| name | parallel-llm-wiki-gap-to-issues |
| description | Use parallel subagents to mine remaining LLM-wiki/document-intelligence gaps, de-duplicate against existing GitHub issues, then create only the strongest bounded follow-on issues. |
| version | 1.0.0 |
| author | Hermes Agent |
| category | workspace-hub-learned |
| tags | ["llm-wiki","github-issues","parallel-research","de-duplication","knowledge-management","workspace-hub"] |
Parallel LLM-Wiki Gap-to-Issues
Use when the user wants more LLM-wiki / document-intelligence improvement issues and the repo already has a large existing issue graph, reports, and prior planning artifacts.
Why this skill exists
A naive single-pass issue brainstorm creates duplicates fast in workspace-hub because many adjacent items already exist as:
- umbrella issues
- batch-pack designs
- architecture/policy docs
- partially overlapping implementation issues
- recently created follow-ons in the same session
The reliable pattern is:
- split the search space across parallel subagents,
- force each subagent to stay read-only,
- have each one produce issue-ready candidates with duplicate assessment,
- synthesize centrally,
- create only the strongest non-duplicate items,
- repeat on the remaining surface until marginal candidates become weak.
Best-use trigger conditions
- User asks to "continue with recommendations using parallel agents"
- You already created some issues and need the next wave without overlap
- Repo has strong prior artifacts (queue docs, reports, architecture docs, issue history)
- Problem is not implementation, but identifying the next best bounded GitHub issues
Core execution pattern
Phase 1 — Load context and recent created issues
Before delegating, capture the already-created issue numbers from the current session and identify the main clusters already covered.
For LLM-wiki work in workspace-hub, common clusters already covered in one session can include:
- provenance / reverse lookup / doc_key
- batch-pack execution waves
- navigation / portals / index quality
- transient-to-durable promotion
- governance / conformance / retention
Always pass the newly created issue numbers into subagent context so they do not rediscover the same work.
Phase 2 — Partition the search space
Use up to 3 parallel subagents, each with a sharply bounded surface. The pattern that worked well was:
-
navigation / discoverability
- wiki indexes
- accessibility registry
- landing pages
- portals / explorers / uplinks
-
source-ingest / promotion pipeline
- batch packs
- registries / ledgers / summaries
- external-source priority queue
- remaining designed-but-unfiled promotion waves
-
tacit knowledge / governance / transient artifacts
- handoffs
- review artifacts
- WRK completions
- retention / conformance / audit trail gaps
Important: tell subagents explicitly to do read-only analysis and NOT create issues.
Phase 3 — Force duplicate-aware outputs
In each subagent prompt, require:
- 2-4 candidates max
- issue-ready wording
- explicit duplicate check against named open issues
- a recommendation whether to file or not file
- scope boundaries to avoid overlap
This is critical. Without it, subagents overproduce vague ideas that collide with open work.
Required evidence sources
Tell subagents to prioritize existing repo intelligence over rescanning blindly.
For workspace-hub LLM-wiki work, the highest-yield sources were:
- docs/reports/2026-04-16-llm-wiki-resource-intelligence-unified-review.md
- docs/reports/llm-wiki-external-source-priority-queue.md
- docs/reports/llm-wiki-staged-batch-packs.md
- docs/reports/engineering-wiki-skill-ingest-readiness-2039-2042.md
- docs/reports/engineering-wiki-skill-ingest-priority-pack.yaml
- docs/document-intelligence/intelligence-accessibility-map.md
- docs/document-intelligence/durable-vs-transient-knowledge-boundary.md
- docs/document-intelligence/standards-codes-provenance-reuse-contract.md
- data/document-index/intelligence-accessibility-registry.yaml
- data/document-index/resource-intelligence-maturity.yaml
- data/design-codes/code-registry.yaml
- knowledge-base/wrk-completions.jsonl
Use live gh issue list/view to verify duplicates rather than trusting local docs alone.
Synthesis rules in the main agent
After subagents return:
- group candidates by theme
- discard anything clearly covered by an existing open issue
- prefer issues that are:
- bounded
- grounded in live repo evidence
- distinct from umbrella issues
- implementation-shaped, not vague strategy restatements
- if two candidates overlap, choose the narrower one with clearer boundaries
- if a residual candidate is too heterogeneous after exclusions, do NOT file it
A key learning: sometimes the correct output is "do not file another issue" for a source family after overlap removal.
Creation workflow
For each surviving candidate:
- draft the body in
/tmp/*.md
- include grounding bullets from live repo state or docs
- explicitly list exclusions to prevent duplicate scope
- create with
gh issue create --body-file
- verify immediately with
gh issue view --json number,title,url,labels,body
Reusable issue families this pattern surfaced well
This approach was especially effective for finding the next wave of:
- wiki reverse lookup / provenance issues
- batch-pack execution issues
- source-title aliasing / index chunking issues
- registry-backed explorer/navigation issues
- closed-issue promotion ledgers
- transient-promotion candidate queues
- WRK normalization / structured seed work
- conformance checks like promotion audit-trail checker and invented-layer detector
Practical heuristics
- If a candidate depends on a policy doc but no enforcement exists, it is often a good issue.
- If a candidate is mentioned as "future work" in a normative doc and not already filed, it is high-signal.
- If an issue family shrinks to only a few mixed leftovers after exclusions, do not file it as a standalone wave.
- Separate curated navigation improvements from canonical index mechanics:
- portals/faceted entry points are different from chunking the authoritative index.
- Separate schema/policy from enforcement:
- field definition issue != checker issue
- queue extraction issue != retention/pruner issue
Anti-patterns to avoid
- Do not ask subagents to search the whole repo with no theme; results get noisy fast.
- Do not create issues directly from subagent outputs without a central duplicate pass.
- Do not file broad residual issues for tiny heterogeneous leftovers.
- Do not rely only on local docs for duplicate checks; use live GitHub issue state.
- Do not merge multiple adjacent ideas into one issue unless the scope boundary is extremely clear.
Minimal template for delegate_task context
Include:
- repo path
- user intent (continue recommendations with parallel agents)
- recently created issues
- allowed focus area for that subagent
- required duplicate-check set
- instruction: read-only only, do not create issues
- expected output: 2-4 issue-ready candidates with rationale and duplicate assessment
Success criterion
You know this pattern worked when each wave produces only a small number of strong, clearly non-duplicate issues, and later waves naturally converge toward fewer worthwhile candidates rather than endless brainstorm sprawl.
