Skip to content

Skill candidates for projio ecosystem

Overview

Identified through analysis of recurring session patterns, task logs (April 1-3), issue patterns, and multi-step MCP tool transaction boundaries. Two skills already implemented: biblio-batch-curate and mcp-tool-scaffold. This note tracks the remaining candidates.

Implemented

  • [x] biblio-batch-curate — ingest → fetch → validate → extract → compile → index
  • [x] mcp-tool-scaffold — implement → wrap → register → update instructions

Tier 2: Clear pattern, moderate frequency

spec-to-tasks

  • What: Read a design spec from docs/specs/, identify affected subsystems, decompose into a coordinated task chain with dependency order, optionally schedule via schedule_queue with after chains
  • Evidence: April 2 batch (6 tasks from bib-architecture.md) was hand-decomposed. Every major spec (flatten refactor, hidden folder consolidation, manuscript design) follows the same decomposition pattern
  • Tools: note_create, note_read, schedule_queue, project_context
  • Why it earns its keep: Encodes the spec → task chain → scheduled execution pattern that currently takes a full planning session

manuscript-preflight

  • What: Bundle all pre-build validation: manuscript_status → manuscript_cite_check → manuscript_validate → manuscript_figure_build_all → biblio_library_quality. Aggregate errors/warnings, recommend proceed or fix
  • Evidence: Citation and figure issues only surface at render time. Every manuscript_build call should be preceded by this bundle
  • Tools: manuscript_status, manuscript_cite_check, manuscript_validate, manuscript_figure_build_all, biblio_library_quality, manuscript_journal_check
  • Why it earns its keep: Prevents wasted render cycles from missing citations or broken figure references

pipeline-validate

  • What: Pre-execution health check: pipeio_registry_validate → pipeio_contracts_validate → pipeio_completion → pipeio_cross_flow → pipeio_run (dryrun). Catches contract mismatches before real runs
  • Evidence: End of any major workflow; implicit in validate-and-deploy but not standalone
  • Tools: pipeio_registry_validate, pipeio_contracts_validate, pipeio_completion, pipeio_cross_flow, pipeio_run
  • Why it earns its keep: Cross-flow dependency issues are expensive to debug mid-run

notebook-to-mod

  • What: Promote exploration notebook to pipeline mod: pipeio_nb_diff → pipeio_nb_promote → pipeio_rule_stub → pipeio_rule_insert → pipeio_config_patch → pipeio_registry_validate
  • Evidence: Step 8-9 of grand routine, done for every new analysis. Currently described in integrate-pipeline.md workflow but not as a reusable skill
  • Tools: pipeio_nb_diff, pipeio_nb_promote, pipeio_rule_stub, pipeio_rule_insert, pipeio_config_patch, pipeio_registry_validate
  • Why it earns its keep: 6 interdependent steps with strict ordering

codelib-onboard

  • What: Add external library: git clone → codio_add(role="external") → update repos.yml → codio_validate → indexio_sources_sync → indexio_build
  • Evidence: April 3 issue — codio_add_urls(clone=True) doesn't actually clone. Manual workaround needed for pycircstat2 onboarding
  • Tools: codio_add, codio_add_urls, codio_validate, indexio_sources_sync
  • Why it earns its keep: Unblocks known codio friction; every external library follows the same path
  • Note: Depends on fixing codio_add_urls clone behavior first (issue-arash-20260403-054500)

Tier 3: Useful but lower frequency

bibliography-rebuild

  • What: Staleness detection + recompile: biblio_merge → pipeio_modkey_bib → biblio_compile. Quick 3-tool chain
  • Evidence: ecosystem_status() already flags staleness but doesn't auto-fix
  • Verdict: Low ceremony — might be better as a one-liner in CLAUDE.md workflow conventions than a full skill

session-bootstrap (as skill)

  • What: Promote the existing workflow prompt to a skill: project_context → ecosystem_status → indexio_status → git_status + worklog agenda
  • Evidence: Already documented in session-bootstrap.md workflow prompt
  • Verdict: Useful for run_prompt sessions that need deterministic startup. Low priority because manual sessions already follow the workflow prompt

Architecture notes

  • Skills should be < 5 sequential steps where possible. biblio-batch-curate is the exception (10 steps) because the pipeline genuinely requires it
  • Skills encode tool orchestration, not prose guidance — they say "call X then Y" not "think about whether to..."
  • Decision gates (human-in-the-loop pauses) should be explicit in the skill, not implicit
  • Skills are loaded via skill_read() at ~2-5k tokens each. An "agent persona" would cost 30k+ tokens for comparable coverage
  • The grand routine workflow prompts remain the session-level routing layer; skills are the tool-level execution layer underneath