude.ai/code/session_01HQ2Qevpwxq4ECfutPuSkyX