DocToSkill

将企业沉淀的各类文档（支持 DOCX、PDF、MD、TXT 等格式）一键转化为可复用的技能包（.zip）。技能包内包含统一规范的 Markdown 正文、提取的图片资源，以及基于 JSONL 格式的结构化知识索引，可直接挂载给 AI Agent 作为专属知识源使用。

Convert every supplied document into one portable, indexed skill. Keep all generated files under the writable task directory supplied by the system prompt. Do not install packages, call OCR or vision models, or depend on application framework code.

Let <skill-dir> be the skill directory supplied by the skill loader. Use it exactly as supplied. Run bundled scripts in place with python3 <skill-dir>/scripts/...; never copy scripts, guess an absolute path, prepend another directory, or add cd.

Let <task-dir> be the writable task directory supplied by the system prompt. Inputs are in <task-dir>/input. Store all workflow artifacts and the final zip below <task-dir>. Every command appends diagnostics to <task-dir>/doc.log.

Proceed directly with the commands below. Do not inventory the skill or task directory, pre-create output directories, or read bundled scripts. The scripts create their own directories and report the bounded data needed for each step.

1. Prepare documents and batches

Run once. The command is resumable after the workflow state exists.

bash

python3 <skill-dir>/scripts/prepare_workflow.py \
  --input <task-dir>/input \
  --markdown <task-dir>/markdown \
  --assets <task-dir>/assets \
  --chunks <task-dir>/chunks \
  --state <task-dir>/index-state.json \
  --log <task-dir>/doc.log

The converter accepts .txt, .md, .docx, and .pdf. It preserves DOCX/PDF images, extracts usable PDF text, and renders PDF pages without usable text as one image per page. It never performs OCR or image interpretation.

After status: prepared, immediately run the batch iterator in step 2. Never open, list, or read chunks/, chunks.jsonl, index-state.json, or individual chunk files through file tools. They are private workflow artifacts; the iterator is the only source of model-visible document text.

Source files staged as 001-original.ext retain that exact value as their document ID. Their normalized Markdown is 001-original.md; no script adds a doc- prefix. A source conversion failure fails the whole task.

Chunks preserve heading context and contain at most 5,000 Unicode characters. Each Chinese character, English letter, digit, punctuation mark, whitespace, and newline counts as one character. Adjacent small sections are packed together; oversized sections are split by semantic boundaries before exact character boundaries. Sections that contain images but no usable evidence text remain standalone and are excluded from model batches, including scanned PDF pages and image-only DOCX or Markdown content.

2. Extract grounded knowledge in batches

Run the iterator:

bash

python3 <skill-dir>/scripts/batch_index.py --next \
  --manifest <task-dir>/chunks/chunks.jsonl \
  --state <task-dir>/index-state.json \
  --parts <task-dir>/index-parts \
  --log <task-dir>/doc.log

When status is pending, read references/indexing.md, generate the complete line-based text part, write the whole returned part_path once with write_file, and rerun the same command. A batch contains multiple chunks and source units identified as u001, u002, and so on. Cite only the allowed IDs; do not copy source text or write model-authored content. The merge script restores final content exactly from the selected units.

When retry_after_invalid is true, the same pending response includes fresh source units and bounded validation errors. Its part_path contains only a small retry placeholder so accidental reads do not fail. Do not read or edit the placeholder; overwrite the whole file once with write_file using the current response. Never use execute, inline Python, heredocs, shell redirection, append operations, read_file, ls, edit_file, or an equivalent operation to create, inspect, or alter a batch part. Continue until status is complete. Never retain multiple pending batches before writing the current part.

Image-only pages are added deterministically during merge. Do not infer their contents from filenames, links, or surrounding text. Images are preserved for the generated skill to inspect later through file operations; this generation workflow does not analyze them.

3. Merge the validated index

After the iterator reports complete, run:

bash

python3 <skill-dir>/scripts/merge_index.py \
  --manifest <task-dir>/chunks/chunks.jsonl \
  --state <task-dir>/index-state.json \
  --parts <task-dir>/index-parts \
  --output <task-dir>/doc-index.jsonl \
  --log <task-dir>/doc.log

If merge reports an invalid, missing, or retry-placeholder batch part, return to step 2 and rewrite only that batch part. For an invalid workflow state, a manifest mismatch, a missing source artifact, or an I/O error, return the reported error instead of rewriting batch parts. Do not hand-edit converted Markdown, chunk manifests, workflow state, or the merged index.

4. Create skill metadata

Use the iterator's final documents, bounded outline, and outline_truncated values to create <task-dir>/skill-metadata.json. Read references/metadata.md for the schema and limits. The outline is sampled proportionally by each document's knowledge-point count, preserves document coverage when the limit permits, and includes deterministic descriptions for image-only content. When outline_truncated is true, describe represented themes without claiming exhaustive coverage. Keep all descriptions, topics, aliases, and coverage notes grounded in the final outline and document IDs. Add a coverage note only for a boundary explicitly stated by the outline; absence of a topic or version is not evidence that the documents exclude it.

Do not reopen all source chunks or load the complete index into model context. If the bounded outline is insufficient for a metadata statement, omit that statement instead of guessing.

5. Build the skill

Use metadata name as <skill-name>:

bash

python3 <skill-dir>/scripts/build_skill.py \
  --index <task-dir>/doc-index.jsonl \
  --metadata <task-dir>/skill-metadata.json \
  --markdown <task-dir>/markdown \
  --assets <task-dir>/assets \
  --zip-out <task-dir>/generate_skill/<skill-name>.zip \
  --log <task-dir>/doc.log

If build validation rejects <task-dir>/skill-metadata.json, correct only that file and rerun the build stage. Do not repeat document preparation, batch indexing, or index merging when their validated outputs are already present.

When --zip-out is relative, the successful command result preserves that relative path instead of resolving it to an absolute path. Return the reported zip_path unchanged.

The generated archive contains:

text

<skill-name>/
|-- SKILL.md
|-- agents/openai.yaml
|-- references/doc-index.jsonl
|-- references/markdown/*.md
|-- references/assets/** (when extracted assets exist)
`-- scripts/search_index.py

The generated skill reads its bounded JSONL index and may inspect packaged Markdown or images through file operations. It has no dependency on DocToSkill's host application.

Completion checks

Before success, confirm from command results:

1. The prepare command reports all expected documents and at least one chunk.
2. The iterator reports complete, with every text part and hard limit validated, every selected evidence ID and final content restored from its own chunk, and image-only chunks counted automatically.
3. Merge produces a non-empty <task-dir>/doc-index.jsonl.
4. Build reports status: complete, action: return_zip_path, and the final zip_path; its deterministic validation guarantees the zip contains SKILL.md, agents/openai.yaml, the index, source Markdown, retrieval script, and every referenced extracted asset.
5. Return only the reported zip_path immediately. Do not call ls or reopen the zip, index, source chunks, or doc.log after a successful build. Inspect them only to diagnose a reported script error.