Prompt
Codex setup review by OpenAI’s GPT-6.2 guide
Maxing out Astra when Sol would do: model, effort, subagent and AGENTS.md problems found, each with a config fix to paste.
Prompt
Maxing out Astra when Sol would do: model, effort, subagent and AGENTS.md problems found, each with a config fix to paste.
Rules
Pixels crisp in your sprites but blurry on screen: textures on Nearest, whole-number scaling, camera and sprites rounded alike.
Prompt
UI bugs in your iOS app: main screens tapped through, Duo poses checked if its simulator exists, smallest changes suggested.
Rules
One stuck step eating your whole night: at most 3 attempts per step, then its error recorded and the next independent step begun.
Skill
For work that takes several agent sessions, a coordinator runs Claude and Codex in Herdr and checks changes before release.
Rules
Agent links that say localhost: Tailscale links to previews, review pages and dashboards that open right on your phone.
Prompt
UI hierarchy and density problems: every page screenshotted, then rebuilt on only a spacing scale, type scale and up to 5 colors.
Prompt
Unfamiliar features written from scratch, offers to continue: GitHub searched first, then the whole job done and a plain recap.
Rules
Questions you’d only rubber-stamp: work you can undo done and shown after, a stop at anything you can’t undo.
Skill
You keep fixing the same design mistakes: each correction in a session turned into a strict rule with reasoning.
Prompt
Broken, inconsistent or unneeded code: how it connects explained before any change, what it’s sure about fixed, the rest listed.
Prompt
Screens that only work when data loads fast and nothing fails: missing states and layout breaks fixed per component, design kept.
Prompt
Half your quota left at reset: dead code, duplicates and unneeded dependencies cut; scheduled tasks and old worktrees sorted.
Prompt
Fable 5.2 burns tokens fast: planning stays with it, the hardest parts may too, the rest goes to an Opus 5.5 subagent or Codex.
Subagents
Picking a model and effort per role: Haiku 5.5 finds files and runs tests, Opus 5.5 edits code, Fable on high only if stuck.
Rules
Rules hard-coded for older models work against Fable 5.2: delegation, think-less, style-limit and ‘never’ lines quoted with a fix.
Prompt
Explore burning big-model quota on every file read: wrong-model subagents and ill-fitting effort found, a config fix each.
Prompt
A feature works, but the code repeats itself: duplicates wiped out and the core logic in one place.
Prompt
Advice with no research behind it: 25-40 competitors indexed, recurring complaints ranked, then where your direction will die.
Subagents
$20 plan on Opus 5.5 by default: Haiku reads code, looks up docs; Opus advises before plans, on repeat errors, before ‘done’.
Prompt
Months of vibe coding left the code in knots: an HTML wiring diagram, the worst numbered in red, then fixes tested one by one.
Prompt
Before bed: each issue and TODO taken as far as it can go without you, then one HTML board of what’s done, stuck or your call.
Prompt
Bugs between a feature’s states: every state from UI to server on one HTML page, then a numbered list with file, line and why.
Prompt
What you call a module isn’t what the code calls it: every mismatch marked in red, those likeliest to cause a wrong edit first.
Rules
A summary saying all went well hides the agent’s choices: one screen of the ones that change the result, those to confirm in red.
Rules
Your agent often forgets to use what you install, so you call each by name: the best fit per task, with one line on which and why.
Prompt
Ask for faster code and you probably get a cache: timings from a profile, one matching fix each, its lines and expected speedup.
Prompt
Code takes minutes, then each PR waits on CI: the 5 slowest jobs split into queue, setup and run, and the line to change per fix.
Rules
Agents that both fix and review can’t stop going back and forth: stop after two failed fixes, review only this round’s changes.
Prompt
With Docker, ufw doesn’t help: visitors reach your site only through Cloudflare, and only you SSH in, over Tailscale.
Rules
Even a tiny change gets null checks at every layer: Codex checks data once where it comes in, then just uses it.
Prompt
Opus 5.5 can’t hear its own soundtrack: beats from librosa, the grid checked below 120 Hz, loudness measured on the MP4.
Skill
You want Van Gogh, Monet or Bauhaus moving: drawn in code, and a delivery check rejects any film where almost nothing moves.
Rules
The agent codes for races that never happen: one line makes it state the risk and the odds, then wait for your OK.
Plugin
Something in Claude Code’s interface you want changed: a prompt that finds an existing Mod or proposes the smallest one.
Skill
A tiny label above every heading gives AI away: one command lists what’s wrong with a page, another rebuilds it, content kept.
Prompt
Where your setup wastes tokens each turn: measured, each instruction line marked Keep, Move or Cut, applied on your OK.
Rules
Astra flags bugs that almost never happen and Opus fixes each one: bugs scored, low scores filed as issues, code left alone.
Skill
A landing page that looks designed: layout, copy, fonts and spacing handled after one round of questions.
Subagents
High effort mostly buys verification: a builder on low, a verifier on high, one CLAUDE.md rule runs each feature through both.
Prompt
Opus 5.5 videos that look like a slideshow: six camera moves by name with what each one does, plus seconds and speed to add.
Prompt
A vibe-coded site that loads slowly: images, fonts, frontend, database, APIs and servers, fixed in one pass and re-measured.
Subagents
Zero sales and you can still be sued: under-13s, Google fonts, session replay, marketing email, renewal terms, DMCA, all fixed.
Rules
Rules trimmed for Astra can drop checks Sol and Luna still need: Astra finds what was lost, each model’s rules kept apart.
Prompt
Whether your app needs them yet: six scaling fixes judged from your code, the file behind each answer, the monthly cost of each.
Prompt
Is your setup as cheap as OpenAI’s harness: pass or fail on each of its 5 rules, with the file and the one thing to change.
Skill
Pages that look AI-made at a glance: every tell at desktop, tablet and phone widths, FAIL before WARN, code untouched.
Skill
Making content solo: skills for topics, writing, visuals, video, sound and scheduling, installed and tried on one topic.
Prompt
Broken features, stuck users, slow steps: ranked by users affected, with log lines; the worst one already has a test.
Prompt
Projects that keep moving while you’re out: OpenAI’s Remote guide applied, what runs alone and what you decide by phone.
Prompt
A pile of sessions and nothing done: today’s merges told apart from busywork, then one sentence on tomorrow’s first task.
Rules
Keep hitting the usage limit: 7 days of session logs scanned, the 3 costliest habits and their share, rules to plug them.
Prompt
What you said you’d do later and forgot: past sessions searched, items ranked, each one a card that opens its own session.
Prompt
Plan written, no code yet: one message for first principles, adversarial review, ablation, Occam’s razor and loose coupling.
Prompt
An animation you love, a one-line idea, a character image or an art style: a prompt for each, one HTML file rendered, plus sound.
Prompt
You can’t find your AI-built site by Googling its name: crawling, every page and the whole site, fixed in one pass.
Rules
Rules from older models make Astra over-read or stop early: the triggers, required reading, checks and stop rules to cut.
Rules
Image options drawn one at a time burn your limits: a global rule puts them all in one numbered grid, then builds your pick.
Skill
Running more agents without leaving Claude Code: Opus 5.5 coordinates up to 10 workers through Herdr, and a Codex worker reviews.
Skill
Making a game: 70+ skills from character movement to shipping on Steam, for Godot, Unity, Unreal, Phaser, three.js and Roblox.
Plugin
Your agent hand-builds what’s already there: the repo, standard library, browser and installed dependencies checked first.
Skill
Your agent’s code hides bugs and its tests fool you: a skill it follows on every change, with crapper, dryer and mutator checks.
Prompt
You only know the surface of what you want: it asks until it’s 95% sure, then runs planning, build and review in order.
Config
Fable quota going fast: 3 settings send reading and lookups to half-price Opus subagents; Fable keeps the judgment.
Rules
What it needs from you is buried mid-report: one AGENTS.md section puts the to-dos, numbered, at the top of every report.
Skill
Spend tokens learning what AI can’t replace: 3 questions about your codebase, then MISSION.md, a curriculum and lessons on it.
Skill
No video model on hand: narration by TTS, music from free sample libraries, every frame in one of 43 styles, 30 to 60 minutes.
Prompt
Disk almost full: a read-only scan, a treemap of what development left behind, and a cleanup script only for what you tick.
Prompt
Before any app code changes: default, loading, empty, error and success on separate artboards, from a design reference repo.
Prompt
Requests you repeat, commands you stop or undo, how you check changes: ranked by how much each would help, your words quoted.
Skill
Motion effects you pick by clicking: 120 labelled samples in 5 packs; tick the ones you like and Claude Code builds them in.
Prompt
Plans made blind to your code and logs: connected to ChatGPT, GPT-6 Pro plans from real PRs and records, then briefs Codex.
Rules
A shot needing old film or old records: the Internet Archive’s free API searched, public-domain items only, each one logged.
Prompt
Codex and Claude Code slowly turn your README into a changelog: what it is and who it’s for first, history in CHANGELOG.md.
Prompt
Code too complicated, or busy a long time with no clear direction: one finds what needn’t exist, one rechecks the direction.
Prompt
Pages that could fit any startup: a Refero style as reference, a design plan you approve, then DESIGN.md and screenshot checks.
Prompt
A product video like the viral ones: the 5 of 703 prompts that fit your product, your pick remade as one HTML file.
Prompt
‘It’s not X, it’s Y’ is gone and it still reads like AI: Anthropic’s Fable 5.1 guide has a more direct one-line prompt.
Command
Today’s pushes to main, reviewed while you sleep: a draft PR per bug with a test that fails first, in the cloud, laptop closed.
Prompt
‘The last 24 hours’ in a scheduled task skips or repeats work: fixed windows, silent read failures and risky posting flagged.
Rules
Someone put 314 malicious skills on ClawHub: check the source, scan the code, question the permissions, in under five minutes.
Plugin
Three lines before something goes wrong: agents sweep the repo, three more try to knock each finding down, two must agree.
Prompt
Did the skill edit help: 10 real tasks, 8 yes/no questions, a blind judge subagent; keep only a fix that raises the score.
Prompt
A quality bar for the site it builds: each round screenshotted, scored on the Awwwards sheet, until every category hits 7.5.
Prompt
Still walking a senior AI through every step: tech lead, requirements, debugging, performance, frontend, backend, reviewer.
Prompt
A finish you’d pay a photographer for: headshot, cinematic still, travel, magazine shoot, old-photo repair, golden hour.
Prompt
One preference typed in three places: a goal wrapper, plus prompts for components, parallel runs, review, states and corrections.
Prompt
AGENTS.md past 200 lines gets followed less: 3 slimmer versions, each more aggressive, nothing lost, shown before changes.
Prompt
A cleanup that could delete your chats: items sorted Keep, History, Work, Clearable or Unknown, each command approved by you.
Prompt
Who finished what across several CLI agents: a goal each, read-only reviewers, worktrees for writers, one owner integrates.
Config
Which work deserves its own Bot: six roles with ownership, handoffs and limits, and what stays a Skill or a routine.
Rules
Astra burns quota faster than Sol: default-subagent rules gone, one line so it does the task and the checks itself.
Config
Claude deleted a config that ‘looked unused’: risky commands blocked, pushes ask first, routine ones run, deny always wins.
Rules
Taglines, helper text and labels on every screen: one line cuts them so the layout speaks for itself; one weighs UX, DX and AX.
Prompt
Pages that scream AI at first glance: every slop sign listed, a five-line direction for your OK, then fixes pass by pass.
Config
The cache expires without warning: the status line shows minutes left, then what the next message re-caches.
Rules
No users yet, but the agent adds compatibility code and migrations: skip both, then docs, tests, smallest change and cleanup.
Prompt
What in your project is worth animating: storyboards by Anthropic’s Motion guide, your pick built in HyperFrames.
Prompt
Users from Google search: 3 free tools people already search for, your pick built as an SEO page with one button to your app.
Subagents
Where a long task stands while you wait: an HTML dashboard of progress, blockers and open questions, built by a side subagent.
Connector
No memory of code, notes, sessions or long tasks: 8 repos in those 4 kinds, one of each installed on your OK.
Prompt
A vibe-coded site about to go live: design, mobile, every state, real-user flows and launch details, fixed in one pass.
Prompt
Trouble behind the page: keys, billing abuse, payments, monitoring and launch day fixed in one pass; what needs you listed last.
Skill
Hover stuck on phones, forms red before anyone types: 47 modern CSS techniques and a list of what your styles get wrong.
Rules
Stuck on style, components, motion, a demo video or polish: 8 sites sorted by that, saved so Claude Code knows which to use.
Skill
Your costliest Claude API feature: a test set first, then cheaper models and lower effort step by step while accuracy holds.
Plugin
Unreal games with Claude Code: Blueprint nodes added, wired and compiled in your open editor; levels, materials and effects too.
Plugin
Browser games with Codex: 9 skills for 2D and 3D, animation, models, menus and health bars; it plays the result to find bugs.
Subagents
A motion video in Claude Code: breakdown, storyboard and review subagents around a main session on HyperFrames; you direct.
Prompt
Which job is worth many agents: repo-wide checks, migrations or fix loops, planned on one directory with at most 10 agents.
Subagents
Fable quota idle while Opus 5.5 does the work: Fable called in before a plan, when an error repeats and before ‘done’.
Skill
Pages still look like default shadcn: a Stripe or Linear style DESIGN.md, every component restyled, a design system page too.
Command
A feature you like and can’t see how it’s built: taken down layer by layer, each finding with evidence, then a plan for your own.
Prompt
Settings carried over from Opus 5 cost more on 5.5: effort, subagents, MCP servers and CLAUDE.md checked, a small fix each.
Prompt
Try to break it yourself first: 16 questions on security, data, flows and environment, issues ranked by severity, then fixed.
Rules
Pages look AI-made since Opus hasn’t seen what you like: 3-5 screenshots become a DESIGN.md; add taste-skill and Impeccable.
Subagents
Astra eats a week’s quota: 6.1 Sol on high as orchestrator, on medium for 3 subagents, Astra only to review big changes.
Subagents
A week of sessions you can’t recall: local logs as an HTML calendar, no-commit sessions flagged, and a 3-line weekly report.
Rules
The agent keeps making one mistake: the right way in a helper, the old way banned by lint, an error that names the helper to use.
Connector
A fast browser for Claude Code: ChatGPT’s computer use as a local MCP server, clicking pages and comparing live with local.
Rules
Background you re-explain every new session: pulled from your last 20 sessions and written up as project instructions.
Prompt
Before a task goes to Codex: GPT-6 Pro finds what you haven’t thought through and writes a brief an engineer could take.
Prompt
Many agents open and the Mac bogs down: a dry-run map of processes and startup items, cleaned only where you approve.
Prompt
Your agent feels dumber over time: every instruction, skill, MCP and memory marked KEEP or REMOVE, cleared on your OK.
Skill
Skills for older models misfire on Fable 5.1: duplicate rules, dead commands and old model names listed, with a diff.
Skill
A skill written from memory drifts from how you work: compared with your past sessions, a dry-run diff, a recurring review.
Prompt
Before it implements on a guess: repo read, goal restated, assumptions and plan listed, then a stop for your approval.
Subagents
Bounded delegated work at max reasoning: a luna_worker custom agent, checked against your Codex version, diff shown first.
Skill
A system you have to explain: an HTML diagram you can search, trace up and downstream, compare across versions, export as PNG.
Skill
Quota unused before it resets: a UI/UX audit, a clickable code map, measurable speedups, a workspace and a Codex cleanup.
Prompt
‘Fix this bug’ hands you defaults: what your diff touches, an end-to-end goal, a red-teamed plan, a user walk, a thinking budget.
Rules
Before you change a repo: click any module for its callers, dependencies, tests and source evidence; a rule keeps it current.
Prompt
Pages designed from text come out plain: Codex’s image model draws the page first, then the coding model builds it, one line.
Rules
Opus 5.5 already thinks before it answers: rules that push extra thinking, dole out steps or stop to confirm, quoted with a fix.
Connector
Dozens of grep and read turns per question: the repo as one graph of functions, files and dependencies; dead code found too.
Rules
The agent over-designs and edits what it wasn’t asked to: think before coding, simplicity first, surgical edits, goal-driven work.
Prompt
A task longer than one session: progress, decisions, files, checks and next steps kept current for the next agent.
Prompt
Fable 5 burns tokens fast: planning and frontend stay with it, the backend goes to Codex by /goal; a redesign prompt too.
Skill
Your Claude Code subscription, used from Codex: Astra runs up to 10 Opus workers through Herdr and hands them the heavy work.
Prompt
Ten questions at once get ignored: assumptions, missing info, the usual mistake, then the one question that changes its approach.
Rules
Two agents, one rules file: plan first, smallest change, subagent roles, own the bug, verify before done, Lessons per correction.
Rules
Banning AI words one by one never ends: 6 rules from 1946 for your global rules file, plus prompts for docs, PRs and reports.
Prompt
Relaying conclusions between windows by hand: the task split, a worktree and outcome per agent, one executes, one reviews.
Prompt
Unused quota overnight: an hour before bed, 3-5 reversible, verifiable tasks from today, prompts ready, nothing changed.
Rules
Asked to change a config, it builds a framework: minimum sufficient change, reuse first, ask before widening, test narrowly.
Skill
Dashboards that look AI-made on sight: ui-ux-pro-max and frontend-design in one redesign prompt, then a four-angle review.
Prompt
Sample data proves little: the real component rendered in every state that matters, fixed only where a failure reproduces.
Config
Usage that lasts longer: Opus 5.5 as main model, Opus 5.5 on medium as default subagent; hard-coded model IDs found first.
Skill
Tired of ‘make it look better’: Impeccable, 21st Design and UX Audit turn it into specific changes, or one line with no skills.
Rules
A reviewer keeps flagging the same things: 50 recent PRs scanned, issues ranked by count, your picks written into AGENTS.md.
Prompt
A review with a designer’s authority: taste baseline, correctness fixed, cheap-looking parts found, big changes proposed.