| name | llm-wiki-distill |
| description | End-to-end runtime skill for LLM Awesome Wiki distillation. Use when an agent needs to plan or run a bounded workflow from raw source material through source packets, semantic drafting, grounding, wiki construction, compare/review, validation, closure, and commit. This skill routes to narrower subskills only when a stage needs deeper execution detail. |
LLM Wiki Distill
Role
Use this as the main runtime entrypoint for an end-to-end knowledge workspace
distillation round.
This skill owns the normal execution sequence. Detailed rules/** files are
reference material for artifact contracts, edge cases, and disputed decisions.
Do not load the full rules directory by default.
Do not use docs/ as the operational workflow entrypoint. docs/ are
maintenance and design records for this system repository.
Preflight
- Run
git status --short --branch.
- Identify the active repository:
- system repo: reusable skills, rules, templates, contracts, tools, tests,
and design records
- knowledge workspace repo: live
raw/, wiki, reports/, plan/,
purpose.md, and schema.md
- In a knowledge workspace, read
AGENTS.md, purpose.md, schema.md, and
KERNEL-MANIFEST.md when they exist.
- Create or update the target plan before editing tracked files. The plan
must name the goal, fixed input set, expected outputs, validation, review
state, and commit intent.
- Identify whether this is source intake, wiki construction, quality gate, or
a multi-stage end-to-end round.
If the workspace is new, do an overview-first initialization before detailed
wiki writing. Create or update wiki/overview.md, a skeleton wiki/index.md,
and a staged distillation plan.
Runtime
For a normal document or PPT corpus round:
preflight and plan
-> fixed input set
-> source inventory
-> source packet
-> semantic draft from packets plus raw/rendered/external reading
-> grounding pass to source anchors, evidence, or review items
-> optional evidence/claim map for important claims
-> wiki construction analysis
-> source/chapter wiki updates
-> compare/review/validation
-> closure decision
-> log, commit, push
Keep the round bounded. Do not process a large corpus in one pass unless the
plan explicitly says the workspace is small enough.
Treat the source packet as the audit baseline, not the semantic ceiling. For
dense technical material, especially PPT/PDF lessons, the round may use raw
source views, rendered pages, external extractor output, or high-density notes
to draft semantics before final wiki prose. Accepted wiki knowledge still must
be grounded to source anchors, evidence, accepted review decisions, or explicit
review items.
Do not mix generated knowledge workspace updates with system-rule, skill,
template, schema, or tool-source edits in the same commit. If the workspace
round exposes a system asset problem, record it as a separate system repo task.
Route Deeper
Load the smallest subskill set needed:
- Source intake, source inventory, source packet, adapter handoff, source-type
packet profile, or generated-field review:
skills/llm-wiki-source-packet/SKILL.md
- Semantic draft, grounding pass, wiki construction analysis, source/chapter
pages, overview, index, or wiki log:
skills/llm-wiki-wiki-round/SKILL.md
- Compare report, review queue, source/wiki coverage, claim audit, validation
note, or round closure:
skills/llm-wiki-quality-gate/SKILL.md
Do not route every task through every subskill. Load the subskill only when its
stage is active.
Outputs
An accepted end-to-end round should leave:
- target plan or updated staged plan
- source inventory and source packet updates when raw sources were in scope
- semantic draft or wiki construction analysis when wiki pages changed
- grounding evidence, review routing, or explicit ungrounded candidate notes
- accepted source/chapter wiki updates
- compare or validation report
- review queue updates when judgment remains unresolved
- closure decision
- workspace log entry and commit
The default wiki surface is source/chapter oriented. Do not create research
object pages or downstream skills/tools unless the plan explicitly selects that
scope.
Progressive Disclosure
Open detailed rules/** only when the active subskill says the round needs
extra detail. Typical triggers are complex source types, generated visual or
OCR content, important claim support, modality conflicts, review queue
lifecycle, or failed validation.
rules/README.md is a reference index, not the default runtime path.
Validation
Run the available checks that match the touched artifacts. Prefer:
python -m llm_wiki_tools workspace-check --workspace . --mode all
If a checker is unavailable, record that state in the validation note. Missing
checker support is not a pass.
Before commit:
- Review the intended diff.
- Update the workspace log.
- Commit only accepted
close-pass or explicitly allowed
close-with-review state.
- Push to the configured branch when the workspace rules require it.