analyze

Repo Analyzer

Multi-agent GitHub repo analysis. Three parallel investigators produce domain-specific reports; a synthesizer combines them into a personalized recommendation grounded in the user's profile (below).

When to use

Trigger on any request that involves evaluating a specific GitHub repo:

• "Analyze https://github.com/owner/repo"
• "What do you think of ?"
• "Should I contribute to ?"
• "Is worth learning?"
• "How does compare to ?"
• A bare GitHub URL with no clear ask — interpret it as a request for analysis.

If the user asks a one-line factual question ("how many stars does X have?"), do NOT trigger this skill — answer directly.

The user's profile

The synthesizer evaluates every repo through four lenses (contribute / learn / compete / influence) — but those lenses only carry weight against the user's own profile (role, expertise, technical taste, what they're optimizing for).

Resolve the user's profile in this order:

1. Memory / CLAUDE.md — if the user has a profile file, memory entries, or CLAUDE.md describing their role, languages, and what they value, use that. Don't re-ask.
2. Inline override — if the user passes --profile "..." or includes profile context in the request itself, that wins for this run.
3. Ask once, briefly — if no profile is available, ask in 2-3 short questions: role/seniority, primary languages or domains, what they're optimizing for (career capital, learning, side-project ideas, hiring signal, etc.). Offer to save the answer for future runs.

The three investigator agents (repo-intel / market / tech-arch) do not use the profile — they produce neutral reports. Only the synthesizer ties findings back to the user. Output preference, regardless of profile: bottom-line first, ranked, concrete, no hedging filler.

Workflow

1. Parse input

Extract <owner> and <repo> from whatever the user gave you (URL, owner/repo shorthand, or a name you have to disambiguate via search). If ambiguous, ask once.

Determine depth: default deep. Honor --depth quick or --depth standard if the user passes either.

2. Set up workspace

TS=$(date +%Y-%m-%d-%H%M)
WORK="./repo-analysis/${OWNER}__${REPO}/${TS}"
mkdir -p "$WORK"/{reports,_raw}

On claude.ai web (no persistent FS), use /home/claude/repo-analysis/... instead and remember you'll be presenting files at the end.

3. Clone (skip if --depth quick)

git clone --depth=200 "https://github.com/${OWNER}/${REPO}.git" "$WORK/_clone"

--depth=200 gives enough history for cadence analysis without paying for a full clone on large repos.

4. Run the deterministic stats pull

bash $SKILL_DIR/scripts/stats.sh "$OWNER" "$REPO" "$WORK/_clone" "$WORK/_raw"

This dumps gh-API JSON, contributor breakdown, language LOC, dep manifests, recent commit/release/issue activity into _raw/. Subagents read from there.

5. Dispatch the three investigator subagents

In Claude Code: spawn three Task subagents in parallel. Each one gets:

• The path to its prompt file (e.g. $SKILL_DIR/agents/repo-intel.md)
• OWNER, REPO, WORK paths
• The output path it must write to (e.g. $WORK/reports/repo-intel.md)
• The depth flag

Give each subagent a prompt like:

Read your instructions from <SKILL_DIR>/agents/repo-intel.md and follow them exactly. The repo is <OWNER>/<REPO>. Raw data is in <WORK>/_raw/. The clone (if present) is at <WORK>/_clone/. Write your final report to <WORK>/reports/repo-intel.md. Depth: <deep|standard|quick>. Return a 5-bullet summary when done.

The three agents are:

Agent	Prompt file	Writes to
repo-intel	agents/repo-intel.md	reports/repo-intel.md
market	agents/market.md	reports/market.md
tech-arch	agents/tech-arch.md	reports/tech-architecture.md

On claude.ai web: no Task tool. Run the three roles sequentially yourself, each as a focused pass: read the agent prompt file, do the work, write the report. Take real breaks of context between them — finish one fully before starting the next.

6. Synthesize

Once all three reports exist, you (the orchestrator) read all three plus the profile above and produce overview.md. Use the template in references/output-templates.md (section "overview.md"). The overview must:

1. Open with a one-paragraph bottom-line: contribute / learn / compete / ignore, ranked by recommendation strength.
2. Cross-reference findings — call out where market and tech evidence reinforce or contradict each other.
3. End with concrete next actions (specific files to read, specific issues to look at, specific competitive angles to explore).

No hedging. If the call is "ignore this, it's a vendored fork of X with no activity", say that.

7. Present results

Claude Code: print the four file paths, in this order: overview.md first, then the three sub-reports. Include a one-line summary per file.

claude.ai web: also call present_files so the user can download them, and additionally print the overview content inline (the sub-reports get linked, not pasted, to avoid wall-of-text).

Depth modes

Mode	Clone?	Web search rounds	Dep analysis	Sentiment dive
quick	no	1 per agent	manifest only	no
standard	yes	2 per agent	direct deps	light
deep	yes	3+ per agent	+ transitive	yes (HN+Reddit)

Default is deep.

Failure modes to guard against

• Private repo / 404. gh returns an error. Tell the user the repo is private or doesn't exist; do not invent data.
• Monorepo with the "real" project in a subdir. If the README points hard at a subdirectory, focus tech-arch analysis there but mention the containing org/monorepo structure in repo-intel.
• Recently renamed / transferred repo. GitHub redirects; record the redirect target in repo-intel.
• Vendored fork with no real activity. Detect via "0 commits ahead of upstream" or near-identical README. Flag and short-circuit — no need for three agents on a dead fork.
• Massive repo (Linux kernel scale). Switch tech-arch to top-level only; do not try to enumerate everything.

Reference files

• references/tools.md — concrete gh / git / tokei / jq invocations the agents and orchestrator rely on.
• references/output-templates.md — markdown templates for all four reports. Use them; don't reinvent structure per run.

Subagent prompt files

• agents/repo-intel.md — health, governance, stats, bus factor, activity.
• agents/market.md — positioning, alternatives, gap analysis, sentiment.
• agents/tech-arch.md — languages, deps, design patterns, build/test/CI.

Each is self-contained; the subagent reads only its own file.