Claude: Workflow

The workflow loop is Research → Plan → Build: Run /last30days [topic] first to ground the plan in current community data. Then /ce:plan, then /ce:work. Example: agent-browser vs Playwright decision informed by 78 Reddit threads + 76 X posts showing 82-93% less context token usage. (from every claude code hack mvanhorn)

Plan files are the checkpoint that survives everything: Context gets lost between sessions? Start a new session, point it at the plan.md, pick up where you left off. 70 plan files and 263 commits. Rule: unless it's literally a one-line change, there's always a plan.md first. (from every claude code hack mvanhorn)

/ce:plan accepts images: Paste screenshots of bugs, error messages, design mockups, or Slack conversations directly into Claude Code with Ctrl+V and run /ce:plan. Claude writes a plan from the image. (from every claude code hack mvanhorn)

Run 4-6 parallel Ghostty sessions: While one session plans, another builds, another researches. Requires bypass permissions — sessions need to run autonomously. (from every claude code hack mvanhorn)

Codex as overflow for token limits: /ce:work --codex delegates implementation to Codex when Claude Max credits ($200/month) run low. Claude for planning, Codex for heavy implementation. (from every claude code hack mvanhorn)

Granola MCP eliminates transcript copy-paste: Claude cross-references meeting transcripts against codebase AND prior strategy plan.md files — compounding context. One-shotted a product proposal from a 90-min lunch transcript. (from every claude code hack mvanhorn)

Plan files work for everything, not just code: Strategy docs, product specs, competitive analysis, articles — same voice → plan → iterate workflow. Compounding context from prior plan.md files. (from every claude code hack mvanhorn)

Parallel experiment execution (910 experiments in 8 hours) achieved 9x speedup over sequential approach, ~$300 compute + $9 Claude API (from parallel gpu autoresearch skypilot)

Agent spontaneously developed emergent optimization: screening on cheaper H100s, promoting winners to H200s — autonomous resource allocation without being told (from parallel gpu autoresearch skypilot)

Run /web-setup to grant scheduled tasks access to repos (Pro and Max plans) — tasks run on cloud infrastructure so local machine doesn't need to be running (from claude code scheduled tasks web setup)

Scheduled tasks accept a repo, cron schedule, and prompt to define recurring jobs — fully async cloud execution (from claude code scheduled tasks web setup)

Multiple Claude Code instances can be connected via inter-process messaging — one session can ask clarifying questions of another session working on related code (from claude code inter instance messaging)

First principles prompting: key phrase "strip each assumption away" forces layer-by-layer analysis; template is "Break [topic] down using first principles thinking. Start by identifying every assumption... Rebuild from only what remains" (from first principles prompt technique)

Dispatch enables directing coding work from anywhere — mental model is not "grind during gaps" but "design your day differently because work runs without you" (from dispatch async day design)

Claude Code workspaces can be purpose-built for non-engineering workflows like SEO content creation — packaging a workspace as a shareable repo (CLAUDE.md + skills + tooling) creates a reusable "job template" others can clone (from claude code seo workspace)

A three-step video-to-UI workflow: (1) record video of target UI, (2) upload to Claude to get a markdown description, (3) feed markdown to Claude Code to build it — using video bypasses the vocabulary bottleneck in describing UI components (from video to ui claude workflow)

The intermediate markdown artifact between vision and code gives the human a checkpoint to review the AI's understanding before committing to implementation (from video to ui claude workflow)

Task prompts should define the end state rather than the process, include uncertainty handling ("flag it instead of guessing"), and batch related tasks into a single session (from claude cowork context architecture checklist)

Subagents can handle parallel work within Cowork: "spin up subagents to research these 4 vendors simultaneously" (from claude cowork context architecture checklist)

An agent-first engineering mandate ("delete your IDE, stop writing code") forces teams to rethink entire workflows rather than incrementally adopting AI tools alongside existing habits (from chintan agent first engineering)

Engineering value shifts to "upstream intent" (knowing what to build and why) and "downstream validation" (verifying correctness), with agents handling the implementation middle layer (from chintan agent first engineering)

Multi-month projects shipping in approximately one day suggests the bottleneck was always implementation speed, not design or decision-making — agents remove that bottleneck (from chintan agent first engineering)

Creating paired Claude Code skills with different personas — one expansive (CEO-like) and one detail-oriented (eng manager-like) — lets you toggle between divergent and convergent thinking within the same workflow (from garry tan mega exit plan skill)

Running expansive and realistic skills back-to-back is an effective two-pass pattern: first generate ambitious possibilities, then pressure-test them with practical constraints (from garry tan mega exit plan skill)

A CPO at Pendo ($2.6B company) uses a single Claude Code command to plan his entire workday, suggesting Claude Code has matured beyond coding into general work orchestration (from claude code daily workflow vibePM)

Power users are choosing Claude Code's terminal-based workflow over GUI editors like Cursor, valuing composability and scriptability over visual convenience (from claude code daily workflow vibePM)

Claude Code /loop command enables non-technical business professionals to set up recurring agentic tasks with a simple cadence + task prompt — no coding required (from claude code loop business use cases)

Meeting prep loop (every 30min) that pulls attendee context, Slack threads, CRM notes, and emails before each meeting is a high-value pattern for managers and account executives (from claude code loop business use cases)

The "watch everything and proactively flag actions" pattern represents the most ambitious loop use case — a unified monitoring agent across email, Slack, CRM, and calendar (from claude code loop business use cases)

Claude Code can generate "Dives" — highly interactive data visualizations including pivot tables — from a single prompt, collapsing the traditional data prep → tool selection → chart building workflow into one step (from claude code data viz dives)

A CEO reports Claude Code as AI Chief of Staff roughly doubles productivity by unifying 6+ inboxes, managing overnight todo lists, and enriching contact records from Granola meeting transcripts (from ceo ai chief of staff claude code)

The "multiplayer todo list that works overnight" pattern represents async agent work — the AI processes the queue while the human sleeps, then the human reviews and reprioritizes in the morning (from ceo ai chief of staff claude code)

The "AI Chief of Staff" framing positions Claude Code not as a developer tool but as an executive productivity layer, with the agent pushing back on decisions and aligning time to goals (from ceo ai chief of staff claude code)

Claude Code is being used as a full outbound sales platform: 11 APIs, 72 automation scripts, handling campaign strategy, list building, and outreach — storing domain expertise in Skills files rather than hardcoding workflows (from claude code outbound sales agents)

Claude Code serves as a prototyping playground for agent systems before migration to Claude Agent SDK for headless production deployment — a "Claude Code to Agent SDK" pipeline is emerging (from claude code outbound sales agents)

Taskmaster enables multi-day continuous Claude Code sessions; the vast majority of users code less than 40 minutes/day, meaning long-running sessions represent a qualitatively different workflow requiring task management infrastructure (from taskmaster claude code marathon)

The /handover slash command solves session amnesia: Claude generates a summary of decisions, pitfalls, and lessons learned when the context window fills up, preserving institutional knowledge for the next session (from claude code handover command)

Audio notifications for Claude Code task completion solve a real workflow problem — developers kick off tasks and context-switch, needing an ambient signal when the agent finishes (from warcraft peon claude notifications)

Specialized harnesses (designer, marketer, etc.) compound in value as skills accumulate; the shift from generalist to specialist harnesses mirrors vertical specialization outperforming horizontal generalization (from claude code designer harnesses)

A practical meeting prep pipeline: cron trigger → calendar API → parallel research (Exa + Perplexity) → Claude formatting → email delivery, demonstrating Claude Code's strength as an API orchestration layer (from meeting prep tool claude code)

Claude Code's biggest UX pain for power users is premature task termination during long runs; Taskmaster monitors and forces resumption, acting as a "meta-agent that supervises the agent" (from taskmaster claude code persistence)

The overnight autonomous coding workflow (plan, kick off, sleep, review in morning) is becoming a standard power-user pattern, but its value depends entirely on task completion reliability (from taskmaster claude code persistence)

Claude Code is being used as an intelligent query layer on top of Linear: natural language queries about sprint velocity, cross-referencing tickets with GitHub to infer blockers, and back-solving point estimates into personalized hours per person (from claude code linear project management)

Adding "research the industry-standard approach to this problem, use it to guide yours" to prompts forces Claude to ground solutions in established patterns, reducing novel-but-fragile implementations (from claude code research industry standard)

Claude is being positioned as a content strategy tool replacing $2K/month ghostwriters — a shift from using AI for one-off writing to strategic editorial planning (from claude content strategy prompts)

Claude generates full PowerPoint presentations in under 5 seconds, collapsing the time-to-output for traditionally tedious knowledge work (from claude powerpoint generation)

Cowork with GSuite and Slack connectors enables Claude to automatically generate weekly status updates by scanning channels and schedule focus blocks based on calendar — proactive time-management agents (from cowork gsuite slack workflows)

The Claude Code creator interview crossed from dev tool into broader product consciousness, with product leaders recommending it as a "masterclass" (from claude code masterclass aakash)

Feynman (Claude-based) generates a cited meta-analysis from a research question in 30 minutes, replicates experiments on Runpod, audits claims against code, and simulates peer review (from feynman claude code for research)

Four-category task classification for AI automation: green (AI handles fully), yellow (AI preps 80%), red (needs human brain), gray (not actionable today) — creates clear human-AI boundaries (from jimprosser chief of staff claude)

The bigger productivity win from AI automation is cognitive load reduction, not time savings — starting the day in "decision mode" instead of "gathering mode" eliminates low-grade anxiety (from jimprosser chief of staff claude)

Stream Deck + Claude Code creates a physical interface for complex AI workflows — pressing a button triggers multi-agent orchestration (from jimprosser chief of staff claude)

AI Chief of Staff cost asymmetry: $5-10/month replaces $400-1000/month of virtual assistant work, with instant access to transcripts, email, and knowledge base (from jimprosser chief of staff claude)

Use @steipete's 'oracle' tool to bundle code and research papers, create an optimized prompt for GPT-5.4 Pro, and capture results back in your coding agent's terminal (from optimizing academic work with gpt 5.4 pro and coding agents)

Run tmux with 6+ Claude agents simultaneously, each with custom status bars showing auto-generated titles, percentage context usage, LOC tracking (+/-), and conversation IDs for cross-referencing between sessions (from claude multi agent productivity workflow)

Use a change-philosophy.md document instructing Claude to "examine the existing system and redesign it into the most elegant solution that would have emerged if the change had been a foundational assumption from the start" — forces architectural thinking over patching (from claude multi agent productivity workflow)

Use --chrome flag to spawn sub-agents that simulate user testing; results feed into a problem space doc → solution space doc → auto-iterate loop — user testing as a schedulable Claude skill (from claude multi agent productivity workflow)

Build session recording shortcuts into apps to collect full internal logs, profiler data, gestures, and state transitions; paste session ID for auto-debugging without manual reproduction (from claude multi agent productivity workflow)

Set up three core folders for compounding Claude context: knowledge folder (static stakeholder info, reference material), projects folder (research, drafts, artifacts that persist across sessions), and people folder (auto-updated from Granola MCP meeting transcripts) (from claude code file structure system)

The people folder compounds by Day 90: Claude knows communication preferences from 30+ real conversations — context no one-time prompt can replicate (from claude code file structure system)

Create a standup command that pulls simultaneously from GitHub, Linear, calendar, and task folder — analyses that would require clicking between 4+ UIs become a single command before your first meeting (from claude code file structure system)

Build a Feature Request chat (powered by Sonnet) that asks simple questions and turns team requests into thorough specs; scheduled Claude Code tasks pick up Spec Ready items nightly, implement them, and move to PR Ready with a link — a separate fresh-context task handles code review (from claude automated feature request workflow)

Claude Cowork automation stack is sequential: skills (single workflow as slash command) → plugins (chain multiple skills + connectors) → scheduled tasks (fire on timer) → dispatch mode (remote from phone) — each layer builds on the last, don't try to build the full machine on day one (from twitter link share hooeem)

Cowork's built-in skill creator automatically tests new skills vs. baseline (without the skill) and compares outputs side by side for iteration — quality evaluation is baked into the creation process (from twitter link share hooeem)

Morning Brief scheduled workflow: Claude builds your daily dashboard before you sit down — calendar summary, outstanding email actions, weather, industry news, draft email replies for review; alone worth the subscription price (from twitter link share hooeem)

LLM Council skill forces 5 AI advisors with different thinking styles (Contrarian, First Principles, Expansionist, Outsider, Executor) to argue about a question, then anonymously peer-review each other — breaks Claude's sycophancy on high-stakes decisions (from link share without context)

The "what did all five miss" question in the LLM Council peer review is the most valuable — every run catches something no individual advisor saw; reading 5 perspectives side by side reveals gaps nobody thought to mention (from link share without context)

Claude Cowork pointed at a "Taxes" folder organizes scattered W-2s/1099s/CSVs/PDFs, builds a Context.md briefing, generates a master bookkeeping spreadsheet (income/expenses/write-offs/home-office), produces a tax summary with refund projection, flags discrepancies, and writes a one-page accountant briefing — same Plan-files-work-for-everything pattern applied to bookkeeping (from claude cowork tax bookkeeping automation)

The /ss screenshot skill: Claude lists newest files in your screenshots folder and grabs the most recent (or N most recent with /ss 4); trailing argument is the action — /ss huh (explain), /ss fix (debug error or design issue), /ss do this (reverse-engineer + remix); saves ~1 hour/week (from claude screenshot skill visual processing)

/ultraplan shifts the planning phase from local terminal to the cloud: Claude drafts the implementation strategy remotely while your terminal stays free; review/edit in a browser with inline comments, section navigation, and visual diagrams; choose cloud execution for an immediate PR or "teleport" the plan back to local CLI — same token allocation as standard local planning (from claude ultraplan cloud development planning)

The Monitor tool turns Claude into an event-driven agent: background scripts wake the agent only when needed (log lines hitting an error pattern, a PR transitioning state) instead of polling — major token saver and the right architecture for long-running session work (from claude monitor tool background scripts)

Build a Claude Code skill that auto-generates and updates Excalidraw architecture diagrams from code or text descriptions — visual documentation as a side-effect of the workflow, not a separate task (from claude excalidraw architecture diagrams)

Premortem prompt: instead of "is this a good plan?" tell Claude "it's 6 months from now and this is already dead — tell me how it died" — flips Claude's training-induced optimism off because there's nothing to validate; returns most-likely failure, most-dangerous failure, the biggest hidden assumption, and a revised plan (from claude premortem technique decision making)

A self-improving Claude Code skill loop: 3-5 binary eval criteria, 10 runs with varied inputs, evaluator scores every output, identify failure patterns, rewrite the prompt, retest, keep the winner — a hook-writer skill went 32/50 → 47/50 overnight; works for hooks, briefs, ad copy, scripts, reports (from claude code self improving skills automation)

Boris Cherny dogfooded Opus 4.7 for weeks and reports feeling incredibly productive — productivity tips for 4.7 are emerging from inside Anthropic, not just the community (from opus 4 7 productivity tips)

Marcus Moretti's PM-as-one-person-team workflow: strategy.md (target problem, audience, 3-5 SMART metrics, 2-4 work tracks, rerun every few months) plus /ce:product-pulse running daily at 8am reading PostHog/Stripe/Datadog/database into ~/pulse-reports/ — replaces 60% of a PM's old week with two files and a cron (from ai agent pm workflow spiral every)

Now/Next/Later kanban with In Progress/Done is the only execution surface — no sprints, no standups, no PRDs, no backlog grooming, no stakeholder updates; the agent writes/moves tickets and keeps statuses live (from ai agent pm workflow spiral every)

Anthropic shipped a 14-minute video guide ("Agents → Structure → Tools → Execution → Systems → Money") plus a 33-page Claude Skills guide focused on persistent instructions and memory — first-party documentation for "what actually works" is now mainstream (from anthropic agent architecture guide)

ZenSpinner replaces Claude Code's spinner verbs with zen quotes — small tooling-as-mood pattern that's part of the broader trend of customizing the entire agent experience surface (from zenspinner claude zen quotes)

HyperFrames video creation through Claude Design takes only 2 prompts when you give Claude the HyperFrames skill — turns Claude into a motion designer for prototypes/slides/one-pagers; Opus 4.7 vision under the hood (from claude design hyperframes video creation)

The complete personal-knowledge-base workflow in three moves: create raw/, wiki/, outputs/ folders plus a CLAUDE.md schema file in 5 minutes, then point Claude at raw/ with the prompt "Read everything in raw/. Compile a wiki following CLAUDE.md rules. Create INDEX.md first." — 45 minutes of weekend setup becomes a continuously improving company asset (from claude personal knowledge base workflow)

Monthly knowledge-base health check: ask Claude to flag contradictions, find unexplored topics, and suggest 3 new articles to fill knowledge gaps — turns the wiki into a self-auditing system (from claude personal knowledge base workflow)

gstack transforms Claude Code from a solo assistant into a 6-specialist AI team — CEO, Eng Manager, Designer, Release Manager, Doc Engineer, QA Lead — installed in 30 seconds; the CEO agent challenges every decision ("why does this need to exist?") before coding, the Designer ships 4-6 variants and picks a winner, and the QA Lead runs real browser tests (from gstack claude code ai team)

Garry Tan's exact Claude Code setup is open-sourced as gstack — 23 opinionated tools for a structured development workflow, forkable as a complete role-based team harness (from gstack claude code ai team)

Multi-LLM development workflow uses Claude Opus 4.7 for feature planning and GPT-5.5 for plan review to catch issues before development starts, with Playwright running automated UX/UI validation after Claude builds and before human review; @conductor_build switches between Claude and GPT models in a single workflow rather than juggling separate interfaces, for ~$400/month full dev-team capability (from multi llm development workflow conductor)

A one-command skill converts a personal PM OS into a team OS without leaking personal context — turning individual productivity gains into team-wide compounding benefits where teammates query shared call summaries, decision logs, and analytics in natural language and get full reasoning back in 15 seconds with no human in the loop (from team os knowledge sharing architecture)

Claude Code's agent view (research preview) consolidates all coding sessions into a single unified list — session management moves toward tracking and revisiting coding interactions across multiple projects rather than per-terminal context (from claude code agent view sessions)

Claude Code can generate static HTML files for reports, explorations, code structure documentation, and mockups as primary workflow outputs, enabling a Google Docs-style iterative commenting system for HTML pages (from claude code interactive html comments)

Claude Code's interactive HTML skill launches a localhost server that allows in-browser commenting on generated pages, automatically updates content based on comments, and provides guided tours of changes — install by asking Claude Code to clone the relevant repo (from claude code interactive html comments)

Treat planning as a search problem: Sean Geng's plan-optimizer skill scores plans against rubrics, critiques them, and rewrites until score plateaus — stopping when improvements become noise rather than signal; particularly effective with Fable 5 (from claude plan optimizer iterative improvement)

Use Dynamic Workflow mode for complex tasks — sends out thousands of subagents to complete work previously taking months; combine with Ultracode mode which sets thinking to xhigh and lets Claude decide when to invoke Dynamic Workflows (from claude opus 4 8 upgrade workflow optimization)

A structured /goal prompt should define one measurable mission, current repo/context, hard constraints, priority, plan, done-when criteria, verification, output expectations, and stop rules, making agent tasks bounded and auditable before execution (from goal command prompt engineering structure)

Include rollback/containment in the VERIFY section and require the agent to state what could not be verified; verification is part of the workflow contract, not just a post-hoc summary (from goal command prompt engineering structure)

Most messy Claude Code workflows come from staying vanilla; installing claude-code-setup lets the official plugin recommend hooks, skills, MCP servers, subagents, and automations so the workflow becomes environment-shaped rather than prompt-only (from claude code setup plugin enhancement)