Staying Current
Last reviewed
IntermediateWhat you'll learn
~12 min- Recognize that AI tools change constantly and plan for it
- Use a 30-minute weekly routine to stay current without burnout
- Distinguish between durable skills and ephemeral tool syntax
- Adopt a new model, feature or tool only after it passes your project's checks (its build and test commands) on one real task
- Prune instructions written for an older model once a new one passes
The tools you learned in this course will change. New models will launch, features will shift, prices will move. This lesson teaches you how to stay current without treating it like a second job.
The weekly routine below is when you look. The test further down is how you decide whether to adopt what you found.
The 30-minute weekly routine
Set aside 30 minutes once a week. Not every day — once a week is enough.
Step 1: Check changelogs (5 minutes)
Check what changed in the tools you actually use:
- Claude Code:
claude update, then read the changelog for what changed - Antigravity CLI:
agy update, thenagy changelogto read what changed (enterprise Gemini CLI:npm install -g @google/gemini-cli@latest) - Codex CLI:
codex updateor check the GitHub releases - Copilot CLI:
copilot update
Skim the changelog for features relevant to your work. Ignore features you don’t need.
Step 2: Try one new thing (15 minutes)
Pick one feature or technique you haven’t tried. Examples:
- A new model that launched this week
- A CLI command you’ve never used
- A prompting technique someone recommended
- A workflow you saw in a blog post or social media
Run it through the test below: one real task, then your project’s checks. Save a note about what worked and what didn’t.
Step 3: Tidy your standing instructions (10 minutes)
Your project’s standing instructions are the files the agent loads without being asked: AGENTS.md (or CLAUDE.md) and the skills folder. Open them. If you sent the same prompt more than once this week, save it there (Giving Good Instructions’s rule: reused prompts become rules or skills). If a new model passed the test, prune what it no longer needs (the last part of the test). Delete any rule or skill that no longer works.
Over a month, you’ll naturally stay current without the anxiety of trying to follow every AI announcement.
The tool-drift reality
AI tools change faster than any technology you’ve used before. In the time since you started this course:
- At least one major model update has probably shipped
- New CLI tools may have launched
- Pricing and free tiers have likely shifted
- Features you learned may have been renamed, improved, or deprecated
This is not a problem. This is the landscape. If you expected these tools to be stable like Microsoft Office, adjust that expectation now.
Tool churn sounds exhausting, and it can be if you try to track everything. You don’t need to. The orchestration skills you’ve built (clear thinking, precise communication, verification, iteration) are durable.
What’s durable vs. what’s ephemeral
Understanding this distinction saves you from wasting energy on the wrong things.
Durable skills (invest heavily):
- Breaking down problems into clear steps
- Writing specific, contextual prompts
- Verifying AI output against reality
- Knowing when to iterate vs. when to pivot
- Choosing the right tool for the task
- Managing project context and memory
Ephemeral details (learn as needed):
- Exact installation commands (
npm install -g @anthropic-ai/claude-codevs. whatever it is next month) - Model names and version numbers (learn the tiers: a deepest-reasoning flagship, a daily coding default, a fast and cheap model. As of 2026-11-19 the flagships are Claude Fable 5.2, GPT-6.2 Astra / Sol and Gemini 4 Carbon / Neon. Expect those names to be superseded by the time you read this, and use
/modelin your tool to see what it offers today) - Specific slash commands and keyboard shortcuts
- Pricing tiers and free limits
- UI layouts and settings locations
When a tool updates, you’ll need to relearn some ephemeral details. That’s a 15-minute exercise, not a course retake. Your durable skills transfer automatically.
When to switch tools
New AI tools launch constantly. Most are not worth your time. None of the reasons below is enough to switch on its own: each is a reason to run the test below, and the test decides.
Count the cost first: if you have habits built around a tool, list what you would have to relearn or rewrite before you spend a task on the test.
Run the test if:
- It solves a problem your current tool can’t. Example: you need multimodal input (images, diagrams) and your current CLI tool doesn’t support it.
- Your current tool is being deprecated. If the maintainers announce end-of-life, test the replacement before you start moving.
- The free tier or pricing works better for your usage. Cost matters, especially for personal or student use, and a cheaper tool still has to pass.
Not a reason on its own:
- It’s just newer. New doesn’t mean better for your specific needs.
- One demo impressed you. Demos are curated. Test it on your own messy, real-world tasks.
- Everyone on social media is excited. Run the test when you have a real task for it, not the day of the announcement.
The test: one real task, your project’s checks
The same test covers a new model, a new CLI feature and a new tool. Commit what you have, then start on a branch, so the trial stays off the branch you work on:
git switch -c try-new-modelFor a new model, type /model and pick it (all four tools have the command); for a new feature or tool, install or turn it on now, on the branch. Then send one real task from your to-do list, one you would have done this week anyway, through the usual loop (Think Before You Prompt’s loop: plan, approve, build, check), and end it with:
When you finish, run the build and the tests and show me the output.Which checks: the build, test and lint commands in your AGENTS.md (Think Before You Prompt’s step 5: commands proving it still works), or, once you have one, the CHECKS.md the Prove It capstone (Lesson 11.3) uses: a list of what must still work.
The bar is one line: adopt it if the checks pass. Commit the trial on the branch, prune (below), then merge it: git switch - takes you back to the branch you started from, and git merge try-new-model brings the trial in. If the checks fail, discard the trial with git restore --staged --worktree . and git clean -fd (that second command deletes every untracked file, so read git status first), then git switch - and git branch -D try-new-model, and keep what you had.
Then prune what the old model needed
A new model often does on its own what you once had to spell out, and a line written for the old one can get in its way. Still on the branch, get the list of those lines. In Claude Code, /doctor prompt-audit produces that list for you (Context Engineering shows what the audit flags). In the other tools, ask:
Read AGENTS.md and our skills. List every line that tells you how to think rather than what this project needs, such as "think step by step", a required visible scratchpad, or a rule in ALL CAPS. Do not change anything yet.Those examples are Let the Model Think’s list of habits to retire. Remove one listed line, rerun the same checks, and commit the removal if they pass. If they fail, git restore AGENTS.md (or the skill file) returns the file to your last commit, which brings the line back. One line at a time, so a failure points at the line that mattered.
You don’t have to pick one tool forever. Using Claude Code for deep codebase work and Antigravity for quick exploration is a valid strategy. Using browser chat for brainstorming and CLI tools for building is standard practice. The orchestrator picks the right tool for each task.
The multi-model advantage
Once you are comfortable with one CLI tool, consider using two or three together. Each frontier model has different strengths: Claude excels at deep reasoning and careful code, and Codex integrates tightly with the OpenAI ecosystem. Antigravity CLI offers Gemini. On a paid (non-trial) Google AI Pro or Ultra plan, it also offers Claude Opus 5.5 and Sonnet 5.5 in the same tool. Running the same question through multiple models and comparing answers is like getting a second and third opinion.
AI counsel: deliberation across models
Advanced practitioners go further: they use tools like OpenRouter (a single API that routes to dozens of models) or purpose-built consensus tools to have multiple AI models debate a decision before committing. Instead of trusting one model’s output, you present the problem to two or three frontier models and let them critique each other’s reasoning. The result is higher-confidence decisions on architecture, design, and complex problem-solving.
This approach is especially powerful for non-trivial decisions: choosing between two database designs, evaluating a security approach, or deciding how to structure a complex workflow. One model catches what the other misses.
The frontier CLI tool subscriptions run from about $10 to several hundred dollars a month each, depending on tier (see the plan tables in Module 6). Running two or three in parallel adds up, so choose tiers deliberately. That sounds like a lot until you compare it to hiring: a junior developer costs far more per month. A senior consultant costs more. If your department funds journal subscriptions, conference travel, or software licenses, AI tool subscriptions belong in the same category. Ask your supervisor: many universities and organizations are beginning to fund these as professional development or research infrastructure.
Staying informed without drowning
Follow (low effort):
- The official blogs of the tools you use (Anthropic, Google AI, OpenAI)
- One or two AI newsletters that summarize weekly developments
Ignore (high noise, low signal):
- Daily AI Twitter/X discourse
- “Top 10 AI tools” listicles
- Predictions about AGI timelines
- Debates about which model is “best”
Engage when relevant:
- Community forums when you hit a specific problem
- Release notes when your tool updates
- Case studies from people in your field
The goal is to be informed, not immersed. AI is a tool for your work, not your work itself.
🧬In Your Field: Biotechclick to expand
For researchers: Follow AI-in-science publications and preprints rather than general AI news. Tools that matter for bioinformatics (AlphaFold updates, new genomics pipelines, lab automation tools) are different from the mainstream AI conversation. Check BioRxiv and PubMed for domain-specific AI applications.
🏛️In Your Field: Government Devclick to expand
For government developers: Follow FedRAMP updates, the federal AI inventory, and your agency’s technology modernization office. Government-approved tools and deployment patterns change on a different timeline than the consumer market. The GSA’s AI Center of Excellence and digital.gov are useful references.
Key takeaways
- Tools change; orchestration skills don’t. Invest in the durable skills — clear thinking, verification, iteration — and learn tool-specific details as needed.
- The weekly routine is when you look. Check changelogs, try one new thing, tidy your standing instructions. Don’t try to follow every AI announcement.
- The test is how you decide. Give a new model, feature or tool one real task on a branch and adopt it if your project’s checks pass. Then prune, one line at a time, what the old model needed.
- Stay informed, not immersed. Follow the official sources for your tools and one good newsletter. Ignore the noise.
- You’re not behind. If you’ve completed this course and are using these tools at work, you’re ahead of the vast majority of professionals.
Your favorite AI CLI tool just released a major update with 15 new features. What should you do?