Menu
Build expert agents from documented domains
Installer avec Codex ou Claude Copiez ce prompt, collez-le dans Codex, Claude ou un autre assistant, puis laissez-le vérifier la page du skill et l'installer pour vous.
Basé sur la classification professionnelle SOC
| name | expert-build |
| description | Build expert agents from documented domains |
| argument-hint | [init|chunk-pdf|chunk-docs|summarize|propose-beliefs|accept-beliefs|pipeline|derive-review-repair|index-sources|cert-coverage|exam|status] |
| allowed-tools | Bash(expert-build *), Bash(uvx *ftl-expert-build*), Read, Grep, Glob |
Build External Epistemic Memory (EEM) from documented domains.
An EEM is knowledge that lives outside the model, carries its justifications with it, and makes "how do you know that?" answerable by traversing justification chains. Unlike RAG (which retrieves by similarity but lacks justification), EEM maintains justified beliefs with truth values (IN/OUT), retraction cascades, contradiction records (nogoods), and derivation depth in a persistent belief maintenance system.
The architecture is hybrid: a symbolic BMS (based on Doyle's 1979 TMS) manages justifications, cascades, and backtracking, while LLMs handle semantic tasks — deriving new beliefs, reviewing existing ones, and detecting contradictions. The derive-then-review cycle intentionally over-derives, then a review pass catches errors and retraction cascades propagate corrections automatically.
This tool automates the EEM construction pipeline: chunk documents → generate summaries → extract beliefs → derive/review/repair → build search indexes.
See https://llmeem.ai/ for the full specification.
expert-build [command]
# or
uvx --from ftl-expert-build expert-build [command]
Bootstrap a new expert agent repo.
expert-build init rhcsa --domain "Red Hat Certified System Administrator"
Creates: CLAUDE.md, expert-build.md, sources/, reasons.db
Chunk a PDF into section-based source documents.
expert-build chunk-pdf sources/paper.pdf --prefix doyle-1979 --source-label 'Doyle 1979, "A Truth Maintenance System"'
expert-build chunk-pdf sources/paper.pdf --dry-run
Writes chunks to sources/chunks/ with provenance frontmatter. Tracks progress in .chunked-{prefix} manifest.
Chunk large .md/.py/.txt files by structural boundaries.
expert-build chunk-docs --input-dir ~/git/my-project --recursive
expert-build chunk-docs --threshold 20000 --dry-run
Splits markdown by headings, Python by class/def boundaries, falls back to fixed-size windows. Writes to sources/chunks/.
Generate summaries from source documents using an LLM.
expert-build summarize --parallel 4
expert-build summarize --input-dir ~/git/my-project --recursive --parallel 4
expert-build summarize --limit 10 --model gemini
Reads .md, .py, and .txt files from sources/, writes summaries to entries/.
Extract candidate beliefs from summaries for review.
expert-build propose-beliefs --parallel 4
expert-build propose-beliefs --batch-size 10 --model claude
expert-build propose-beliefs --entry entries/2026/06/01/specific-file.md
Writes proposed-beliefs.md with [ACCEPT] or [REJECT] markers. LLM recommends accept/reject.
Import accepted beliefs from the proposals file.
expert-build accept-beliefs
Reads proposed-beliefs.md, adds each [ACCEPT] belief to reasons.db.
Run end-to-end EEM construction (9 stages).
expert-build pipeline --sources-dir ~/git/my-project --parallel 4 --recursive
expert-build pipeline --pdf paper.pdf --domain "TMS"
expert-build pipeline --resume
Stages: ingest → summarize → extract → derive/review/repair/dedup loop → export → index.
Run derive/review/repair loop on existing beliefs.
expert-build derive-review-repair --rounds 5
expert-build derive-review-repair --model gemini
Build FTS5 search database from source documents.
expert-build index-sources --input-dir ~/git/my-project --recursive
expert-build index-sources --input-dir entries/ --type summary
expert-build index-sources --rebuild
Creates rag_fts.db compatible with reasons search-sources.
Map certification objectives to beliefs, report coverage gaps.
expert-build cert-coverage objectives/ex200.md
expert-build cert-coverage objectives/ex200.md --model claude
Run practice questions through LLM with belief context, discover gaps.
expert-build exam questions/practice.md
expert-build exam questions/practice.md --limit 20 --output results.md
Show pipeline progress.
expert-build status
If the user says:
expert-build init Xexpert-build chunk-pdf <path>expert-build chunk-docs --input-dir <path>expert-build summarizeexpert-build propose-beliefsexpert-build pipelineexpert-build derive-review-repairexpert-build index-sourcesexpert-build cert-coverage <file>expert-build exam <file>expert-build status