bytedance/deer-flow
21 skills31.7k installs1M starsGitHub
Install
npx skills add https://github.com/bytedance/deer-flowSkills in this repo
1Claude To Deerflowclaude-to-deerflow documents how a solo builder or coding agent talks to ByteDance DeerFlow over HTTP using the repository's chat.sh helper: verify gateway health, open or continue a LangGraph thread, pick a reasoning mode, and collect the streaming reply. It suits builders who run DeerFlow locally or behind a custom DEERFLOW_URL and want Claude Code, Cursor, or Codex to delegate research or multi-step agent work to DeerFlow instead of re-implementing those graphs in the IDE. The skill is intentionally thin integration glue—bash, curl, and environment-driven base URLs—not a full DeerFlow operator manual. Use it when DeerFlow is already running (for example via make dev in the DeerFlow project) and the task is to hand off a prompt or continue a thread. Intermediate complexity reflects streaming APIs, thread IDs, and mode selection rather than a single static REST call.2.4kinstalls2Ppt GenerationPPT Generation is an agent skill for solo builders who need a credible slide deck without hiring a designer or manually aligning fonts on every page. When you ask to create, generate, or make a PPT or PPTX, the skill plans the slide flow, picks a presentation style, generates images sequentially with style continuity, and composes the final file. It targets founders preparing investor updates, product walkthroughs, course content, or internal demos inside agentic workflows such as Deer Flow. You choose a visual lane—from business-clean to glassmorphism or dark-premium—so the deck matches SaaS positioning rather than generic clip art. The workflow depends on image generation per slide, so runs take longer than text-only outlines but produce shareable artifacts. Use it in Build for documentation and demo collateral, and again in Launch when you need polished visuals for distribution.2.2kinstalls3Frontend Designfrontend-design is a ByteDance deer-flow agent skill for solo builders who want distinctive, production-grade web interfaces instead of repetitive “AI slop.” Before writing code, it walks you through purpose, tone, constraints, and a memorable differentiator, then applies concrete aesthetics rules for typography pairing, committed color themes, motion, spatial composition, and background depth. It explicitly bans generic font stacks, predictable purple gradients, and layout templates that could belong to any product. The skill fits whenever you are building or polishing web components, dashboards, landing experiences, or React views and need an agent to both think like a designer and implement working markup and styles. It is editorial and procedural—no external APIs—so it pairs naturally with your existing stack and design tokens while pushing you toward one coherent bold direction per surface.2.1kinstalls4Data Analysisdata-analysis is a Deer Flow agent skill backed by a Python script that uses DuckDB to analyze Excel and CSV files locally. Solo builders often receive metrics as spreadsheets—ads exports, Stripe dumps, survey CSVs—and need fast answers without spinning up BigQuery or pandas notebooks for every question. The workflow loads files into an in-process DuckDB database, exposes schema discovery, runs SQL, produces statistical summaries, and can export results. A file-hash cache avoids reloading unchanged uploads, which matters when an agent iterates on the same export during a debugging or growth review session. Dependencies are bootstrapped on first run if duckdb or openpyxl are absent. Invoke it when you have concrete files on disk and need repeatable analysis the agent can drive from the terminal. It is strongest for exploratory analytics and validation of exported datasets, not for streaming production ETL.2.1kinstalls5Video Generationvideo-generation packages a small Python routine for solo builders and agent pipelines that need programmatic video clips from prompts rather than manual editor work. It reads a UTF-8 prompt file, optionally attaches one or more reference images encoded as base64 JPEG assets with referenceType asset, and submits a JSON payload to Google's Gemini API Veo 3.1 generate preview model via predictLongRunning. The function extracts the operation name from the initial response and loops on GET status until the job finishes, then writes results to the caller's output path. Aspect ratio defaults to 16:9 but can be overridden for different channels. This is narrowly scoped automation: no SKILL.md narrative beyond the code, so agents should treat it as a generator integration when Deer Flow or similar repos already depend on GEMINI_API_KEY and local image assets. It fits builders stitching generative demos, marketing loops, or agent-produced storyboards into a backend or CLI tool.1.9kinstalls6Find Skillsfind-skills in the deer-flow repo documents a concrete install-and-link ritual for agent skills: pass a fully qualified skills.sh coordinate, run npx skills add globally, confirm the skill landed under ~/.agents/skills, then symlink into the project's skills/custom directory after resolving the repo root via deer-flow.code-workspace. Solo builders using Deer Flow or similar agent frameworks can invoke this pattern whenever they discover a catalog skill and want it versioned beside project customs without copy-paste SKILL.md files. The skill is procedural infrastructure—shell, npx, and filesystem—not domain logic. It pairs with discovery workflows elsewhere in deer-flow (search, evaluate) and keeps custom skill paths consistent so Claude Code, Cursor, or Codex pick up the same tree. Expect to adapt the root-marker if your repo name differs.1.9kinstalls7Image GenerationImage-generation packages a Deer Flow–style Python workflow for solo builders who need repeatable visuals without a separate design tool chain. The agent loads a prompt from disk, optionally attaches reference images, and runs generation into a named output file. Reference images pass through validate_image using Pillow verify and load so bad assets are dropped with clear warnings rather than breaking the whole batch. Aspect ratio is parameterized for common landing and social formats. It fits builders shipping landing pages, docs illustrations, or in-app marketing assets who want the coding agent to own the scriptable path from brief to file. You should still supply API keys and endpoints per your environment—the readme excerpt focuses on validation and orchestration hooks. Pair with content and launch work when you need hero images or ad creatives, but catalog it under build integrations as the technical hook.1.8kinstalls8Consulting AnalysisConsulting-analysis is an agent skill for solo and indie builders who need consulting-grade research artifacts—not informal blog notes—when exploring markets, competitors, brands, or financial narratives. Phase one turns a research subject into a rigorous framework: chapter skeleton, per-chapter data requirements, analysis logic, and a visualization plan so you know what to collect before writing prose. Phase two runs after other skills or workflows supply data, synthesizing everything into a polished Markdown report with structured narratives, embedded charts, and strategic insights. It fits early journey research and later strategic updates when you already have inputs to merge. The skill is procedural and document-centric; it does not replace data pipelines or live APIs unless you pair it with collection skills. Expect intermediate familiarity with research scope and willingness to iterate on framework before final report generation.1.7kinstalls9Github Deep ResearchGitHub deep research is an agent skill framed as a rigorous report template for investigating open-source repositories before you bet a solo project on them. It standardizes repository metadata—stars, forks, issues, languages, license, topics, and activity dates—then forces an executive summary, phased historical narrative, and thematic analysis sections, each tied to inline citations so answers remain traceable in AI search surfaces. A dedicated architecture section expects a Mermaid diagram, which helps you compare system shape against your own stack quickly. Confidence level and research date fields make the output citable on Prism-style detail pages and in founder notes. Use it when you are scouting dependencies, studying competitor OSS, or documenting due diligence for a validate-stage prototype decision, and treat the result as a research artifact rather than automated install advice.1.6kinstalls10Deep ResearchDeep Research is a journey-wide agent skill from Deer Flow that teaches your coding agent how to investigate topics on the web with discipline rather than a single search snippet. Solo builders should invoke it whenever a user question needs current, multi-source evidence—definitions, comparisons, market facts—or whenever they are about to produce deliverables such as slide decks, landing copy, UI concepts, documentation, or multimedia outlines that must reflect real-world examples and up-to-date data. The skill positions itself as the default replacement for casual WebSearch on substantive queries and says to load it proactively before content generation starts. That makes it equally valuable while validating an idea, scoping a feature, writing launch SEO, or drafting growth content, as long as the task depends on verified external information. Intermediate complexity reflects the agent’s responsibility to plan searches, synthesize conflicts, and cite breadth. Pair it with planning or writing skills after research converges on a spec.1.6kinstalls11Chart VisualizationChart Visualization from Deer Flow is a generator-oriented agent skill that turns tabular intent into rendered charts—starting with area charts for continuous trends (often time) and bar charts for category comparisons and Top-N rankings. Solo builders wire it when they need agent-driven analytics assets inside research workflows, internal dashboards, or growth reporting without hand-authoring chart code for every slice. Inputs are explicit JSON fields: required `data` arrays with `time`/`value` (and `group` when stacking), or `category`/`value` for bars, plus toggles for stack versus group layouts, themes, textures, and axis titles. The skill emphasizes data hygiene—consistent time formats and filled gaps before stacked areas—and delivers both a public image URL and a machine-readable spec in `_meta.spec` so you can reproduce or tweak visuals later. It fits Grow analytics and Validate prototype demos; it is not a full BI platform or arbitrary D3 customization layer.1.5kinstalls12Skill CreatorSkill Creator (Post-hoc Analyzer) is an advanced agent skill for builders who maintain SKILL.md packages and run disciplined evaluations. After a blind comparator picks side A or B, this agent unblinds the result: it ingests the comparison JSON, loads both skills and their execution transcripts, and explains what the judge valued and where the losing skill fell short. The workflow is intentionally forensic—structural diff of instructions, tooling patterns, example coverage, and edge cases—so improvements are specific edits rather than vague “make it better” notes. Solo skill authors use it inside Deer Flow–style creator pipelines when iterating on procedural knowledge for Claude Code, Cursor, or Codex. It assumes you already ran paired executions and stored paths to transcripts and comparison output; it does not replace writing the initial skill or running the blind test harness itself.1.4kinstalls13Podcast GenerationPodcast Generation is a Deer-flow agent skill aimed at solo and indie builders who want to produce spoken episodes from machine-readable scripts instead of recording manually. The bundled Python workflow defines Script and ScriptLine structures, loads JSON-shaped dialogue, and calls Volcengine text-to-speech with environment-backed credentials. Parallel workers synthesize each line so male and female speaker parts can be generated faster before you assemble the final track. It fits the Grow phase when you are turning research, product updates, or newsletters into podcast collateral. You need Volcengine TTS credentials configured; without them the skill raises a clear configuration error rather than failing silently.1.4kinstalls14Web Design GuidelinesWeb Design Guidelines is an agent skill that turns Vercel’s Web Interface Guidelines into a repeatable UI audit on your codebase. Solo builders shipping landing pages, SaaS dashboards, or marketing sites install it when they want an objective pass over components and pages without hiring a separate UX review. Before each run the agent retrieves the latest rules from the canonical GitHub raw URL, reads the files or patterns you name, and checks every applicable rule. Output stays machine-friendly—short file:line messages aligned with the guideline document—so you can fix issues in Cursor, Claude Code, or Codex in one sitting. It fits after you have UI code worth reviewing: during ship-phase polish, before a launch, or when refactoring frontend work in build. It does not replace automated test suites or WCAG tooling with legal sign-off; it encodes interface best practices from a maintained public checklist.1.4kinstalls15Surprise MeSurprise Me is a journey-wide Deer Flow agent skill that turns a vague craving for delight into a single polished gift artifact. After you say “surprise me”, admit you are bored, or ask for something interesting, the agent reads every skill in your enabled catalog, picks one to three capabilities, and designs an unexpected but coherent mashup rather than a pile of disconnected demos. The workflow emphasizes emotional wow—juxtaposing research with presentation design, algorithmic art with storytelling, or data with canvas-style illustration—while optionally leaning on what the agent already knows about you from memory. It is built for solo and indie builders who want their agent stack to feel alive, not only utilitarian, and who are okay spending tokens on a creative detour that still demonstrates real procedural skills. Use it when you need inspiration across Idea, Validate, or Build without choosing a lane first; skip it when you already have an approved spec and need execution only.1.3kinstalls16Vercel DeployVercel Deploy is an agent-oriented bash workflow from the deer-flow kit that ships a built web project to Vercel through a claimable deploy API endpoint, returning structured JSON with preview URL, claim URL, deployment id, and project id. Solo builders use it when they want a fast path from a repo folder to a shareable preview without manually clicking through the Vercel dashboard for every agent iteration. The script inspects package.json dependencies in a fixed priority order—Next.js, Remix, Astro, SvelteKit, Nuxt, Hydrogen, TanStack Start, and more—so the upload matches Vercel’s expected framework preset when possible. Run it from Ship when tests pass and you need stakeholders or beta users on a live URL, or when an agent loop should close with deploy artifacts instead of stopping at local dev. It is not a full infrastructure skill: it assumes you already have a deployable tree and network access to the remote endpoint. Pair it with build and test skills earlier in the journey; after success, claim the deployment in Vercel to attach it to your account. Intermediate complexity reflects bash, JSON parsing, and trusting an external deploy service.1.3kinstalls17Academic Paper ReviewAcademic Paper Review is an agent skill for solo builders and technical founders who need publication-quality critique without hiring a domain expert for every preprint. It triggers on requests to review, analyze, critique, or summarize academic papers, research articles, preprints, or scientific publications—including paper URLs, uploaded PDFs, arXiv links, or phrasing like “review this paper” or “write a peer review.” The workflow parses the source, then produces balanced assessments covering summary, strengths and weaknesses, methodology assessment, contribution evaluation, literature positioning (including targeted literature search where applicable), and actionable recommendations grounded in evidence from the paper. Use it during Idea when surveying ML or systems research before choosing an approach, during Validate when evidence must support a technical bet, or during Build when documenting why an architecture choice is sound. It does not replace human peer review for submission, legal opinions, or clinical decisions.1kinstalls18Systematic Literature ReviewSystematic Literature Review is an agent skill for solo builders and indie researchers who need publication-style evidence maps without hand-curating every paper. It guides a structured SLR pipeline: read SKILL.md, query arXiv with tight keywords and optional category constraints, delegate metadata work to subagents, read the APA template, and write a dated markdown report. The deliverable reads like a real review—executive summary, cross-paper themes, agreement and disagreement, gaps, and annotated sources—so you can cite decisions in specs, blog posts, or model-selection memos. Use it when exploring a technical niche (for example diffusion models in computer vision) before you validate scope or commit engineering time. It pairs naturally with agent tooling stacks that already use task delegation and fixed output paths. Confidence is high on workflow shape from bundled eval expectations; treat network and script execution as operator-reviewed before production repos.871installs19Code DocumentationCode Documentation is a Deer Flow skill that turns repository analysis into professional docs matched to project complexity, following patterns seen in mature open-source and API products. It supports comprehensive README files, API references derived from code, architecture write-ups, developer and contribution guides, changelogs from history or release notes, and language-appropriate inline documentation. Solo builders use it after features land to make repos legible for future-you, collaborators, or agent handoffs; it also helps before Launch when SEO-facing project pages need accurate setup instructions. Trigger it when you ask to document code, create a README, generate API docs, or improve guides. Output can be a single README or a small doc set; pair with code review skills for accuracy checks rather than treating generated prose as authoritative without verification.827installs20Newsletter GenerationNewsletter Generation is an agent skill for solo builders and creators who need professional, well-researched email or web digests without hiring an editor on day one. You specify a topic, audience, or format—weekly tech roundup, industry briefing, multi-topic digest—and the skill researches the open web, curates credible sources, writes headlines and commentary, and ships a cohesive Markdown artifact you can paste into Beehiiv, Substack, or your CMS. It emphasizes readability, scanning, and consistent voice across editions so Grow content stays on cadence. Use it when Launch distribution needs a repeatable editorial product, when Validate tests whether an audience cares about a niche, or when Operate iteration means refreshing a subscriber-facing channel. The skill does not replace your positioning strategy; it accelerates production once you know what you want to say and who should hear it.788installs21BootstrapBootstrap is a Deer Flow conversation guide skill that structures first-run onboarding for a personal AI agent. Instead of jumping straight into tasks, it walks through lightweight Hello language selection, then a You phase that surfaces who the user is, what drains them, what to call the agent, and how they relate to it emotionally—assistant, partner, co-pilot, or digital twin. Those choices directly shape SOUL.md, so the same stack feels different for a founder vs an operator. Solo builders adopting agent workflows in Claude Code or similar environments use it when spinning up a persistent helper they want aligned to real pain points, not generic chat. It spans setup in Build and early Operate iteration when you refine persona after living with the agent. The skill is procedural: read the guide before the first reply, keep Phase 1 minimal, and merge rounds when users volunteer context to avoid repetitive questioning.565installs