ITADN
synthesisengineering/synthesis-skills
synthesisengineering/synthesis-skills · 文件 下载 ZIP
文件最后提交记录最后更新时间
README.md

Synthesis Skills

Proven AI agent skills for code review, content creation, project management, and more. Built on the Agent Skills open standard and portable across Claude Code, OpenAI Codex, Cursor, GitHub Copilot, and other capable agents.

What's new

Agent correspondence, generalized (August 2026). New synthesis-agent-correspondence skill (Communication family): how AI agents compose and send correspondence on a human principal's behalf, honestly. v2 models it as three lanes on one axis — how much of the principal is in the words: principal-direct (their words, their hands — no disclosure), the assistant lane (their words, the agent's hands — one authorship signature), and the bot lane (their direction, the agent's words — a handled-for-me signature) — with review depth demoted to internal governance and the bot-vs-assistant archetype binding a persona to its lane. Recipients learn the legend from the emoji alone. Includes the persona-registry config schema, verified channel-disclosure facts, and the three compose/send gates that pair with synthesis-message-guard. See the 4.16.0 release notes.

Executive communication for technical leaders (August 2026). New synthesis-executive-communication skill (Communication family): translating technical work for the non-technical executives who fund it — the every-noun persona test, the six-category kill-list, mechanism-to-consequence translation patterns, upward-report structure, and an in-persona adversarial review protocol. See the 4.14.0 release notes.

One-command onboarding for people and whole organizations (August 2026). New skill synthesis-onboarding v1.0.0: a convergence engine that takes a bare Mac to a working synthesis setup — plugin into Claude Code and/or Codex, an ai-knowledge-<workspace> scaffold, receipts-backed idempotent re-runs that repair half-finished installs and never overwrite files you edited, skill rename/removal reconciliation, and a built-in doctor. An organization onboards its members by shipping one declarative .agents/onboarding.yaml manifest in its knowledge-base repo — no installer code — and the curl-able onboard.sh covers individuals. See the 4.13.2 release notes.

Portable drift detection, mechanically enforced (August 2026). The synthesis-git-hooks v2.3.0 doctor no longer assumes where the skill source lives: its drift baseline resolves through an explicit override, the running copy itself, or documented install locations, so the same health check works from a fresh machine, a worktree, or a plugin cache — and a misconfigured override fails closed instead of silently skipping. synthesis-agent-conformance now scans the repository for personal workspace paths so this class of defect cannot return. See the 4.10.0 release notes.

Disclosure governance by precedent, not blacklist (July 2026). synthesis-disclosure-policy distinguishes the names you deliberately publish in your own biography from disclosures nobody approved, backed by an evidence-cited precedent ledger. synthesis-git-hooks v2.2.0 enforces it by publication surface: your site repos get full protection minus your ledgered names, public OSS repos stay pinned strict, and a missing or unverifiable ledger fails closed. See the 4.9.0 release notes.

Trustworthy resumption and safe retirement (July 2026). Activation, handoff, and SessionStart context now detect stale project checkouts by comparing the record with its fetched upstream, and handoff verifies that both client envelope formats carry identical context. synthesis-project-management v1.8.0 makes lease-managed boards self-declaring — a machine without the lease config refuses to write rather than losing changes silently, with a sanctioned lease-disable path — and adds retire_worktree.py for fail-closed, remote-verified retirement of merged worktrees. See the 4.7.0 release notes.

Symmetric verification and cross-machine coordination (July 2026). Conformance checks, doctors, and installers now resolve the Claude and Codex CLIs through overrides, PATH, and documented install locations, so the same verification runs from either client's shell. synthesis-project-management v1.7.0 adds an opt-in git-backed coordination lease — an atomic ref compare-and-swap on a shared remote — for safe same-resource sessions across machines, and claim-overlap detection now matches mixed absolute, ~, and relative claim spellings. See the 4.6.0 release notes.

A stable inbox engine across native clients (July 2026). synthesis-inbox-cleanup v1.4.0 installs verified, immutable engine releases under ~/.synthesis/inbox-cleanup/engine/. Claude Code and Codex private workflows now share engine/current instead of depending on either client's version-numbered plugin cache. See the 4.5.0 release notes.

Clean handoff after project completion (July 2026). A completed synthesis project now emits no pending actions during activation or SessionStart; checked items are never relabeled as future work. See the 4.4.4 release notes.

One repository contract for every agent (July 2026). The public source now tracks its own AGENTS.md, with Claude Code importing that same contract through a one-line adapter. CI verifies both files so contributors using Codex and Claude Code receive the same repository rules. See the 4.4.3 release notes.

Trustworthy native-plugin status (July 2026). Installer status now reads the checked-out or plugin-packaged skill tree directly, verifies that the Claude Code and Codex plugins are enabled, and fails clearly when no authoritative source is available. A stale legacy cache can no longer produce filesystem errors followed by a false pass. See the 4.4.2 release notes.

Accurate cross-client project recovery (July 2026). Active-project activation and SessionStart now share one parser that selects pending, multiline project actions. See the 4.4.1 release notes.

Parallel Claude Code and Codex sessions without shared-state collisions (July 2026). synthesis-project-management v1.6.0 registers the machine, project, heartbeat, isolated worktree/branch, source claims, and context role for every root session. Different projects can proceed independently. Same-project sessions use one canonical context owner plus non-overlapping contributors with session-specific reconciliation artifacts. The helper refuses shared worktrees, branches, claims, and context ownership. See the 4.4.0 release notes.

First-class Codex interfaces and agent-neutral automation (July 2026). Every public skill now carries a Codex interface with an explicit invocation prompt. Day-end automation installs under stable ~/.synthesis/ ownership and can launch Codex or Claude Code from one configuration, while the correspondence safety doctor validates both clients independently. Source conformance rejects client-bound runtime paths before they ship. See the 4.3.0 release notes.

One knowledge-base contract across agents (July 2026). New synthesis-kb-edit reads .agents/knowledge-base.yaml for editable and generated paths, topic routing, the one frontmatter schema, confidentiality controls, Git host, and review policy. synthesis-okf v1.1.0 adds a config-driven seven-point consistency checker, and synthesis-knowledge-capture v1.1.0 hands validation and shipping to those shared layers. A knowledge-base edit can now move between Claude Code and Codex without tool-owned workflow copies or date-field drift. See the 4.2.0 release notes.

A skip is not a pass (July 2026). synthesis-implementation-integrity v1.1.0 adds a Test Honesty check for a specific reading error: "X passed, Y skipped, 0 failed" gets read as "tests pass," but a skip is an absence of information, not a green light. The new step asks whether the skipped set could plausibly contain the one test that validates the exact property a decision depends on — most load-bearing for security, data-integrity, and irreversibility claims, where a skip in that territory is never neutral. See the 3.17.0 release notes.

Multi-agent dispatch hygiene for project management (July 2026). synthesis-project-management v1.2.0 adds a "Parallel Sub-Agent Dispatch" section covering two risks specific to concurrent writers on one project: git-index collisions (a bare git commit after git add <your files> commits everything currently staged, not just what you added, so checking git status --short / git diff --cached --name-only first has to be a mechanical prefix, not a judgment call) and tracking-doc aggregation (sub-agents that each correctly leave siblings' work alone also mean no single agent sees the combined result, so the orchestrator reconciles all reports before updating the shared CONTEXT.md/index.yaml). Project Discovery also gains a scope re-verification step: before dispatching work against a paused project's stated "N items remaining," re-derive that count from live disk/repo state — the claim goes stale the moment anything else touches the same corpus, even a workstream unaware the paused project exists. See the 3.16.0 release notes.

Google's Open Knowledge Format, validated and converted (July 2026). New skill synthesis-okf v1.0.0 fills the one gap Google's own OKF repo leaves open: a conformance validator and an idempotent frontmatter converter for OKF v0.1 (announced 2026-06-12 by Google Cloud's Sam McVeety and Amir Hormati). okf_validate.py checks the spec's three hard rules plus soft-guidance warnings and link-checking; okf_convert.py backfills frontmatter onto an existing markdown corpus without ever overwriting what's already there. Proven across several real conversions, from a small public reference repo up to a 72-doc personal knowledge base. See the 3.15.0 release notes.

Day-end that survives tired evenings (July 2026). synthesis-daily-rituals v2.16.0 provides full and Quick Close modes, owed-weekly loose-ends review, explicit Decays: dates, and a state-aware notification. Its launcher and nudge live under ~/.synthesis/day-end/, independent of client skill caches, and the launcher can open Codex or Claude Code. See the 4.3.0 release notes.

Autonomous execution as a mode (July 2026). New skill synthesis-autopilot v1.0.0 encodes the delegation contract users otherwise retype per task: one explicit phrase ("take care of this for me," "autopilot this," "handle this end to end") engages a mode that sequences the existing stack — thinking framework for decisions, plan file + context lifecycle + checkpoint for compaction survival, anti-shortcuts for quality, implementation integrity before "done." Strict trigger discipline (explicit delegation only — ambiguity resolves to not engaging), batched user-only questions at checkpoints instead of blocking, and an explicit rule that standing gates survive autonomy: delegating a task never delegates authority the user has reserved. See the 3.10.0 release notes.

Agent attribution for multi-agent projects (July 2026). When Claude Code, Codex, Cursor, or subagents contribute to the same project, git history alone cannot tell you which agent did what: different tools commonly commit under the same human author identity. synthesis-context-lifecycle v1.3.0 defines the convention — one compact line per contributing agent at the end of a session-log entry, recording agent, model, effort, scope, verification performed, and a durable ref. Unknown values stay the literal word unknown (never inferred from git trailers, which are authored claims rather than verified facts), and secrets never go in attribution fields. synthesis-project-management v1.1.0 adds the convention to its Session End and Cross-Agent Handoff protocols, so a receiving agent knows who did what, with what verification. See the 3.8.0 release notes.

Slop detection is now a free hosted tool. Slopcheck at tools.synthesiswriting.org/slopcheck/ runs the upgraded synthesis-content-quality and synthesis-fact-checking skills as a web app, with zero data collection and no signup. Same engine that ships with these skills, available without installing anything.

Two major skill upgrades shipped in May 2026. synthesis-content-quality reached v4.0 with model-family fingerprinting across eight LLM families (Claude, GPT, Gemini, Llama, Grok, DeepSeek, Mistral, Qwen), a substance-and-depth section grounded in the Frankfurt-Pennycook-Hicks-Humphries-Slater framework, the compounding-archive principle that retains patterns across the LLM era, and per-family two-axis calibration with an ESL safe-harbor. synthesis-fact-checking reached v2.0 with nine new protocol sections covering nested attribution, paraphrase drift, composite quotes, position-shifting, source-translation drift, URL rot vs hallucination, AI-generated synthetic sources, citation laundering chains, and tool-specific hallucination patterns by LLM family. See CHANGELOG.md for the full release history.

Install

One-command onboarding (new machines, non-engineers, whole ecosystem)

curl -fsSL https://raw.githubusercontent.com/synthesisengineering/synthesis-skills/main/onboard.sh | sh

The synthesis-onboarding engine converges your machine: plugin installed into whichever of Claude Code / Codex is present, optional ai-knowledge-<workspace> scaffold, verified by its built-in doctor. Re-running is always safe — it updates and repairs, and never overwrites files you edited. Organizations layer their own knowledge bases and shared skills on the same engine with one declarative manifest (no installer code); see skills/synthesis-onboarding/references/org-manifest.md.

Native plugin for ChatGPT, Codex, and Claude Code

The repository is a dual-runtime plugin. The same skills/ source tree is packaged for the ChatGPT/Codex plugin system and Claude Code.

# Codex / ChatGPT desktop
codex plugin marketplace add synthesisengineering/synthesis-skills
codex plugin add synthesis-skills@synthesis-engineering

# Claude Code
claude plugin marketplace add synthesisengineering/synthesis-skills
claude plugin install synthesis-skills@synthesis-engineering

The Codex package also restores the active synthesis project after context compaction. Run synthesis-agent-conformance to verify both runtime installations, instruction discovery, and project handoff.

Agent Skills installer

One command — installs all skills to every AI agent on your machine:

npx skills add synthesisengineering/synthesis-skills --global --all --copy

This works with Claude Code, OpenAI Codex, Cursor, GitHub Copilot, and 40+ other agents.

No Node.js? Use the shell installer

curl -fsSL https://raw.githubusercontent.com/synthesisengineering/synthesis-skills/main/install.sh | sh

Or clone and run directly:

git clone https://github.com/synthesisengineering/synthesis-skills.git
cd synthesis-skills
./install.sh install

Install specific skills only

npx skills add synthesisengineering/synthesis-skills --global --copy --skill synthesis-pr-review,synthesis-codebase-review

Update / Uninstall

npx skills update          # Or: ./install.sh update
npx skills remove synthesis-skills  # Or: ./install.sh uninstall

Durable Project Memory

The project-management skills use a three-tier memory structure:

  • CONTEXT.md for current working state
  • REFERENCE.md for stable project facts
  • sessions/ for historical session archives

That structure keeps project memory in the repo, not inside one assistant's chat transcript or tool-native memory. A project can move between Claude Code and Codex, and between synced workstations, because every agent reloads the same durable project files.

When multiple agents work on one project, the session log also records provenance: one attribution line per contributing agent, capturing agent, model, effort, scope, verification performed, and a durable reference. Git authorship cannot make that distinction on its own, because different tools commonly commit under the same human identity.

Available Skills

All skills are prefixed with synthesis- to prevent namespace collisions with skills from other repositories.

Engineering

SkillDescription
synthesis-codebase-reviewEnterprise-scale codebase audit with tiered review system
synthesis-code-audit10-dimension quality scan of code diffs with scored PASS/WARNING/FAIL verdicts
synthesis-pr-reviewDelta review methodology with security scanning and AI-analysis verification
synthesis-review-triagePR queue prioritization: scoring, author-response detection, and review routing
synthesis-code-integrationAdopt-and-adapt pattern for integrating multi-contributor code with cherry-pick safety
synthesis-code-planningStructured multi-approach evaluation before coding
synthesis-preflightPre-merge quality gate: tests, types, audit, commit hygiene, go/no-go verdict
synthesis-implementation-integrityAdversarial self-review: verify implementations are genuinely complete before shipping

Content Creation

SkillDescription
synthesis-article-writingTwo-phase workflow: research/validation then strategic writing
synthesis-content-distributionStrategic content sharing and distribution across platforms with quick-start templates
synthesis-link-researchFind authoritative links for people, organizations, and entities

Content Enhancement

SkillDescription
synthesis-content-qualityv4.0 slop-detection methodology: model-family fingerprinting (8 families), substance-and-depth tests, two-axis calibration, compounding archive
synthesis-fact-checkingv2.0 fact-checking with 9 new protocols: nested attribution, composite quotes, paraphrase drift, citation laundering, AI-synthetic sources, tool-specific hallucination signatures
synthesis-article-refreshRefresh old blog posts while maintaining temporal integrity

Communication

SkillDescription
synthesis-agent-correspondenceHow AI agents compose and send correspondence on a human's behalf — the three-lane authorship model (my words / my words via my agent / my agent under my direction), a persona-registry schema with binding archetypes, channel disclosure facts, and the compose/send gates
synthesis-concise-messagingHigh-Five Habit — condense messages to 5 sentences or less
synthesis-executive-communicationTranslate technical work for non-technical executives — the every-noun test, the six-category kill-list, and upward-report structure for CTOs and product/engineering leaders

Project Management

SkillDescription
synthesis-autopilotAutonomous-execution mode for explicitly delegated work: plan-file protocol, batched decisions, standing gates preserved
synthesis-agent-conformanceCross-agent control plane: native plugin/runtime checks, instruction migration, lifecycle-hook health, and durable handoff verification
synthesis-context-lifecycleThree-tier context architecture for managing AI working memory, with agent attribution for multi-agent provenance
synthesis-project-managementLightweight PM system for human-agent collaboration, with cross-agent handoff, agent attribution, and parallel sub-agent dispatch protocols
synthesis-daily-ritualsDay-start and day-end checklists with dependency-ordered rituals

Knowledge Bases

SkillDescription
synthesis-kb-editConfig-driven plain-language editing, validation, branching, review, and synchronization
synthesis-knowledge-captureReconcile durable session facts into the correct knowledge base with provenance
synthesis-okfValidate OKF conformance, metadata consistency, taxonomy use, and convert existing bundles

Synthesis Engineering

SkillDescription
synthesis-anti-shortcutsDeterministic enforcement of anti-shortcut discipline: costume-vocabulary catalog, constraint-first protocol, sub-agent hygiene, case studies
synthesis-content-framingContent framing with topic, sophistication, and engagement gates

Reasoning & Templates

SkillDescription
synthesis-thinking-frameworkFive-mode thinking methodology: first principles, systems, complexity, analogical, and design thinking
synthesis-voice-profilerGenerate a structured writing voice profile from samples for agent instruction files
synthesis-tree-of-thoughtMulti-expert collaborative reasoning technique
synthesis-llm-setupConfigure Claude Projects, ChatGPT GPTs, and Gemini Gems
synthesis-creative-writerCreative writer persona template
synthesis-technical-advisorTechnical advisor persona template

DevOps & Sync

SkillDescription
synthesis-git-hooksYAML-driven pre-commit policy: auto-classifies each repo by push remotes (personal vs strict), enforces tiered patterns for credentials and exposure-sensitive content
synthesis-inbox-cleanupManifest-driven email cleanup across iCloud / generic IMAP (Python), Microsoft 365 / outlook.com (Mail.app AppleScript), and Gmail (workspace-mcp API + native server-side filters). Public engine + private rules. Ships with prompt-injection defenses and adversarial test fixtures for any LLM-augmented path. macOS.
synthesis-mac-syncMulti-Mac config sync via iCloud with git repo sync and machine inventory
synthesis-meeting-transcriptsFetch AI-generated meeting notes and transcripts into local working files
synthesis-repo-guardWorkspace sync guard: detect unsynced repos, confidentiality-safe alerts, event-driven checkpoint auto-commits for private context repos
synthesis-slack-syncSlack channel sync protocol: read channels, threads, DMs to local transcripts
synthesis-skills-managerAgent-native skill installer: drift detection, synthesis merge, provenance tracking

Background Instructions

SkillDescription
synthesis-clean-textProduce text without AI watermarking patterns
synthesis-response-mergerCombine multiple LLM responses into one unified document

How Skills Work

Skills use progressive disclosure:

  1. Tier 1 (always loaded): name + description (~50 tokens) — matches your requests
  2. Tier 2 (on activation): SKILL.md body — the actual instructions
  3. Tier 3 (on demand): reference files for detailed material

The plugin layout is:

.codex-plugin/plugin.json
.claude-plugin/plugin.json
hooks/hooks.json
skills/<skill-name>/SKILL.md

When you ask your AI assistant to do something that matches a skill's description, it loads automatically. Skills that involve writing include defaults that work standalone. If you have personal preferences in agent instruction files such as CLAUDE.md or AGENTS.md, those override the defaults.

Many of these skills are practical artifacts of synthesis engineering, including synthesis coding, synthesis writing, and synthesis project management.

Licensing

  • CC0 1.0 — methodology and content skills (no attribution required)
  • Apache 2.0 — skills with executable scripts

Learn More

Read the launch article: Synthesis Skills: Install Methodology Into Your AI Workflow

Part of the Synthesis Engineering Ecosystem

Contributing

See CONTRIBUTING.md.

Author

Rajiv Pant — technology executive, AI practitioner, and creator of synthesis coding.