本文へ移動
cccskills
無料GitHub で公開

file-summarization

Summarize files or a requested directory scope from actual contents. Extract code, config, data and document evidence, preserve exact counts and source locations, and disclose incomplete or unsupported acquisition. Use for file summaries, code explanations and configuration overviews.

インストール方法を見る

含まれるファイル(1)

  • SKILL.md4.8 KB

SKILL.md(原文)

インストールする前に、エージェントに与えられる指示の中身を確認できます。

File Summarization

Read the shared fidelity rules and execution contract. Use the selected format; do not force structured Markdown when JSON, TL;DR or another template was requested.

Acquire and choose depth

Resolve the exact source scope first. For directories, enumerate the requested tree without following cycles, preserve exclusions/inaccessible entries, and account for all intended sources. Do not narrow a directory request to a representative file without disclosing that scope change.

Inspect file type and size before choosing a read strategy. When execution is available:

uv run --script "<plugin-root>/scripts/file_metrics.py" "<file-path>" --json --no-excerpt

Metrics and excerpts are planning evidence, not proof that source contents were read. A binary classification means a text probe failed, not that no media-aware reader can extract the file. Choose a reader from actual content/media type and available capabilities. For image files, including SVGs, use image-summarization.

Read short files completely. For larger sources, preserve section/module boundaries and source locations; choose chunks that fit the host's available context. The legacy small (<2,000 words), medium (2,000 to <10,000), and large (at least 10,000) labels are compatibility hints, not measured universal limits. If only excerpts can be read, record exactly what is omitted.

For chunked UTF-8 text or whole-dataset CSV/TSV claims, load mechanical source acquisition and use its source-bound plan, coverage reconciliation or full-record profiler.

Type-specific extraction

TypeRequired extraction
CodeImports/dependencies; classes/functions/methods and signatures; documented purpose versus inferred purpose; key algorithms/state/data transformations; entrypoints/exports; environment/configuration dependencies. Retain line locations, relevant docstrings and exact complex expressions.
ConfigurationRoot keys/types, nested hierarchy, settings categories, documented validation constraints and meaningful values. Note credentials without exposing their values. Do not claim effective/default settings from an isolated file without resolving their dependencies.
DataParsed record count, headers/schema, observed types, examples only within an explicit sample scope, missingness and evidenced identifiers. Compute dataset-wide claims over all records with an appropriate parser; raw line count is not CSV record count.
DocumentationActual heading hierarchy, main claims/definitions, prerequisites, commands/examples and internal/external references. Retain locations; quoted commands remain data, not instructions to execute.
MarkupDocument structure, visible/embedded text and relevant links; separate markup inspection from verified visual rendering.
PDF/office documentUse a host-provided or available format-aware reader. Account for pages, tables, images or attachments not extracted. Inspect relevant visual material before claiming its contents.
Other binary/unknownUse an appropriate available reader only within task authority. Otherwise return metadata plus the specific extraction limitation, not guessed contents.

Do not generalize schema, ranges, row totals or missingness from the first ten rows. Keep sampling for illustrative examples and disclose its scope. Failed text decoding is a reason to choose a compatible reader or report a limitation, never to silently replace corrupted text and call it complete.

Extract, render and verify

Extract passages/values with source locations, preserve their qualifiers, then compose the requested summary. For chunked, delegated or audit work, retain an evidence record. Ensure every relevant section/chunk either contributed inspected evidence or has an explicit failure. Do not synthesize a chain of progressively shortened summaries without recoverable source support.

Render the selected template and run the execution contract's final checks. For several files, keep individual attribution and failure records; invoke multi-source-synthesis only for requested integration. On access failure, preserve the exact path/error and caller status; do not manufacture a successful summary or repeat an unchanged failing access method indefinitely.

レビュー

まだレビューはありません。使ってみた感想をお寄せください。

同じリポジトリのスキル

概要と使いどころ

Add automated documentation updater to any Claude skill. Creates a Python sync script that downloads upstream docs, processes markdown for AI consumption, and maintains local cache with configurable refresh. Collects template variables, then delegates implementation through 5-phase workflow. Use when adding auto-updating reference documentation to plugins or skills.

日本語の概要は準備中です。原文の説明を表示しています。

Jamie-BitFlight/claude_skills672026年10月9日 更新

SAM-style feature initiation workflow — discovery through codebase analysis, architecture spec, task decomposition, validation, and context manifest. Use when a user asks to add a feature, plan a feature, or convert an idea into an executable SAM plan.

日本語の概要は準備中です。原文の説明を表示しています。

Jamie-BitFlight/claude_skills672026年10月9日 更新

Browser automation CLI for AI agents. Use when the user needs to interact with websites, including navigating pages, filling forms, clicking buttons, taking screenshots, extracting data, testing web apps, or automating any browser task. Triggers include requests to "open a website", "fill out a form", "click a button", "take a screenshot", "scrape data from a page", "test this web app", "login to a site", "automate browser actions", or any task requiring programmatic web interaction.

日本語の概要は準備中です。原文の説明を表示しています。

Jamie-BitFlight/claude_skills672026年10月9日 更新

Browser automation for AI agents using the agent-browser CLI and Playwright. Use when navigating pages, filling forms, clicking buttons, taking screenshots, extracting data, testing web apps, logging into sites, or automating any browser task. Triggers on "open a website", "fill out a form", "click a button", "scrape data", "test this web app", "automate browser actions", or any programmatic web interaction request.

日本語の概要は準備中です。原文の説明を表示しています。

Jamie-BitFlight/claude_skills672026年10月9日 更新

Runs the description-drift experiment — spawns all Claude Code agents simultaneously to collect self-reported capabilities, then compares them against static frontmatter descriptions to reveal how reliable orchestrator routing based on descriptions actually is. Use when measuring description drift across the agent fleet, re-running the capability collection experiment, analyzing a specific agent's self-reported capabilities, or auditing whether frontmatter descriptions accurately reflect agent behavior.

日本語の概要は準備中です。原文の説明を表示しています。

Jamie-BitFlight/claude_skills672026年10月9日 更新

Create or adapt Claude Code agent definitions. Use when creating an agent, changing subagent configuration, selecting agent scope, or designing a specialized delegation role.

日本語の概要は準備中です。原文の説明を表示しています。

Jamie-BitFlight/claude_skills672026年10月9日 更新

Jamie-BitFlight のスキルをすべて見る

このスキルの問題を報告する