Oct 25, 2026

Over the past year, you’ve gotten better at explaining what you want, and the AI misreads you less. Lately the Streak redesign pulled the other way: you’ve been checking prototypes less and starting over more often.

Improve your prompts and setup

Vague prototype starts take the most time. Syncing your rules is the quickest fix, so it comes first. Five fixes below, all in your plan until Nov 8.

Did the fixes from your last review work?

  • KeepConfirm the plan before editing. You’re correcting the AI about as often as before, at 1.2 per 10 messages. The fix is only 2 weeks old, too soon to tell.
  • KeepSend research comparisons to a helper agentA helper agent works on the side and brings back only the result, so your main session stays short.. Sessions using one rose from 11% to 18%, but sessions crowded with lookups stayed about the same.
  • ReviseBrief before big prototype tasks. You haven’t tried this yet, so it’s smaller now: only longer prototype sessions. It’s fix 2 in your plan.
1

You retype the same rules across your tools

Claude Code · Codex · Cursor · claude.ai · ChatGPT Sure

What happens
You type the same rules by hand, because only some of your tools have them saved.
Evidence
You retyped rules 27 times in the last 4 weeks. “just show the diff, skip the recap” alone came up in 9 sessionsOne conversation with an AI, from your first message until you start a new one.. Codex and ChatGPT are missing all 4 of your rules, Cursor 2.
Why it matters
You lose time every session, and Codex, Cursor and ChatGPT still answer without your rules.
Which tools have which rules
RuleClaude CodeCodexCursorclaude.aiChatGPT
Confirm the plan before editing✓—✓✓—
Show changes, not whole files✓——✓—
Link official docs for how a tool works✓——✓—
Prototype stack: React + Tailwind✓—✓✓—

Try this: Add your rules to the tools missing them

Saves about 5 minutes a weekAbout 7 rules retyped a week × under a minute each.

Copy the block once. Codex: paste it into a new session and say “Save these as my rules for every project.” Cursor: Cursor Settings, then Rules. ChatGPT: Settings → Personalization → Custom instructions. Claude Code and claude.ai already have them.

- Confirm the plan before editing more than one file.
- Show changes, not whole files. Skip the recap at the end.
- Link the official docs when you tell me how a tool works.
- Prototypes use React + Tailwind.
2

Prototype sessions start vague, then restart

Claude Code · Codex · Cursor · claude.ai Likely

What happens
Your first message names what to build, but not who it’s for or what done looks like. You fill in the gaps over the next few messages, and sometimes start a new session.
Evidence
In the last 4 weeks, 16 of your 25 restartsA new session that opens with mostly the same request as one from the last two days, in the same tool. It usually means the first try went off track. were prototype sessions, including 4 Streak settings-screen sessions on Oct 14–15. In those 4 weeks, 44% of prototype sessions opened with a brief.
Why it matters
Each restart repeats the setup, and without a definition of done it’s hard to tell when a prototype is finished.
“build the settings screen for streak, make it look good and hook up reminders”

How 3 of the 4 settings-screen restarts opened. The next 5–8 messages each added a missing piece, like who it was for or the stack.

Try this: Brief before big prototype tasks

Saves about 30 minutes a weekAbout 2 restarts avoided a week × 15 minutes of repeated setup.

Starting a prototype that’s more than a quick tweak? Copy this as your first message and fill in the brackets. If you don’t know what done looks like yet, say so.

Goal: [what this prototype should let someone do]
User: [who uses it]
Done when: [what you'll check to call it finished]
Limits: [tools to use, what to leave alone]
Before building, list up to 3 questions you need answered.
3

Prototypes often go out without a last check

Claude Code · Codex · claude.ai Hunch

What happens
When a prototype looks done, the session often ends with no critique or testA check that the code really works, run by the AI or by you..
Evidence
3 prototypes went into the Streak beta with no check in the session, Oct 6–20. In the last 4 weeks, checking in prototype sessions fell from 43% to 23%, and Claude Code and Codex saved code they hadn’t tested 10 times.
Why it matters
Beta testers use what you shipPut it where other people can use it., so a broken screen does more harm there than in a sketch.

Try this: Check beta work before it ships

Means fewer fixes after launch

Only for work others will see. Paste this before you share it.

What are the 3 most likely ways this breaks for a real user, and what did you assume that I didn't tell you?

If your pr-review skillA skill is a saved set of steps the AI follows the same way each time. is fixed, type /pr-review first for a fuller check.

How much to check, by type of work
Type of workRiskLight check
Anything beta users will seeHighThe “3 ways it breaks” question, and pr-review once it’s fixed
Facts for decisions or docsMediumAsk for sources; spot-check one
Code changesMediumRun it, and ask what it assumed
Routine writing and formattingLowNone
4

Long sessions drift across projects and eat your usage

Claude Code · Codex Likely

What happens
Some sessions jump from one project to the next without starting fresh. Every message in a very long session makes the AI re-read everything before it, including files from the earlier project.
Evidence
In the last 4 weeks, 14 of 72 Claude Code and Codex sessions were very long (past about a novel’s worth of text) and took 61% of your usage. A message in them took 2.6× the usage of your typical one. 2 ran past 60 messages across the onboarding flow, a Streak sync bug and a pricing page.
Why it matters
You reach your limit sooner, and leftovers from the last project can steer the next one.

Try this: Start fresh between projects

Saves about a tenth of your Claude Code and Codex usageVery long sessions took 61% of that usage. Starting fresh could cut about a sixth of it.

When you switch projects or finish a piece of work, open a new session. In Claude Code, /clear does the same. If the next task needs context, first ask: “Summarize this in five lines for a new session.” Paste that into the new session.

5

Your strongest model does your simplest work

Claude Code Likely

What happens
Every Claude Code session runs on Opus, including quick questions and lookups a cheaper modelThe version of the AI you’re talking to. Stronger ones use up your limit faster. handles just as well.
Evidence
In the last 4 weeks, 19 of your 62 Claude Code sessions were short and changed no code. They took 14% of your usage on Opus.
Why it matters
Simple work on a cheaper model leaves more of your limit for the work that needs the strongest one.

Try this: Match the model to the job

Saves about a tenth of your Claude Code usageShort sessions took 14% of it × the price gap to a cheaper model.

Switch before you start. In Claude Code and Codex, type /model. In Cursor, use the model menu. If you correct a cheaper model twice, move up a level.

UseFor
Cheapest (on Claude, Haiku)Lookups, summaries, renames, formatting, quick questions
Middle (on Claude, Sonnet)Everyday building, edits and first drafts
Strongest (on Claude, Opus)Hard bugs, big design decisions, work others rely on

Your plan until Nov 8

TimeSavesNov 8 review checks
  1. Add your rules to the tools missing them5 minutes, onceAbout 5 minutes a weekRules retyped, and corrections in Codex and Cursor
  2. Brief before big prototype tasksA minute before any prototype that’s more than a quick tweakAbout 30 minutes a weekBriefed starts and restarts in prototype sessions
  3. Check beta work before it ships5 minutes per piece others will seeFewer fixes after launchOutput checking and untested changes shippedBetter with your pr-review skill. Fix it when you have 5 minutes.
  4. Start fresh between projectsA habit, whenever you switch projectsAbout a tenth of your Claude Code and Codex usageLong-session share
  5. Match the model to the jobA few seconds per sessionAbout a tenth of your Claude Code usageCorrections on cheaper models, usage on Opus

When these are done: the other skill fixes, then the repeats you could automate.

Keep your skills current

Three of your seven skillsA skill is a saved set of steps the AI follows the same way each time. give wrong information today.

Checked Oct 25: 7 skills you wrote, 17 you installed. Couldn’t read your Cursor skills folder, so any skills there aren’t counted.

Skills to fix

IssueFix
pr-review names a tool that doesn’t exist, so the visual check can get skippedFix 3 in your plan works better with it.Paste the fix (5 min)
release-notes still lists last quarter’s pricing tiersIts output goes into every changelog.Paste the fix (5 min)
beta-feedback-synth says the team has 4 engineers; your notes say 6In claude.ai, paste: “Change the team size in my beta-feedback-synth skill from 4 to 6.” (1 min)
api-docs is installed twice, and the claude.ai copy is olderTwo versions give different advice.In claude.ai, paste: “Help me delete the older copy of my api-docs skill.”
mermaid-diagrams is missing its main file, so it never loadsIn Claude Code, paste: “Reinstall mermaid-diagrams.”

pr-review: paste this into a new Claude Code session.

Update my pr-review skill. Its tool table names the wrong agent. Change the two column headings to "Claude desktop (built-in browser)" and "Claude Code, Codex, Cursor and others".
Then tell me in plain words what changed before you save it.

release-notes: paste this into a new claude.ai session.

Update my release-notes skill. Delete the pricing table and put this line in its place: "Read prices from docs/pricing.md. Never keep a copy here."
Then tell me in plain words what changed before you save it.

Worth a look

IssueWhy it mattersHow sureNext step
Your team details are copied into beta-feedback-synth, ui-review and metrics-checkThe copies already disagreeSureIn claude.ai, paste: “Move my team details out of those three skills and into one shared file.”
Your skills are set up for Claude, with none for Codex. Cursor’s folder couldn’t be read.Codex and Cursor can run the same filesLikelyIn Claude Code, paste: “Copy my skills to Codex, then check my Cursor skills folder.”
ui-review doesn’t know about Claude Design, and loads all three platform guides every timeIt misses a current tool and loads more than it needsLikelyIn claude.ai, paste: “Add Claude Design to my ui-review skill and have it load one platform guide at a time.”
No skill has a testA model update could change how a skill behaves without you noticingHunchIn claude.ai, paste: “Write 3 test prompts for my release-notes and ui-review skills.”
commit-message and pdf-tools are behind their source.You may be missing fixesSureIn Claude Code, paste: “Show me what changed in commit-message and pdf-tools since I installed them.”
Skills you wrote3 stale · 1 to review · 1 minor · 2 current
SkillWhereStatusWhat’s out of date
release-notesclaude.aiStalePricing tiers, description, no tests
pr-reviewThis computerStaleNames a tool that doesn’t exist
beta-feedback-synthclaude.aiStaleTeam size, no tests
ui-reviewclaude.aiReviewTool list; loads too much at once
metrics-checkclaude.aiMinorExample query names an old table
explain-like-pm, test-planclaude.aiCurrentNothing found
Skills you installed from others2 behind their source · 1 broken · 1 duplicate · 2 current · 11 not checked
SkillStatusDetail
commit-messageBehindYou have v2.7.0 from Jul 13. The source was rewritten through Oct 16.
pdf-toolsBehindSource updated Jul 2, after you installed it Jun 4.
mermaid-diagramsBrokenMissing its main file, so it never loads.
api-docsDuplicateThe claude.ai copy is older than the one on your computer.
skill-creator, tailwind-componentsCurrentMatch their sources.
11 othersNot checkedNo source link, so there’s nothing to compare against.

Update these from their source. Edits you make to them get overwritten.

How your habits are changing

Activity since your last review

Oct 12–25

You hit 1 usage limit.

By tool
ToolSessionsMessagesTypical message (characters)Corrections
claude.ai3620421220
Claude Code3124914833
Codex654968
Cursor654417
ChatGPT3151581
Skills and commands6 used · 1 unused for 30+ days
SkillRuns
ui-review7
pr-review5
beta-feedback-synth3
metrics-check2
release-notes2
explain-like-pm1
test-plan0 · unused 30 days

Found a wrong number or a broken chart? Report an issue on GitHub. Leave your prompts out of it.