artifact-analysis（工件分析）

Artifact Analysis

Turn a set of local paths (or a beagle project's conventional knowledge locations) into a cited, structured extraction of insights, context, decisions, and raw detail.

The deliverable is always on disk: a written scan plan the caller can audit, one findings file per slice, and a synthesized report with path-anchored citations. Nothing returns as inline prose, and no claim ships without a source path + verbatim excerpt behind it.

When to use

• A user asks for a local-document read — "analyze the docs folder", "scan the project for context", "extract what's in .beagle/concepts/".
• Another beagle skill invokes this one programmatically as a grounding companion (see references/companion-contract.md).
• The caller wants auditable output: a plan written before extraction, findings files per slice, and a citation-backed synthesis report.

When NOT to use

• Codebase lookups ("where is this function defined", "grep for this symbol"). Use Grep/Glob.
• Web research. Use web-research.
• Comparative evaluation of two implementations or source credibility adjudication. Use llm-judge.
• Rewriting or editing the scanned documents. Use humanize-beagle or the file tools.
• PDF / image OCR / format conversion. First version reads plain text and markdown only. beagle-core:docling is the future path.
• Paywalled or authentication-gated remote sources. This is a local-filesystem primitive.
• Coaching, challenge, or reshaping of the caller's question. That belongs to the caller.

Workflow

Four steps, in order. No step is skippable.

Hard gates

Advance to the next step only when the pass condition is true—confirm using files under output_dir (and tool output), not memory.

1. Write plan.md — resolved paths (with any auto-discovery applied), intent summary (when provided), per-slice briefs, skip patterns, and how findings will be synthesized.
2. Dispatch subagents — spawn 1-3 parallel subagents over non-overlapping slices of the resolved paths. Each writes findings/<slice-slug>.md under output_dir.
3. Synthesize report.md — fold findings into the seven fixed sections with path-anchored citations.
4. Verify before returning — satisfy the last Hard gates row; execute the numbered checklist in references/failure-modes.md (Verification checklist (orchestrator runs at end)). Any check that fails becomes an entry in Gaps & Limitations per that file—do not return a deliverable with silent checklist failures.

Receive paths + optional intent ──→ Auto-discover if paths empty
                                   ↓
                                 Write plan.md (no user-confirmation pause)
                                   ↓
                                 Dispatch subagents (up to 3 parallel)
                                   ↓
                                 Collect findings/<slice>.md files
                                   ↓
                                 Synthesize report.md
                                   ↓
                                 Return paths to caller

Unlike web-research, artifact-analysis does not pause for a plan review gate. Local scanning is cheap; plan.md is written for auditability so a reader weeks later can tell what each subagent was told. Unlike web-research, there is no fail-fast on missing tools — filesystem tooling (Read, Glob, Grep) is assumed present in the Claude Code environment.

Inputs

The skill does not parse caller-specific structures. Callers pass an intent string and/or a path list.

Auto-discovery

When paths is absent or empty, scan beagle's conventional knowledge locations:

• .beagle/concepts/ — concept specs and analysis folders.
• .planning/ — roadmap, state, and phase artifacts.
• docs/ — project documentation.
• Top-level files matching README*, BRIEF*, OVERVIEW*, CONTEXT*, CLAUDE.md, AGENTS.md.

Resolved paths (including any auto-discovery) are listed verbatim in plan.md so the caller can see exactly which files were included.

Intent modes

• intent present — extraction is targeted to what is relevant to the intent. Off-topic material goes into Raw Detail Worth Preserving only when it is a specific quote or metric worth keeping.
• intent absent — generic-salient-extraction mode. Subagents extract anything structurally important (insights, decisions, technical constraints, user/market context) without an interpretive filter.

Output location

If the caller passes output_dir, use it verbatim. Otherwise derive the default:

.beagle/analysis/<slug>/

Slug derivation (stable so re-running the same input on the same day lands on the same folder):

1. If intent is present, slug from the intent string: lowercase, strip punctuation, collapse whitespace to single hyphens, truncate to 60 characters on a word boundary (cut at the last hyphen before 60; if no hyphen exists before position 60, hard-cut at 60).
2. If intent is absent, slug from the first scanned path's basename using the same rules.
3. Prepend YYYY-MM-DD-.

Re-run protection. Before writing anything, check whether output_dir already contains plan.md or report.md. If it does and refresh is not true, refuse with a message naming the existing folder. When refresh: true, archive the prior contents into <output_dir>/.archive-<timestamp>/ before starting fresh. See references/failure-modes.md.

Callers embedding artifact-analysis in a concept-folder convention (e.g. prfaq-beagle) pass output_dir explicitly so the analysis sits next to its consumer: .beagle/concepts/<concept-slug>/analysis/.

The scan plan (plan.md)

The plan is written before any subagent runs. It is the audit trail, not a review gate — the skill does not pause for user confirmation.

plan.md contains:

• Intent — the input string, verbatim, or "generic salient extraction" when intent is absent.
• Resolved paths — every path that will actually be scanned, with a note next to any entry that came from auto-discovery vs. caller-specified.
• Slices — how the resolved paths are partitioned across 1-3 subagents. Slices are non-overlapping.
• Per-slice briefs — one paragraph per slice summarizing what that subagent is told to extract. Derived mechanically from the spec so a reader of plan.md can predict what each subagent was told.
• Skip patterns — the denylist applied to this run (see references/skip-patterns.md).
• Budgets applied — subagent count for this run (1-3) and the skim threshold in effect (see Budget defaults below).
• Synthesis approach — how the per-slice findings will combine into report.md.

Report: Wrote plan.md and proceed to dispatch. No pause, no gate.

Subagent dispatch

Up to 3 subagents run concurrently over non-overlapping slices. Each gets a mechanically-derived brief built from plan.md — no interpretation drift between the plan and the briefs. The brief template lives in references/subagent-brief.md.

Each subagent:

• Scans its assigned slice of paths, honoring skim strategies (sharded-doc: read index first; large-doc: TOC/headings first) and skip patterns.
• Writes findings/<slice-slug>.md under output_dir.
• Returns one terse status line to the orchestrator (path + status), never inline findings.

The orchestrator waits for all subagents to finish, then verifies every expected findings file exists before synthesis. A missing file is a silent failure, recorded in Gaps & Limitations — see references/failure-modes.md.

Skim strategies

Subagents do not read everything end-to-end. They apply:

• Sharded documents (folder with index.md plus multiple sub-files) — read index.md first, then only the sub-files the index points to as relevant.
• Large documents (single file > ~50 pages or > ~2000 lines) — read the TOC, executive summary, and section headings first; pull full content only from sections relevant to the intent (or structurally important when intent is absent). The findings file records which sections were skimmed vs. read fully.
• Short documents (single file, moderate size) — read end-to-end.

Skip patterns

Sensitive, binary, and vendor/build paths are skipped silently without the caller re-specifying them. The default denylist lives in references/skip-patterns.md. Each subagent records the paths it skipped under the paths_skipped frontmatter field so the audit trail shows exactly what was excluded.

Citations

Every claim in a findings file and in report.md carries a citation. The shape lives in references/citation-schema.md. At a glance:

• Required fields: path (relative to the scanned root when possible), excerpt (a verbatim quoted string from the document).
• Optional fields: lines (line range or single line, only when the subagent naturally has them — never synthesized), heading (the nearest enclosing heading), document_type.

Inline references use [^n] footnotes; the full citation sits in the numbered Sources section at the bottom of the report.

Synthesis (report.md)

The report has a fixed seven-section layout, in this order. Every section is required, every time — even when a section is thin, the report includes a bullet saying so.

1. ## Documents Found — each included path with a one-line note on its relevance.
2. ## Key Insights — the highest-signal observations from the corpus, grouped by theme.
3. ## User / Market Context — users, customers, competition, market data surfaced from the documents.
4. ## Technical Context — platforms, constraints, integrations, dependencies.
5. ## Ideas & Decisions — each tagged accepted, rejected, or open with rationale. Rejected ideas are preserved deliberately so future work does not re-propose them.
6. ## Raw Detail Worth Preserving — specific quotes, data points, metrics, and other detail that would be lost to summarization.
7. ## Gaps & Limitations — what the corpus could not establish; which paths were empty, skipped, or unreadable; which subagents failed.

Gaps & Limitations is required even when the scan looks complete. Honest accounting of what was and was not in the corpus is part of the product. The full literal skeleton the skill copies from lives in references/report-template.md.

Failure modes

• Partial success — one or more subagents fail. The skill continues with what succeeded and enumerates each failed slice under Gaps & Limitations, including the last-known brief and the stub-file reason.
• Empty corpus — path resolution (auto-discovery + explicit paths) yields zero readable documents. The skill writes plan.md and a minimal report.md with a single "no documents found" bullet under Gaps & Limitations, does not spawn subagents, and returns cleanly to the caller. Callers decide how to proceed.
• Silent-failure detection — every subagent writes at least a stub findings/<slice-slug>.md with status: frontmatter (ok, empty, failed) before returning. Missing file after dispatch = silent failure, recorded in Gaps & Limitations.
• Re-run protection — covered under "Output location" above; details in references/failure-modes.md.

Unlike web-research, artifact-analysis does not fail-fast on missing tools. Filesystem tooling (Read, Glob, Grep) is assumed present in the Claude Code environment. If it is somehow absent, the skill will surface that as a subagent failure under partial-success rather than aborting the whole run.

Full rules and the structured error shape live in references/failure-modes.md.

Budget defaults

Tunable knobs, not hard-coded invariants:

A caller that needs narrower or broader scope overrides by passing a more specific paths list — one path per subagent forces narrower slicing; a single folder lets the skill partition internally.

Companion invocation contract

Other beagle skills invoke this one via a small, documented contract. The minimal call passes only intent; the full call adds paths, output_dir, and refresh.

Worked examples for the three known callers (prfaq-beagle, brainstorm-beagle, strategy-interview) plus the success and refusal return shapes live in references/companion-contract.md. Callers are expected to honor the contract verbatim rather than invent parallel invocation styles.

Tone

This skill is a tone-neutral primitive. It does not:

• Coach the caller on which documents matter.
• Adjudicate document quality or claim credibility (that is llm-judge's job).
• Reshape the caller's intent into a different question.
• Adopt a posture (hardcore, Socratic, warm) — that is the caller's job.
• Editorialize in findings or the report.

If the caller is a coaching skill (prfaq-beagle, brainstorm-beagle), the coaching happens before and after this skill runs. Inside this skill, the intent is treated as final.

Out of scope

• Scanning paywalled or authentication-gated remote sources. Use the file tools to extract content first, then pass paths in.
• LLM-as-judge evaluation of document quality or claim credibility. Use llm-judge.
• Coaching, challenge, or opinionated reshaping of the intent.
• Rewriting or editing the scanned documents. Read-only by design.
• Binary / image OCR, PDF text extraction, or format conversion. First version reads plain text and markdown only; beagle-core:docling is the future path.
• Multi-language analysis. English-only today.
• Caching or re-use of prior findings across invocations.
• Long-running or scheduled scans.

Reference files

• references/subagent-brief.md — template the orchestrator mechanically fills from plan.md when dispatching each subagent.
• references/citation-schema.md — required and optional citation fields, footnote convention, and a well-formed example.
• references/report-template.md — literal report.md skeleton with all seven fixed sections.
• references/failure-modes.md — partial-success, empty-corpus, silent-failure detection, and re-run protection rules.
• references/companion-contract.md — programmatic invocation shape with worked examples for prfaq-beagle, brainstorm-beagle, and strategy-interview.
• references/skip-patterns.md — default denylist (sensitive, binary/media, vendor/build) applied to every run.