Competitors Analysis

Build competitor intelligence that can be shared, re-run, and audited later. This skill has two layers:

1. Repository evidence: clone or update the competitor code under the durable competitors workspace, then cite facts from actual files and commits.
2. Landscape synthesis: explain which product choices the evidence changes, including trade-offs and what would overturn the judgment. Keep positioning, pricing, strengths and gaps as supporting evidence.

This skill intentionally subsumes lightweight "competitor scan" workflows. A scan is useful for the landscape table, but it is not enough for technical conclusions.

Resolve the analysis target (read this first)

The executing agent resolves two separate inputs: the code source to analyze and the our-product context used for comparison. A repository URL identifies the first; a separately typed product name is not required to profile it.

1. Honor the current request's explicit product/market and exclusions first. If the caller requests a standalone profile, do not add an our-product comparison.
2. Otherwise, read the current project's AGENTS.md / CLAUDE.md and follow its product-contract or research entry to confirm the product and analysis scope. The caller's current project is valid context for this invocation. A directory basename alone is not enough; record the authoritative file used to identify it. A delegated caller should pass the project path and relevant entry when available.
3. A supplied repository URL without a confirmed our-product context selects a standalone Profile. Continue repository analysis; omit comparison and opportunities specific to our product. A missing customer segment does not block code facts or sourced positioning from the repository's own docs.
4. For Discover or Landscape, use the explicit request or confirmed current project to establish the scope. If comparison context remains unknown, complete any supplied repository profiles and report the exact missing context to the caller. An interactive host may ask only for the unresolved comparison input; a noninteractive fork reports it without blocking independent profiles.

Stop only when neither an analysis target nor an actionable current-project scope is available. Report that missing input to the caller. Do not invent a task, browse $COMPETITORS_BASE to choose an unrelated product, or start discovery from a self-generated market. A fork receiving neither a task request nor caller-bound project scope must stop; its inability to ask is not a reason to reject a supplied repository URL.

Entry Router

Resolve scope above before selecting a mode. A bare repository URL selects Profile, using the confirmed current project for comparison when available. Known competitors are optional for a scoped Discover request.

Use the user's wording and historical qualifiers to choose the path. A request to recap earlier work selects the inventory route before keyword-based discovery:

Before saving source profiles, resolve the user's selected destination. When the request explicitly selects a team collection, read collection destination routing and use its existing producer/review entry. Keep project-specific adoption judgments in the existing project research entry. Without that selection, retain the current project-backed learning and durable source layout below.

Durable Source Layout

Use a durable workspace, not /tmp. The default base is:

COMPETITORS_BASE="${COMPETITORS_BASE:-$HOME/workspace/competitors}"

Directory convention:

$COMPETITORS_BASE/
└── {product-slug}/
    ├── {owner-repo}/
    └── ...

Use owner-repo for GitHub repositories so forks and similarly named projects do not collide. Derive product-slug from the explicitly named or authoritatively confirmed current product. If neither applies, use standalone for a supplied repository URL; this is a storage namespace, not an invented market. Reuse the matching product directory and clone after checking its remote. Do not browse other product directories to manufacture the analysis scope.

Preflight

Bind the requested edition to the supplied source before comparing capabilities. For Git, record the remote and immutable analyzed commit. For an archive or app bundle, record its origin (or unknown), exact package/member path, content digest, and evidenced edition/version (or unknown). Keep separately distributed editions separate; mark each requested edition as found or unresolved. A digest identifies bytes, not the requested edition, latest release, or completeness of the collection. Reuse the bound local package; do not clone a substitute repository or copy an archive merely to satisfy the Git storage convention. Consult prior analysis only when the request calls for history or reuse.

Apply the following commands to Git sources only. For first ingestion or an explicit freshness/update request, establish these facts from commands, not memory:

repo="$COMPETITORS_BASE/{product-slug}/{owner-repo}"
test -d "$repo/.git"
git -C "$repo" remote -v
git -C "$repo" fetch --all --prune
git -C "$repo" log -1 --format='%H%x09%cI%x09%s'

For synthesis or continuation, reuse verified profiles and their pinned commits. Confirm the matching remote and required objects locally; do not fetch or repeat unchanged checks merely because the context was compressed. Refresh only for requested freshness, changed inputs, missing evidence, or an unresolved concern. Record the analyzed commit rather than describing it as current upstream.

If the repository is missing, clone it first. Prefer SSH for GitHub when possible:

mkdir -p "$COMPETITORS_BASE/{product-slug}"
git clone --depth 1 <git-ssh-url> "$COMPETITORS_BASE/{product-slug}/{owner-repo}"

If SSH fails for a public repository, report the failure and retry with the repository's HTTPS URL only when that keeps the work moving.

Discovery Workflow

Use gh search repos for GitHub repository discovery. Search multiple query phrases; do not trust one keyword.

gh search repos "product keywords" \
  --limit 30 \
  --archived=false \
  --json fullName,url,description,stargazersCount,forksCount,openIssuesCount,language,pushedAt,updatedAt,defaultBranch

For each candidate, record:

Clone only candidates that are relevant to the user's product or analysis goal. For broad markets, first present a shortlist with evidence and then analyze the strongest set.

Repository Fact Gathering

Before making mechanism-level claims or our-product gap judgments, read references/mechanism_evidence.md. Follow its entry-path, lifecycle and evidence-depth checks while reading the bound source; apply its execution boundary before running competitor code. A standalone Profile continues without an our-product baseline.

Read files in this order and capture exact sources:

1. Project metadata: package.json, pyproject.toml, Cargo.toml, go.mod, or equivalent.
2. README and docs: positioning, screenshots, installation, pricing links.
3. Entry points: main, bin, scripts, src/, app/, packages/.
4. Core implementation: renderer, parser, storage, export, sync, auth, API, or domain-specific modules.
5. Tests and fixtures: they often reveal supported data structures and edge cases.
6. Releases/changelog: current direction and recent changes.

Use nl -ba <file> or an editor with line numbers before citing. Every technical claim about implementation needs file:line evidence.

Landscape Workflow

Before synthesizing existing profiles, read references/landscape_synthesis.md. Follow its evidence → causal explanation → product choice → counterexample/falsifier chain and its project-document continuation contract. A capability matrix or a list of features to borrow does not satisfy Landscape. A supported conclusion that no change is warranted is valid; do not manufacture a new direction. Before closing a project-backed comparison, follow its material-opportunity disposition contract. When checking an existing JSON decision projection, load references/opportunity_disposition.md and use scripts/reconcile_opportunities.py; its structural result does not establish semantic coverage. Standalone source profiles need no adoption table.

Report Structure

For a single competitor, use references/profile_template.md. Record the current project's authoritative scope source when used. Without our-product context, omit its comparison section and keep opportunity judgments repository-specific.

For a landscape summary, use this structure. For Update, apply the change-classification procedure when filling the change table.

# {Product} Competitor Landscape

## Decisions Changed By The Evidence
For each material judgment: cited observations, causal explanation (inference),
concrete choice and its cost, alternative explanation/counterexample, and the
observation that would overturn it. State scope and unresolved assumptions.

## Current Understanding And Next Check
Link the existing project research entry; retain accepted/rejected judgments,
their evidence versions and failure conditions, and the next decision-bearing check.

## Source Register
| Competitor / edition | Carrier | Local source / member | Git remote + commit or package origin + digest | Retrieved / freshness |
|---|---|---|---|---|

## Changes Since The Prior Analysis (for Update)
| Competitor / edition | Change type | Earlier and current evidence | Consequence for the judgment |
|---|---|---|---|

## Positioning
| Competitor | User segment | Primary promise | Source |
|---|---|---|---|

## Product And Technical Comparison
| Dimension | Competitor A | Source | Competitor B | Source | Our product | Source |
|---|---|---|---|---|---|---|

## Strengths
| Competitor | Strength | Evidence | Why it matters |
|---|---|---|---|

## Weaknesses And Gaps
| Competitor | Gap | Evidence | Opportunity |
|---|---|---|---|

## Opportunities
| Opportunity | Evidence base | Product implication | Confidence |
|---|---|---|---|

## Risks And Assumptions
| Item | What is known | What still needs verification | Next check |
|---|---|---|---|

Evidence Rules

Required

Forbidden

Do not present unsupported technical claims as repository facts. Strategic inference is allowed when labeled and supported by the Landscape chain; words such as “可能 / likely” do not by themselves make a claim valid or invalid. Check what the sentence claims and its evidence, rather than banning a word:

When an implementation or market fact is unavailable, write 待验证 and state the exact next check that would verify it. Do not promote an unresolved assumption into a requirement, claimed competitive advantage, or implementation authorization.

Output Quality Bar

Before finishing, run the checks in references/analysis_checklist.md:

• Git clones exist under $COMPETITORS_BASE/{product-slug}/; supplied packages retain their bound location.
• Source register identifies the carrier and requested edition; Git remote/commit or package/member/digest is recorded, with freshness and unresolved editions explicit.
• Each technical claim has a file:line citation.
• Market facts have a source and retrieval date.
• Landscape judgments explain a concrete choice, its trade-off and falsifier; facts and inference remain separate.
• For project-backed analysis, new understanding is saved in the existing project research entry before a stage ends or context handoff; unchanged evidence is reused on continuation.
• For an explicitly selected team collection, follow collection destination routing: verify the producer/review result or report the exact unresolved handoff; retain project judgments and stable source links.
• The final answer names gaps, opportunities, and risks without pretending they are code facts.

Script

Use scripts/update-competitors.sh as the starting point for durable competitor repository management:

COMPETITORS_BASE="$HOME/workspace/competitors" \
PRODUCT_NAME="{product-slug}" \
./scripts/update-competitors.sh status

./scripts/update-competitors.sh discover "claude code viewer"
./scripts/update-competitors.sh clone-url https://github.com/org/repo
./scripts/update-competitors.sh pull

The script is a template. For a long-running product, copy it into that product's own repo or operations directory and fill the persistent competitor list.

Relationship To Product Analysis

product-analysis may invoke this skill for compare mode. Keep this skill focused on competitor discovery, repository evidence, and competitive synthesis. Do not turn it into a general product audit orchestrator.