菜单
Convert a PDF to clean Markdown with tables preserved and images dropped. Use when the user wants to turn a PDF into Markdown or extract a PDF's text/tables for further processing. Input is a public URL, a workspace S3 key, or a container file path.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
| name | psd-pdf-to-markdown |
| summary | Convert a PDF (from a URL, workspace S3 key, or container path) into clean Markdown with tables preserved — no image files, no model download. |
| description | Convert a PDF to clean Markdown with tables preserved and images dropped. Use when the user wants to turn a PDF into Markdown or extract a PDF's text/tables for further processing. Input is a public URL, a workspace S3 key, or a container file path. |
| allowed-tools | Bash(/opt/agentcore-venv/bin/python3:*) |
Convert a PDF into clean, well-structured Markdown. Tables become Markdown tables;
images and graphics are dropped (no image files are written). Runs entirely inside the
container against the pre-installed pymupdf4llm engine — there is no ML model download
at runtime and no external API key.
The agent harness cannot receive a file a user drops into chat. Provide the PDF one of three ways (exactly one):
| Flag | Source | Use when |
|---|---|---|
--url <https> | A public http(s) URL | The user pasted or linked a PDF URL |
--s3-key <key> | The caller's own public-images/<email>/ prefix (requires --user) | The PDF is in the caller's workspace prefix |
--path <path> | A file already in the container/workspace | Another step already fetched the PDF |
--url is SSRF-guarded: only http/https, and the host must resolve to a public address
(loopback/link-local/private/metadata targets are refused; redirects are re-validated).
--s3-key is scoped to the caller's own public-images/<email>/ namespace (so it requires
--user and cannot read another user's objects). Input is validated by the %PDF- magic
header regardless of source.
/opt/agentcore-venv/bin/python3 /opt/psd-skills/psd-pdf-to-markdown/scripts/convert.py --url "https://example.com/report.pdf"
/opt/agentcore-venv/bin/python3 /opt/psd-skills/psd-pdf-to-markdown/scripts/convert.py --user <email> --s3-key "public-images/<email>/report.pdf"
/opt/agentcore-venv/bin/python3 /opt/psd-skills/psd-pdf-to-markdown/scripts/convert.py --path "/home/node/workspace/report.pdf"
Options:
| Flag | Description |
|---|---|
--out <path> | Output .md path (default /tmp/<stem>.md) |
--pages "0,5-10" | Convert specific 0-based pages only |
A single JSON object on stdout:
{ "status": "ok", "source": "...", "output_path": "/tmp/report.md", "chars": 8123, "markdown": "# Report\n..." }
markdown — use it directly.preview is inlined; Read the output_path file for the full document.empty_output. (Amazon Textract is wired elsewhere in the platform and is the future
fallback for scanned documents.)psd-summarize).bad_args — missing/invalid input flag, non-PDF file, or a refused --url.forbidden — --url resolved to a non-public address (SSRF guard), or
an artifact key was outside the signed caller's public prefix.misconfigured — the trusted artifact broker is unavailable.upstream_error — the URL fetch or brokered artifact download failed.convert_error — the PDF could not be parsed (corrupt or unsupported).too_large — the input exceeds 100 MB.empty_output — no extractable text (likely a scanned PDF needing OCR).Use this to read AI Studio's live capability catalog and repository catalog, execute assistants, and work with decisions. Prefer the caller-bound OAuth connection; current scopes and repository/item/segment ACLs are enforced server-side on every request.
Resolve a district email address or a Google Chat `users/{id}` to the real person in the Peninsula SD directory — name, title, department. Use this whenever you are about to refer to someone by a raw email address or a Chat sender id, instead of guessing who they are.
Schedule an agent task to run later using Amazon EventBridge Scheduler — either a recurring cron/rate, OR a ONE-SHOT follow-up at a specific time. Use this for morning briefs, weekly summaries, reminders, cron jobs, AND ALSO for any "I'll get back to you in N minutes" promise you make to the user. Do NOT use OpenClaw's built-in cron subsystem; it is disabled in this deployment.
Google Workspace (Gmail, Calendar, Drive, Docs, Sheets, Slides, Meet, Chat) as the user's agent account. Use for reading or writing email, calendar events, files, or any Workspace data.
Read-only access to PSD's Redshift data warehouse, authenticated with the caller's Cognito identity. Lists tables, inspects schemas, runs RLS-rewritten SELECT queries, and manages cross-session "lessons" learned about the data. The MCP server enforces all access control — the skill simply forwards JSON-RPC calls with the right bearer token.
Read the caller's Plaud voice recordings — list/search recordings and fetch transcripts and AI summaries. Use for Plaud, voice memos, recorded meetings, "my recordings/notes", transcript or summary of a recording.