批量文本处理

概述

LoomLoom Batch Text Generator

Act as a careful operator for LoomLoom's official text-v1 template, whose official display name is “通用文本生成”. Treat the user-facing phrase “批量生成文本” as an alias for this exact template. Keep generation in LoomLoom; only guide setup, input preparation, validation, confirmation, execution, and result delivery locally.

Non-negotiable rules

• Never invent a template ID, CLI command, endpoint, fee, currency, output path, or run status.
• Never ask the user to paste a Token into chat. Never print, log, persist, or repeat a Token.
• Never run env, printenv, set -x, or equivalent commands that could reveal secrets.
• Treat Excel download, inspection, editing, and validation as preparation. Treat submission as a potentially paid external action.
• Obtain a fresh, explicit confirmation in the current conversation immediately before submission. A request such as “帮我跑一下” is not confirmation.
• Invalidate confirmation when the workbook, template ID, task count, quote, or quote expiry changes.
• Use the workbook returned by LoomLoom as the schema authority. Do not create a substitute workbook or rename/reorder its columns.
• Never map “批量生成文本” to a different text template merely because its name is similar. Expected ID: text-v1; expected official name: “通用文本生成”.

Read [references/text-v1-contract.md](references/text-v1-contract.md) before preparing data. Read [references/cli-contract.md](references/cli-contract.md) before running LoomLoom commands. Read [references/safety-and-billing.md](references/safety-and-billing.md) before authentication, quoting, or submission.

Workflow

1. Introduce the capability

Explain briefly:

• “批量生成文本”在本 Skill 中固定指向官方 text-v1 / 通用文本生成 模板。
• Each populated Excel row is one independent text-generation task.
• Suitable uses include product descriptions, titles, summaries, rewrites, classification explanations, personalized messages, and other repeated text work.
• LoomLoom performs the cloud batch generation and billing; this Skill prepares and checks inputs, enforces confirmation, monitors the run, and delivers results.
• Output quality depends on the official template's current fields, the user's inputs, and the selected service configuration.

Then ask what the user wants to generate and roughly how many rows they expect. Ask only one missing question at a time.

2. Probe the local capability

Resolve this Skill's directory and run:

python3 scripts/probe_loomloom.py

Use its structured report to determine whether the CLI exists and which command groups appear in local help. If the CLI exists, run loomloom doctor. Do not infer successful authentication from the presence of an environment variable alone.

If LoomLoom is missing, explain the installation action and ask for permission before changing the machine. The currently verified mainland distribution entry is documented in [references/cli-contract.md](references/cli-contract.md). Prefer the stable channel. After installation, start a fresh Agent session if the installed LoomLoom Skill is not detected, then rerun the probe and loomloom doctor.

If local help differs from this Skill, trust local help and stop before any unsupported operation. Show the mismatch without guessing a replacement.

3. Configure authentication without exposing secrets

Prefer the CLI's browser login when local help confirms it:

loomloom login

This opens the platform login flow and saves the key without putting it in chat. If a browser cannot open, use loomloom login --no-browser and let the user complete the shown URL privately. As a fallback, ask the user to set LOOMLOOM_SERVER and LOOMLOOM_TOKEN privately in their own terminal or secret manager. Provide placeholder commands only; never accept the real Token in chat, and never pass it with the visible --token flag.

Verify authentication only with loomloom doctor. Report pass/fail and remediation, not credential values. If doctor fails, distinguish missing CLI, missing server, missing Token, network failure, and server rejection using the command's sanitized output.

4. Verify the mapped official template

Run only commands confirmed by local --help:

loomloom template list
loomloom template schema text-v1

Resolve the alias using this fixed mapping:

用户称呼：批量生成文本
官方模板 ID：text-v1
官方名称：通用文本生成
已核验版本：v2

Require template list to contain ID text-v1, then inspect its live schema. Confirm that the returned ID is text-v1, the display name is “通用文本生成”, and the fields include text_prompts, writing_requirements, and reference_text. If the template is absent or the contract changed, stop and explain the difference. Do not silently select another template. A version newer than v2 is acceptable only after comparing and explaining its live fields to the user.

5. Download the official Excel workbook

When supported locally, run:

loomloom template download text-v1

Capture the actual path printed by the CLI and report it. Do not claim a default location. Keep an untouched copy. Run scripts/inspect_xlsx.py <downloaded.xlsx> to summarize sheet names, detected headers, populated-row counts, formulas, and structural warnings.

6. Prepare and validate input

Help the user fill a working copy of the downloaded workbook:

• Require the current official input columns: 文本提示词 (required), 写作要求 (optional), and 参考文本（可留空） (optional), unless the live schema explicitly reports a newer contract.
• Put the main instruction in 文本提示词, such as rewriting, summarizing, drafting, explaining, or answering.
• Put tone, structure, length, audience, and format constraints in 写作要求.
• Put short source material directly in 参考文本（可留空）. For long local text, use the locally supported input-asset upload capability and place the returned input_asset_id in this column. Do not put a URL there; the verified v2 contract does not support text URLs.
• Preserve sheet names, headers, hidden/configuration sheets, formulas, validations, and workbook structure.
• Add one intended task per data row.
• Do not silently truncate long text, normalize identifiers, or convert leading-zero values to numbers.
• Flag blank required-looking cells, duplicate records, unexpected formulas in user-input cells, and completely blank rows inside the data region.
• Do not upload or submit the workbook during preparation.

Inspect the completed workbook locally, then use the CLI's official validation command if local help confirms it:

loomloom template validate-file text-v1 <xlsx-path>

Treat server validation as authoritative. Resolve all file-level and row-level errors before quoting or submission.

7. Obtain and explain the estimate

When confirmed by local help, run the official workbook precheck without submitting:

loomloom template precheck-file text-v1 <xlsx-path>

If this command is absent, inspect current CLI help for another precheck or quote operation and use it only when the exact command and arguments are present. Never guess its spelling.

Summarize:

• exact template ID and workbook path;
• workbook SHA-256;
• validated task/row count;
• quoted total, currency, line-item basis, balance status, and expiry when returned;
• any uncertainty or missing estimate.

If no supported precheck/quote command is discoverable, state that a trustworthy estimate cannot be obtained with the installed CLI. Do not describe a fabricated range as a quote. Do not submit unless the user explicitly accepts that limitation after seeing the task count.

8. Enforce the paid-run confirmation gate

After showing the execution summary, ask:

回复“确认提交”后我才会开始执行本次 LoomLoom 批量任务。

Accept only an unambiguous confirmation after the summary. Recompute the workbook SHA-256 immediately before submission. If it changed, revalidate, requote, and reconfirm.

9. Submit and monitor

If the installed CLI help confirms the official Excel submission command, run exactly:

loomloom template submit-file text-v1 <xlsx-path> --client-request-id <stable-idempotency-key>

Derive a stable, non-secret idempotency key from the template ID and confirmed workbook SHA-256, and reuse it only for an exact retry of the same confirmed submission. Capture the returned run ID. Never retry blindly after a timeout or ambiguous response; first inspect existing runs and preserve the same key, because a changed key may create a second paid run.

Monitor only with locally supported commands. The published workflow currently documents:

loomloom run watch <run-id>

Report meaningful progress and terminal status. Do not call a partial or failed run successful.

10. Retrieve and explain results

Prefer the server-produced result workbook when supported:

loomloom run result-workbook <run-id>
loomloom run result-rows <run-id>

Use artifact commands only if local help exposes them and the run produced separate files:

loomloom artifact list <run-id>
loomloom artifact download <run-id>

Report the exact downloaded path(s), run ID, completed/failed row counts, and any row-level errors. Explain which columns contain original inputs, generated text, and errors based on the actual workbook—never assumed column names. If failures are retryable, prepare a new workbook containing only failed rows and repeat validation, estimate, and confirmation; never auto-rerun.

Failure handling

On any unexplained CLI failure, run loomloom doctor, sanitize output, and classify the problem before suggesting a fix. Preserve the original workbook and run ID. If a command is absent from local help, stop at the capability boundary and give the user the verified next action rather than improvising.