UX · CX · BX
A field guide for designers and builders
Use Claude properly.
Practical commands, tools and workflows, with sources you can check and starting points you can use. Build better habits with the Claude Code playbook, then find what you need in the guide.
Edition 29 September 2026 · Claude Opus 5.5 and Sonnet 5.5 · GPT-6 cost comparison
What’s new in Claude
The last four weeks, drawn from Anthropic’s own release notes and checked on 29 September 2026. Each card names its source and date. A release is not the same as availability on your account, so confirm in your own settings.
29 September 2026 Checked against Anthropic’s platform release notes. Benchmark scores are not reproduced here; test a model on your own work.
claude-sonnet-5-5, the second model in the Claude 5.5 family. $2 in and $10 out per million tokens, cache reads $0.20. On the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry and Claude Platform on AWS.Officialclaude-opus-5-5, built for long-running agentic coding and knowledge work. 1M token context by default, 128k maximum output and always-on adaptive thinking. $4 in and $20 out, against $5 and $25 for Opus 5. Fast mode is a research preview on the Claude API.Officialclaude-fable-5-1, successor to Fable 5 for long-running agentic coding, knowledge work and research. 1M context and 128k output. $10 in and $50 out, with cache reads cut to $0.25 per million tokens.Officialclaude-mythos-5-1, offered to Project Glasswing participants. Same limits and prices as Fable 5.1.OfficialSource: Claude Platform release notes ↗. Prices for these models are compared with GPT-6 in Cost per token.
Source: Claude release notes ↗
compact-2026-09-04.Officialdiagnostics object on a Messages request to opt in.Officialinline-tools-2026-09-15.Officialbio, frontier_llm and reasoning_extraction categories, at the rates of the model that ran.OfficialSource: Claude Platform release notes ↗
/config.Official/checkup prompt-audit.OfficialavailableModelsMatch and deniedModels let an organisation block specific models, so new releases stay blocked until listed.OfficialSource: Claude Code changelog ↗. This section is a dated snapshot. For the live record, use Follow what changes.
The Claude Code playbook
A route from a vague request to work you can trust. Start with one task, add reusable guidance when patterns repeat, and bring in automation only when it solves a real bottleneck.
Start here For a real feature or fix in Claude Code. No extra plugin required.
Brief
Name the user outcome, constraints and visible evidence of success. A screenshot or real example makes the target concrete.
Explore & plan
Read the existing implementation. Plan when the scope or approach is uncertain; make a tiny, clear fix directly.
Build a slice
Deliver one usable path through the feature. Keep unrelated changes out so the result stays easy to review.
Verify
Run relevant checks, try the important edge case and inspect the interface. Ask for evidence, including anything untested.
Review & save
Read the diff against the brief. Save a known good point in Git. Agree separately when the work should be published.
Source: Anthropic: best practices ↗
The five-step checklist is Pow Digital’s adaptation. Use the shortest process that still proves the result.
Help me deliver this change in the current project. Outcome: [what a person should be able to do] Audience and scenario: [who, when and why] Reference material: [files, screenshots, designs or source URLs] Scope: [what to change and what to leave alone] Constraints: [existing components, accessibility, data, dependencies] Done means: - [observable behaviour] - [important edge case] - [evidence I can review] Read the relevant files and project instructions first. Identify assumptions that could change the solution. For a substantial or unclear change, propose the approach before editing; make a small, clear fix directly. Implement a coherent slice. Run the relevant checks and inspect the result. Report what changed, the evidence, and anything you could not verify. Ask before publishing, sending messages or making destructive changes.
Review this change against the original brief and the current diff. Focus on correctness, missing requirements, regressions and accessibility. For each actionable finding, give the affected file or screen, the user impact, and evidence or a reproduction step. Separate confirmed problems from hypotheses. Do not invent findings to fill a quota. Check which tests actually ran and what they covered. Name any behaviour that still needs a real browser, device, user or service to verify. Do not edit files or publish anything. If there are no material findings, say so.
Use the smallest addition that solves the problem. These mechanisms have different jobs.
Adapt this to your repo and review every command before saving it as CLAUDE.md.
# Project guide ## Purpose [Who this product serves and what it helps them do.] ## Working commands - Install: [the actual command for this repository] - Run locally: [the actual command] - Validate: [the relevant test, lint or build commands] ## Conventions that are easy to miss - [Where shared components and design tokens live.] - [One important domain rule the code does not make obvious.] - [A repeated mistake to avoid, with a concrete example.] ## Completion Check the requested behaviour and relevant edge cases. Report the checks run and any unresolved limitations. Review the diff before handing it over. ## Boundaries Ask before publishing, sending external messages, deleting data or changing credentials. Never commit secrets. Keep this file concise. Replace placeholders with verified project details; remove anything that does not apply.
Source: Context management ↗
Source: Checkpoint limits ↗
Source: Parallel work ↗
Shan Raisshan’s Claude Code Best Practice ↗ is a community-maintained map of techniques, examples and further reading. Use it to explore how people work, then check product details in Anthropic’s documentation.
- Tips and tricks ↗: prompts, context and everyday working habits.
- Orchestration example ↗: study how reusable work is divided.
- Feature map ↗: follow the links through to official documentation.
- Always planning is the author’s preference. Tiny, clear changes can be direct.
- The roughly 40% context threshold is a heuristic, not a universal quality boundary.
- Popularity and elaborate agent setups are not evidence of reliability.
- Read instructions and scripts before adopting a configuration.
Reviewed 7 September 2026. Repository snapshot used for this review ↗. Examples are linked rather than installed. This guide is an independent adaptation, not endorsed by the repository author or Anthropic.
Source: Planning tradeoffs ↗
Designing with Claude
Bring the reference, design rules and user scenario. Use Claude to explore, implement and critique, then check the result against the real interface and the people who will use it.
Write down tokens, component choices, interaction states, tone and accessibility requirements in a design document. Reference that file in the task or project guide. The filename DESIGN.md alone is not a guarantee that an agent loads it.
Use existing components where they fit. shadcn/ui ↗ provides editable component source; Refero ↗ offers design references. Neither establishes that a design is usable or accessible.
/design-login authorises access; /design-sync converts and uploads a React design system for Claude Design. Review what you are sharing and check the supported environment in the official command reference ↗.
Share the exact frame or component URL, name the target route and reference your existing components. Ask for loading, empty, error and small-screen states. Compare the built result with the design and test keyboard use.
Visual similarity does not establish working behaviour. Check real content, interactions and component reuse before sign-off.
Source: MCP configuration ↗
Judge the result
A model’s usefulness depends on the task, tools, brief and review process. Start with what it costs, then compare outcomes with a repeatable check on your own work before changing your workflow.
29 September 2026 US dollars per million tokens (MTok). List prices from Anthropic ↗ and OpenAI ↗. Prices change, so confirm before you budget.
| Model | Tier | Input | Output | Cached input | Above 272K input tokens (input / output / cached) |
|---|---|---|---|---|---|
| Claude Fable 5.1 | Frontier | $10 | $50 | $0.25 | No surcharge |
| GPT-6 Astra | Frontier | $10 | $50 | $1 | $20 / $75 / $2 |
| Claude Opus 5.5 | Flagship | $4 | $20 | $0.20 | No surcharge |
| Claude Sonnet 5.5 | Balanced | $2 | $10 | $0.20 | No surcharge |
| GPT-6 Sol | Balanced | $2 | $10 | $0.20 | $4 / $15 / $0.40 |
| Claude Haiku 4.5 | Fast | $1 | $5 | $0.10 | No surcharge |
| GPT-6 Luna | Fast | $0.10 | $0.50 | $0.01 | $0.20 / $0.75 / $0.02 |
Pick a scenario or enter your own token counts. The chart prices one request on every model, using the rates above.
- Cost per token is not cost per task. Models differ in tokens used, retries and review time. Run the same task on two models and compare the whole job: see Compare two models on your own work.
- Long context. GPT-6 rates rise above 272K input tokens: input and cached input double, output is 1.5 times. Claude Fable 5.1, Opus 5.5 and Sonnet 5.5 keep one price across the 1M window. The calculator counts all input tokens, cached included, toward the threshold; confirm on OpenAI’s page.
- What is not modelled. Cache writes (Claude charges a premium), tool fees, regional pricing and subscription plans. Batch applies 50% to every rate shown.
- No scores here. Benchmark results are vendor-reported and depend on tools and settings. Judge on your own work.
Sources: Anthropic pricing ↗ · OpenAI API pricing ↗ · Claude Platform release notes ↗ (Opus 5.5 at $4 and $20, against $5 and $25 for Opus 5). Both vendors’ pages were read on 29 September 2026.
Pow Digital’s evaluation worksheet: choose a representative task and give each candidate the same inputs, constraints and time budget. Record the model, date, tools and revisions used.
| Test | Look for | Keep as evidence |
|---|---|---|
| A real feature | Required behaviour and edge cases work | Test output and a working example |
| A design implementation | Readable layout, useful states and keyboard access | Screens at target sizes and a manual check |
| A written deliverable | Accurate claims, useful structure and your voice | Source audit and a reader’s feedback |
| A repeated task | Consistency across several inputs | Failure rate, review time and total usage |
Public benchmarks answer narrower questions. SWE-bench ↗ measures software issue resolution; LiveBench ↗ evaluates several model capabilities. Read the evaluation method, model version and tool setup before comparing scores.
Flagship releases move faster than published comparison tables, and a score means little without a model version, a date and the tool setup attached. OpenAI announced GPT-6 Astra on 3 September 2026 ↗. Anthropic released Claude Fable 5.1 on 1 September, Opus 5.5 on 22 September and Sonnet 5.5 on 28 September, all listed in the platform release notes ↗ and in model configuration ↗. Availability varies by plan, region and provider. Confirm the exact model name in your own account before repeating someone else’s result.
Use Cost per token for the price side of the comparison. Quality has to be measured on your own task, because vendor-reported benchmark results depend on tools, effort settings and scaffolding you may not share.
The previous 2025-to-2026 scores out of ten were removed because the page did not provide a reproducible calculation or dated underlying results.
Resources
Understand the product surface first, then add only the context and tools your task needs. These notes and downloads are also usable as learning material in your own Claude.
Claude chat is useful for discussing and drafting work. Claude Code operates on a project with tools. Availability, approvals and connected data depend on the product, account and settings.
Start with the official extension overview ↗. For the practical choice between memory, skills, agents, hooks and MCP, use What belongs where?
Anthropic’s skills repository ↗ contains examples and skill packages. Read each skill’s scope, supporting scripts and licence. A community directory is a discovery aid, not a safety review.
Follow the provider’s setup guide. Check what can be read or changed, and remove connections you no longer need. Never paste an API key into a shared prompt or project guide.
Source: MCP setup and trust ↗
Use CLAUDE.md for recurring project facts and conventions. Check /memory to inspect saved context. Chat memory and Claude Code project memory should not be assumed to have the same scope or settings.
Source: Claude Code memory ↗
Subscription limits and API billing are different. Check the current cost guide ↗ for how your account measures usage.
Use /context to inspect the session. Compact a long, relevant conversation; start fresh when the task changes. Save a short handoff with decisions, files, checks and the next step.
Use /model to see the options your account supports. Compare the smallest suitable model on a representative task, including the time spent correcting it.
Track retries, tool output, review work and subagent usage as well as response length. Terse output can save output tokens, but a claimed percentage saving for a writing style is not a reliable total-cost estimate.
Current Claude Code documentation lists /cost and /stats as aliases of /usage. What is displayed depends on account and environment.
Source: Cost controls ↗
Source: Current command behaviour ↗
Each of these is a copy and paste prompt built on a well known way of working through a problem. Use these prompts to request questions, alternatives and explicit assumptions. Each is a small markdown file you can download and reuse.
Integration ideas, not tested end-to-end recipes. Follow the provider’s documentation for access, permissions, supported clients and current pricing.
These products can be ingredients in a workflow. Their presence here does not imply an official Claude integration. Check the provider’s documentation before you choose a connection method.
Recipes
Small tasks with a useful output, a starting prompt and a source to explore. Adapt them to your files and available tools; the prompt is a starting point, not an automatic integration.
Command reference
Commands below are for Claude Code. A slash command is typed inside a session; a CLI command starts with claude in your shell. Plugin commands only appear after the relevant plugin is installed.
This is a reference snapshot. Most entries were reviewed on 7 September 2026; additions were checked against the Claude Code changelog up to version 2.1.283 on 29 September 2026. Availability can depend on version, plan, platform and settings. Open the linked official reference and confirm the command in your own session.
How this guide is maintained
See what was checked, where the recommendations come from and how to recheck an item.
29 September 2026 Content review snapshot. This page does not update itself. Sections were last reviewed on different dates, listed below.
Official identifies the publisher: Anthropic or the product’s own documentation. Community identifies an independent source. Pow practice is our editorial advice or a reusable prompt. A label is not a security audit, a universal availability promise or evidence that an integration has been tested.
The Claude Code playbook and key command corrections were checked against official documentation. Design and Figma entries were checked against their official listings. The wider tool directory is a set of integration ideas; pricing and complete workflows have not been certified. Model prices and release facts were checked on 29 September 2026 against Anthropic’s and OpenAI’s own pages. Benchmark scores are deliberately not reproduced.
- Added a practical playbook with links to Shan Raisshan’s repository and the reviewed snapshot.
- Removed unsupported numerical rankings, traffic claims, precise token-saving claims and speculative release stories.
- Corrected /fork, /agents, /hooks, /design and the scope of the budget flag.
- Restored missing downloads and made learning prompts independent of a public hosting address.
- 17 September 2026: added Anthropic’s official course path and a note on confirming current model releases.
- 29 September 2026: added What’s new, list prices for seven Claude and GPT-6 models with a request calculator, six commands, four compatibility notes and six recipes. Refreshed every beginner session and starting prompt. Replaced the globe. Claims about releases and prices were read on the vendors’ own pages that day. The rest keeps its earlier review date.
Open the source attached to the item, check its version or account conditions, then verify the behaviour in your environment. If a command changes, record the date, observed change and replacement instead of guessing.
Sources worth keeping
Official references for product behaviour, then selected independent material for ideas and judgement.
Anthropic publishes free courses on its own training platform. Create an account at anthropic.skilljar.com ↗, then take these three in order. Each is free, ends with an assessment and issues a certificate of completion. The course account is separate from your Claude account.
Launching Cowork: it runs in the Claude desktop app, and the course opens with a setup lesson and a first end to end task. Anthropic’s course page now states that Cowork is becoming part of Claude itself, rolling out to Pro and Max, so newer menus may simply say Claude. The course content still applies. Verified 17 September 2026 against the course pages linked above.
Compatibility notes
Older advice may describe a different version. A legacy alias can remain valid even when a newer workflow is preferred.
Follow what changes
Go to the primary source for new releases. A feature announcement is not the same as availability on your account.