[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M7AX7DNP0175CPV21F228J4B":3,"skill-same:01M7AX7DNP0175CPV21F228J4B":73},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":18,"publishedAt":19,"linkOnly":20,"markdownContent":21,"markdownBody":22,"frontMatter":23,"refName":24,"commitSha":25,"skillDirectory":26,"forksCount":14,"repositoryHtmlUrl":27,"repositoryHomepageUrl":6,"ownerHtmlUrl":28,"ownerBlogUrl":6,"ownerDescription":29,"updatedAt":30,"install":31,"resources":44},"01M7AX7DNP0175CPV21F228J4B","better-skill-creator",null,"Author, repair, validate, measure, and package DSH skills. Use when turning a repeated workflow, prompt pattern, team convention, or model-facing template into a skill; when SKILL.md does not trigger, parses incorrectly, is too long, or produces inconsistent results; or when comparing a skill with a clean no-skill baseline. Covers DSH frontmatter, project\u002Fuser\u002Fbundled precedence, \u002Fname invocation, paired evaluation, blind grading, evidence artifacts, and private handoff.","cnidrs\u002Fdeepseek-dovetail","cnidrs","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F36650836?v=4","01M7AX6YA4ZT9XXWH0KCXSGPB5","Cnid.rs","01M7AX6YARFSJC4G9JK5CYYV41",0,"Apache-2.0","https:\u002F\u002Fgithub.com\u002Fcnidrs\u002Fdeepseek-dovetail\u002Fblob\u002Fmain\u002Fports\u002Fdsh\u002Fskills\u002Fbetter-skill-creator\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: better-skill-creator\ndescription: >-\n  Author, repair, validate, measure, and package DSH skills. Use when turning a repeated workflow, prompt pattern, team convention, or model-facing template into a skill; when SKILL.md does not trigger, parses incorrectly, is too long, or produces inconsistent results; or when comparing a skill with a clean no-skill baseline. Covers DSH frontmatter, project\u002Fuser\u002Fbundled precedence, \u002Fname invocation, paired evaluation, blind grading, evidence artifacts, and private handoff.\nlicense: Apache-2.0\n---\n\n# Better Skill Creator for DSH\n\nA `SKILL.md` is model-facing instruction. Before drafting its bytes, load `prompt-engineering` and `hypershot-protocol` with the lowercase DSH `skill` tool unless their `\u003Cskill_content>` blocks are already present. If either is unavailable, state that limitation once and continue only if the collaborator wants an unassisted draft. Never reload a body already injected by an explicit `\u002Fname` gesture.\n\n## Workflow\n\n1. Define the unreliable behavior, realistic trigger language, output evidence, and done condition. Decide whether judgment plus procedure warrants a skill; a fixed action may belong in a script, ordinary instruction, tool, or preset instead.\n2. Draft the smallest `SKILL.md` that carries the workflow. Put directly linked detail in `references\u002F`, executable helpers in `scripts\u002F`, and output assets in `assets\u002F`.\n3. Validate immediately by executing `python scripts\u002Fquick_validate.py \u003Cskill-dir>` from this installed skill's DSH-reported resource base. Fix every parser\u002Fpolicy\u002Fresource error before model runs.\n4. Prepare realistic cases with identical task bytes for treatment and baseline. State objective assertions and subjective grading criteria before outputs exist.\n5. Run independent, bounded DSH treatment\u002Fbaseline sessions. Treatment begins with `\u002F\u003Ctarget-skill>`. Baseline disables the exact `tool-skill` consumer so it receives neither catalog nor body. Never tell the baseline that a skill or comparison exists.\n6. Verify session evidence: treatment contains one target `\u003Cskill_content>` body; baseline contains neither `\u003Cavailable_skills>` nor `\u003Cskill_content>`. A contaminated arm is retained and excluded, never silently rerun as though it had not happened.\n7. Randomize candidate position and blind graders to arm identity. Reveal the mapping only after verdicts settle. Record output, stop reason, duration, available usage, contamination checks, and missing measures.\n8. Review results with the collaborator, improve the general instruction rather than naming the test case, widen the held-out set, and repeat until evidence or the collaborator says stop.\n9. Revalidate, package privately if needed, inspect the archive, and install through the DSH bundle\u002Fprofile workflow. Do not copy into project or user skill homes.\n\nStart at the stage the collaborator actually needs. Evals are the default for effectiveness claims, not a barrier to a requested drafting-only session.\n\n## DSH skill format\n\n- Directory: `\u003Ckebab-name>\u002FSKILL.md`; DSH filesystem discovery is one level deep.\n- Required frontmatter: string `name` matching the directory and nonempty string `description`.\n- Description: function plus trigger conditions, parsed\u002Fnormalized length at most 480 characters for this port.\n- Invocation: omitted fields allow both surfaces. `disable-model-invocation: true` makes a skill explicit-only but still available through `\u002Fname`. `user-invocable: false` removes the human gesture. Camel-case spellings are invalid and fail closed.\n- DSH project and user candidates outrank bundled rank. Never bypass that precedence.\n- The model sees only name\u002Fdescription before selection. A user `\u002Fname` gesture injects a canonical `\u003Cskill_content>` block; the model must not call `skill` again for it.\n- Resolve every relative resource against the loaded `\u003Cskill_resources>` base. Never hardcode a user home or development checkout.\n\nRead `references\u002Fdsh-frontmatter.md` for exact fields and validation. Read `references\u002Fdsh-distribution.md` for private package\u002Fprofile handoff.\n\n## Evaluation runner\n\n`scripts\u002Frun_eval.py` is the DSH runner abstraction and CLI. It does not create another agent loop or model-facing tool; it invokes the installed `dsh` headless profile as bounded subprocesses. Configuration names the executable, profile, provider\u002Fmodel, work fixture, artifact root, time\u002Fbyte\u002Frun limits, and allowed environment names. Credential values remain inherited only in memory and never enter command lines or artifacts.\n\nUse `--plan` first. A live run is opt-in and self-reports `UNMEASURED` when required credential variables are absent. It writes independent arm workspaces and session roots, sanitized outputs, stop reason, duration, usage when available, catalog\u002Fbody checks, randomized blind-grading inputs, verdict, and the post-verdict arm map. Any unsettled arm, missing target body, contaminated baseline, premature unblinding, or exceeded bound fails loudly.\n\nRead `references\u002Fdsh-evaluation.md` before a paired run. Execute each retained script to validate it rather than treating source inspection as proof.\n\n## Measurement language\n\n- `0` is a measured value.\n- `UNMEASURED` means the run may exist but the named metric or check was unavailable.\n- `CONTAMINATED` means a withheld skill\u002Fcatalog\u002Finput reached an arm.\n- `FAILED` means a required process, control, arm, or grader did not settle.\n\nNever turn missing timing\u002Fusage into zero, a load into evidence of benefit, or a no-skill delta into proof that the skill works. Report what was measured, what was withheld, and what remains an owner decision.\n\n## Provenance\n\nThis is an Apache-2.0 adaptation of OpenCnid's Better Skill Creator, itself derived from Anthropic's Skill Creator. Preserve `LICENSE.txt` and `NOTICE`. The DSH rewrite is identified in the package notice and does not imply Anthropic or OpenCnid endorsement.\n","# Better Skill Creator for DSH\n\nA `SKILL.md` is model-facing instruction. Before drafting its bytes, load `prompt-engineering` and `hypershot-protocol` with the lowercase DSH `skill` tool unless their `\u003Cskill_content>` blocks are already present. If either is unavailable, state that limitation once and continue only if the collaborator wants an unassisted draft. Never reload a body already injected by an explicit `\u002Fname` gesture.\n\n## Workflow\n\n1. Define the unreliable behavior, realistic trigger language, output evidence, and done condition. Decide whether judgment plus procedure warrants a skill; a fixed action may belong in a script, ordinary instruction, tool, or preset instead.\n2. Draft the smallest `SKILL.md` that carries the workflow. Put directly linked detail in `references\u002F`, executable helpers in `scripts\u002F`, and output assets in `assets\u002F`.\n3. Validate immediately by executing `python scripts\u002Fquick_validate.py \u003Cskill-dir>` from this installed skill's DSH-reported resource base. Fix every parser\u002Fpolicy\u002Fresource error before model runs.\n4. Prepare realistic cases with identical task bytes for treatment and baseline. State objective assertions and subjective grading criteria before outputs exist.\n5. Run independent, bounded DSH treatment\u002Fbaseline sessions. Treatment begins with `\u002F\u003Ctarget-skill>`. Baseline disables the exact `tool-skill` consumer so it receives neither catalog nor body. Never tell the baseline that a skill or comparison exists.\n6. Verify session evidence: treatment contains one target `\u003Cskill_content>` body; baseline contains neither `\u003Cavailable_skills>` nor `\u003Cskill_content>`. A contaminated arm is retained and excluded, never silently rerun as though it had not happened.\n7. Randomize candidate position and blind graders to arm identity. Reveal the mapping only after verdicts settle. Record output, stop reason, duration, available usage, contamination checks, and missing measures.\n8. Review results with the collaborator, improve the general instruction rather than naming the test case, widen the held-out set, and repeat until evidence or the collaborator says stop.\n9. Revalidate, package privately if needed, inspect the archive, and install through the DSH bundle\u002Fprofile workflow. Do not copy into project or user skill homes.\n\nStart at the stage the collaborator actually needs. Evals are the default for effectiveness claims, not a barrier to a requested drafting-only session.\n\n## DSH skill format\n\n- Directory: `\u003Ckebab-name>\u002FSKILL.md`; DSH filesystem discovery is one level deep.\n- Required frontmatter: string `name` matching the directory and nonempty string `description`.\n- Description: function plus trigger conditions, parsed\u002Fnormalized length at most 480 characters for this port.\n- Invocation: omitted fields allow both surfaces. `disable-model-invocation: true` makes a skill explicit-only but still available through `\u002Fname`. `user-invocable: false` removes the human gesture. Camel-case spellings are invalid and fail closed.\n- DSH project and user candidates outrank bundled rank. Never bypass that precedence.\n- The model sees only name\u002Fdescription before selection. A user `\u002Fname` gesture injects a canonical `\u003Cskill_content>` block; the model must not call `skill` again for it.\n- Resolve every relative resource against the loaded `\u003Cskill_resources>` base. Never hardcode a user home or development checkout.\n\nRead `references\u002Fdsh-frontmatter.md` for exact fields and validation. Read `references\u002Fdsh-distribution.md` for private package\u002Fprofile handoff.\n\n## Evaluation runner\n\n`scripts\u002Frun_eval.py` is the DSH runner abstraction and CLI. It does not create another agent loop or model-facing tool; it invokes the installed `dsh` headless profile as bounded subprocesses. Configuration names the executable, profile, provider\u002Fmodel, work fixture, artifact root, time\u002Fbyte\u002Frun limits, and allowed environment names. Credential values remain inherited only in memory and never enter command lines or artifacts.\n\nUse `--plan` first. A live run is opt-in and self-reports `UNMEASURED` when required credential variables are absent. It writes independent arm workspaces and session roots, sanitized outputs, stop reason, duration, usage when available, catalog\u002Fbody checks, randomized blind-grading inputs, verdict, and the post-verdict arm map. Any unsettled arm, missing target body, contaminated baseline, premature unblinding, or exceeded bound fails loudly.\n\nRead `references\u002Fdsh-evaluation.md` before a paired run. Execute each retained script to validate it rather than treating source inspection as proof.\n\n## Measurement language\n\n- `0` is a measured value.\n- `UNMEASURED` means the run may exist but the named metric or check was unavailable.\n- `CONTAMINATED` means a withheld skill\u002Fcatalog\u002Finput reached an arm.\n- `FAILED` means a required process, control, arm, or grader did not settle.\n\nNever turn missing timing\u002Fusage into zero, a load into evidence of benefit, or a no-skill delta into proof that the skill works. Report what was measured, what was withheld, and what remains an owner decision.\n\n## Provenance\n\nThis is an Apache-2.0 adaptation of OpenCnid's Better Skill Creator, itself derived from Anthropic's Skill Creator. Preserve `LICENSE.txt` and `NOTICE`. The DSH rewrite is identified in the package notice and does not imply Anthropic or OpenCnid endorsement.\n",{"name":5,"license":15,"description":7},"main","09a4abbbd75fc5d14ef2f17d67a27f6b7070917a","ports\u002Fdsh\u002Fskills\u002Fbetter-skill-creator","https:\u002F\u002Fgithub.com\u002Fcnidrs\u002Fdeepseek-dovetail","https:\u002F\u002Fgithub.com\u002Fcnidrs","In Rust We Trust.","2026-09-04 13:32:02",{"source":8,"skillName":5,"commands":32,"prompt":42,"zipAvailable":43},[33,36,39],{"tool":34,"command":35},"npx","npx skills add cnidrs\u002Fdeepseek-dovetail --skill better-skill-creator",{"tool":37,"command":38},"bunx","bunx skills add cnidrs\u002Fdeepseek-dovetail --skill better-skill-creator",{"tool":40,"command":41},"pnpm","pnpm dlx skills add cnidrs\u002Fdeepseek-dovetail --skill better-skill-creator","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fcnidrs\u002Fdeepseek-dovetail\u002Fblob\u002Fmain\u002Fports\u002Fdsh\u002Fskills\u002Fbetter-skill-creator\u002FSKILL.md",true,[45,51,55,59,63,69],{"path":46,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":49,"sourceHtmlUrl":50},"references\u002Fdsh-distribution.md","REFERENCE","参考文档",1131,"https:\u002F\u002Fgithub.com\u002Fcnidrs\u002Fdeepseek-dovetail\u002Fblob\u002Fmain\u002Fports\u002Fdsh\u002Fskills\u002Fbetter-skill-creator\u002Freferences\u002Fdsh-distribution.md",{"path":52,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":53,"sourceHtmlUrl":54},"references\u002Fdsh-evaluation.md",3460,"https:\u002F\u002Fgithub.com\u002Fcnidrs\u002Fdeepseek-dovetail\u002Fblob\u002Fmain\u002Fports\u002Fdsh\u002Fskills\u002Fbetter-skill-creator\u002Freferences\u002Fdsh-evaluation.md",{"path":56,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":57,"sourceHtmlUrl":58},"references\u002Fdsh-frontmatter.md",2254,"https:\u002F\u002Fgithub.com\u002Fcnidrs\u002Fdeepseek-dovetail\u002Fblob\u002Fmain\u002Fports\u002Fdsh\u002Fskills\u002Fbetter-skill-creator\u002Freferences\u002Fdsh-frontmatter.md",{"path":60,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":61,"sourceHtmlUrl":62},"requirements.txt",109,"https:\u002F\u002Fgithub.com\u002Fcnidrs\u002Fdeepseek-dovetail\u002Fblob\u002Fmain\u002Fports\u002Fdsh\u002Fskills\u002Fbetter-skill-creator\u002Frequirements.txt",{"path":64,"resourceType":65,"resourceTypeDesc":66,"sizeBytes":67,"sourceHtmlUrl":68},"scripts\u002Fquick_validate.py","SCRIPT","脚本",6648,"https:\u002F\u002Fgithub.com\u002Fcnidrs\u002Fdeepseek-dovetail\u002Fblob\u002Fmain\u002Fports\u002Fdsh\u002Fskills\u002Fbetter-skill-creator\u002Fscripts\u002Fquick_validate.py",{"path":70,"resourceType":65,"resourceTypeDesc":66,"sizeBytes":71,"sourceHtmlUrl":72},"scripts\u002Frun_eval.py",32867,"https:\u002F\u002Fgithub.com\u002Fcnidrs\u002Fdeepseek-dovetail\u002Fblob\u002Fmain\u002Fports\u002Fdsh\u002Fskills\u002Fbetter-skill-creator\u002Fscripts\u002Frun_eval.py",[74],{"skillNo":75,"name":76,"nameZh":6,"description":77,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":78,"licenseFileUrl":6,"sourceHtmlUrl":79,"categories":80,"keywords":81,"publishedAt":19},"01M7AX7TT5F9H5T8D5Z081X3CH","prompt-engineering","Design or improve prompts, templates, system instructions, agent instructions, and output schemas with semantic tags, explicit hierarchy, structured placeholders, attention control, uncontaminated ground, and iterative refinement. Use when precision matters or the user asks to write, debug, restructure, or review prompt bytes. Pair with hypershot-protocol when an example should teach form without leaking content.","CC-BY-4.0","https:\u002F\u002Fgithub.com\u002Fcnidrs\u002Fdeepseek-dovetail\u002Fblob\u002Fmain\u002Fports\u002Fdsh\u002Fskills\u002Fprompt-engineering\u002FSKILL.md",[],[]]