Design Meta Skill

Design Reusable Meta Skill Bundle

Design a parameterized Creator workflow for future distillation anchors. This skill does not produce the operating graph for the current task and does not execute the final research or software task. The paper distinguishes only task-agnostic and task-oriented distillation; design-reusable is a Creator construction strategy, not a third distillation form.

distill-ml-knowledge owns anchor identification, Q scoping, adequacy assessment, composition assessment, and strategy selection. This skill owns the reusable workflow design, validation, approval, and managed Creator-infrastructure import after it receives that exact handoff.

Required Handoff

Require the current routing handoff with:

• decision revision and construction strategy: design-reusable;
• anchor kind, anchor identity, and distillation form;
• current task tau = (q, D, E, g) when the gap came from a task-oriented run;
• future anchor class and intended operating output when it came from a task-agnostic source run;
• scoped capabilities Q, required evidence/source behavior, and verification/recovery needs;
• Creator-visible capability matrix and inventory snapshot;
• uncovered recurring construction contract, recurrence evidence, and why extension or bounded composition is inadequate;
• construction constraints, unknowns, assumptions, and approval state.

If invoked without this handoff, read ../distill-ml-knowledge/SKILL.md and return through canonical routing. Do not start another reuse, composition, or gap decision here. If the visible inventory or requirements changed materially, return the evidence for a new routing revision.

Workflow

1. Validate the handoff and stop with the exact missing blocking fields. Do not inspect Researcher-only skills or hidden paths.

2. Read reusable-bundle-specification.md. Draft the exact future-anchor and four-stage specification, preserving exclusions and assumption owners. Obtain approval before material source exploration, file generation, or compute use.

3. Stage all artifacts outside live skill roots, for example:

   textCopy code
   ~/.disco/agent/creator-runs/<run>/<timestamp>/
     routing-handoff.md
     reusable-bundle-spec.md
     evidence-plan.md
     draft/<meta-skill-id>/
     verification/
     construction-record.md
   

4. Read generation-verification-and-review.md and generate a workflow that later performs scope, ground, construct, and verify for caller-supplied anchors.

5. Parameterize all source, task, environment, permissions, budget, graph, and handoff inputs. Never hardcode the current checkout, credentials, temporary task state, or machine-specific paths.

6. Keep every draft meta SKILL.md explicitly metadata.disco-role: meta. Do not include target-specific agents/ or agents/openai.yaml. Require every operating root, router, and sub-skill produced by the future workflow to declare metadata.disco-role: operating.

7. Verify the candidate workflow on representative anchors: one task-agnostic source anchor, one task-oriented task anchor with source discovery, one source/access failure and recovery, one graph verification failure and local repair, one task-bound/project deployment, and one self-contained/managed deployment. Confirm uncertain reuse defaults to project scope, each graph stays in one scope, and repo output uses its specialized locked transaction.

8. Run the validator before review:

   bashCopy code
   node scripts/validate_meta_skill.mjs <draft-meta-skill-dir>
   

9. Present the routing revision, reusable-bundle specification, evidence plan, candidate files, verification results, unresolved gaps, exact managed target, collisions, and overwrite impact. Obtain approval for this exact artifact revision.

10. After approval, import only the reviewed runtime meta bundle as Creator infrastructure:

    bashCopy code
    node scripts/import_meta_skill.mjs \
      --agent-dir ~/.disco/agent \
      <creator-run-dir>/draft/<meta-skill-id>
    

    Use --overwrite only after separate approval. The helper refuses to replace an operating or unclassified target.

11. Tell the user to run /reload. The next invocation of the imported meta skill, not this design run, constructs and verifies the operating graph and writes its researcher-handoff.md.

Required Outputs

• validated routing handoff and recurring-gap evidence;
• approved reusable-bundle specification;
• staged, parameterized meta-skill draft and evidence plan;
• verification results, repairs, unresolved points, and construction record;
• separate managed Creator-infrastructure import proposal or result;
• contract for later operating graph verification, one-scope deployment, locked import, and Researcher handoff.

The future workflow's output is accepted operating graph G and construction record R; the current meta bundle itself is not an operating graph. Its reusability is assessed against recurring anchors and the generated workflow's ability to support both project deployment under .agents/skills/ and managed deployment under ~/.disco/agent/skills/.

Failure Rules

• Missing or stale handoff: return to distill-ml-knowledge.
• Unresolved source access, credentials, hardware, permissions, or budget: keep the item in the future construction environment and stop before use.
• Failed strict verification: report the unverified bundle and do not import it.
• Invalidated recurring gap: return evidence for a new routing revision.
• Collision, overwrite, expanded source access/cost, or weakened verification: obtain separate approval.
• Never write drafts or review artifacts into live skill roots, split one future operating graph across scopes, send repo metadata through the generic importer, or execute the final downstream task here.