carl-tools

Scan Claude/Codex session logs to find agent behavior patterns, Toolsmith adoption gaps, repeated frustrations, and candidates for new skills/tools.

Generate user-facing changelog entries from git history — plain language, audience-segmented, with optional CHANGELOG.md update.

Audit chezmoi dotfiles for drift, unmanaged files, and broken agent skill symlinks across Claude Code, Codex, Gemini, and other harnesses.

Consolidate/dedupe contacts from macOS, iCloud, Google, Zoho, or VCF with provenance-aware review, backups, and optional approved dossiers.

Capture a technical or product decision with chosen option, rejected alternatives, and rationale — in a format a future agent can read to reconstruct context.

Audit repos for SHA/digest dependency pinning and release cooldowns across Docker, CI, and major language ecosystems; report violations, fix with approval.

Run and triage Django/DRF security smoke checks into real risks, hygiene, and false positives using clean-scan discipline.

Review code through four empathy lenses — user, machine, developer, support — to surface quality issues that pure technical review misses.

Red-team a product's onboarding and first-run experience to find where new users get confused, think it's broken, or abandon setup.

Autonomously handle GitHub PR review comments — evaluate, implement HIGH/MEDIUM changes, run tests, commit, reply to all threads.

Run the full pre-launch gauntlet (first-contact → support-storm → trust-audit → pre-mortem) and get a single GO/CAUTION/NO-GO verdict.

Design parallel isolated test lanes for desktop apps and local tools with shared state — maps collision surfaces and splits non-colliding test lanes.

Analyze a real failure — reconstruct what happened, find root cause, extract learnings, and feed them back into the skill collection to prevent recurrence.

Multi-agent project pre-mortem — parallel agents with different failure-finding mandates, synthesized into ranked risks with mitigations.

Build a personal AI profile from your digital footprint — portrait, working-with-me guide, and compact system prompt for any AI assistant.

Repeatable multi-LLM codebase hardening sweeps: map under-reviewed surfaces, patch fixes, document learnings, and loop.

Run a CLI/app release end-to-end: verify state, update changelog/version, test packaging, commit, tag, push, publish, and verify installs.

Verify commands across this machine and named SSH hosts, comparing versions, install paths, config, and behavior with clean local/remote evidence.

Research a real person (friend or colleague) from public sources into a confidence-marked, cross-linked Obsidian People dossier.

Ruthlessly compress brainstorm output into a shippable MVP — classify ideas as DELETE / MOCK / ALREADY EXISTS / SHIP and surface the shortest path to launch.

Get validation from a different AI model before committing major changes — detects available LLM CLIs and routes to the best one.

Audit CLI/app status output for confusing, inconsistent, or trust-eroding wording; verify idempotent repeated runs and align labels across clients.

Simulate the support emails, reviews, and complaints a launch will generate, then identify product fixes that cut maintenance drag.

Audit whether a product feels trustworthy or unsafe — covering permissions, privacy, billing, file mutation, and silent-failure surfaces.

Run a repeatable before/after visual QA loop for local web/app UI changes, using stable screenshots, artifact folders, and concise visual findings.

Multi-model brainstorming room for product strategy and experience design — serious plus whimsical, multi-altitude, multi-round ideation.