Articles tagged with #Debug
Chrome (and now Firefox) blocks an otherwise-valid cert with NET::ERR_CERTIFICATE_TRANSPARENCY_REQUIRED. The cert was never logged to CT, or its SCTs are missing, stale, or stripped. Here is how to confirm it and fix it at the source.
Upload a CSV and ChatGPT lists weird columns, merges two into one, or treats the header as data. Delimiter and encoding detection is the usual culprit. Here is the fast fix.
You changed your ChatGPT email, got signed out, and now neither address logs you in. The fastest fix is the 6-digit code in the new inbox (or a password reset). Recover without losing chats.
Encrypted PDF uploads succeed visually but ChatGPT answers as if the file is empty. The rejection is silent — no error banner, just hollow replies.
Upload an XLSX and ChatGPT reports plain cell values fine, but formula-driven cells come back as the literal text '=SUM(A1:A10)' or as 0. Here is why openpyxl does this and the fastest fix as of June 2026.
You replaced the file but ChatGPT keeps quoting the old text. The fix: delete the old copy, watch the Skip / Upload anyway prompt, rename to bust the cache, or open a new chat.
Scanned or handwritten PDFs return 'No text could be extracted from this file' or hallucinated content because the extractor only reads the text layer. Fastest fix: upload page images, not the PDF.
ChatGPT read your JSON as a wall of text and string-matched an answer instead of running pandas. Here's how to force the code sandbox and get real aggregates.
ChatGPT stopping mid-output is usually the response length budget, a structural stop the model chose, or a dropped stream. Type continue first, then fix the real cause.
Upload a big spreadsheet, ask 'how many rows are there,' and ChatGPT says 1000 — that's sampling, not counting. Here's how to force a full-file analysis on every row.
The reply streams half a sentence then freezes. Usually a dropped SSE connection, a backgrounded tab, or a backend hiccup. Refresh first, then diagnose by bucket.
Upload a ZIP and ChatGPT sees only the filename, never the contents. The fix: trigger Data Analysis with an explicit unzip prompt, or extract locally and upload the files. Verified June 2026.
Claude shows the attachment tile but no preview thumbnail. Usually a render lag, an unsupported format (HEIC/TIFF), or a stale cache. Fastest fix and full diagnosis here.
A Claude citation link 404s or lands on the wrong article. Usually URL reshaping, an expired snapshot, or a paywall — how to rescue the real source and verify the fact.
A PreToolUse hook keeps rejecting Edit calls, even safe ones. The fix is almost always exit-code 2 logic, the deny-JSON format, stdin parsing, or matcher scope.
MCP tool calls in Claude Code time out or hang while a direct curl works. Usually stdio stdout pollution, the wrong startup vs tool timeout, a never-sent initialize, or a proxy that buffers SSE. Fixes verified June 2026.
A long Claude Code reply cuts off mid-sentence. Fastest fix: tell it to continue from the last line. Real cause is usually auto-compact, the per-reply output cap, or one huge tool result eating the budget.
The orange/grey cloud in Cloudflare DNS decides whether traffic hits the edge or your origin. The wrong state breaks SSL issuance, WebSockets, or origin masking. Here's how to pick the right one.
Cursor reopens with an empty chat list, only the latest thread, or stuck on 'Loading Chat'. Usually a workspace-hash change, a botched update migration, or a corrupted state.vscdb. How to recover and prevent it.
Cmd-K (Ctrl+K) spins forever or shows "Taking longer than expected…" with no diff. The fast fix: Reload Window or update Cursor. Walk the six verified causes.
Cursor's Extensions panel shows a network error, empty list, or a stuck spinner. Cursor uses Open VSX, not Microsoft's marketplace. Pick the right cause and fix it fast.
Added an MCP server in Cursor but it stays disconnected, or no tools appear in chat. 90% of the time it's a bad PATH, a JSON typo, or a stdout leak. Diagnose and fix in five minutes.
Cursor uses system Python instead of your project venv, breaking imports, linting, and agent edits. Fix the interpreter picker, venv discovery, shell PATH, and the new Python Environments panel.
Cursor's SSH Remote session disconnects while you edit, losing agent context and unsaved buffers. Fix it with the right keepalive values, and check for the v1.0.51 reconnect bug.
Cursor Agent runs terminal commands with no confirmation even after you turned auto-run off. As of June 2026 it's almost always Run Mode, a wide allowlist, or a workspace permissions.json. Fix it in a minute.
You turned on DNSSEC and validating resolvers now return SERVFAIL. Read the EDE code, regenerate the DS from the live DNSKEY, replace it at the registrar, or pull the DS to roll back. Verified June 2026.
IPv4 users reach the site fine but IPv6 users get timeouts. Either AAAA is absent, points to a dead address, or your firewall blocks v6. Fix dual-stack.
Delegated a subdomain to nameservers inside itself and resolution fails? The parent zone needs glue A/AAAA. Here is the dig diagnosis and the exact registrar fix.
Monthly image/video/music quota gone in a day. The three habits that burn fastest, with current per-second credit costs and the exact toggles to fix.
A normal-looking image prompt gets refused — one word tripped the filter. Binary-search the trigger and rewrite it in 60 seconds.
Subject lit from the left, sun on the right, shadows pointing nowhere consistent. Pin one key light by direction, quality, source, and color temperature.
You wanted a photograph, got plastic-skinned CGI. The fast fix: strip render keywords, drop CFG/guidance, switch to a true-photo model. Five stacking causes, six steps.
Skin is highlighter orange, grass is neon lime. Style words, CFG, and color LoRAs stack into a billboard. Strip the boosters, generate flat, then push saturation in post.
Skin looks plastic, jaundiced, over-smoothed, or doll-like. Beauty LoRAs, color-cast lighting, missing negatives, and Flux's default smoothing are the usual culprits. Here is the diagnosis and fix order.
Arm bends backward, neck twists past human range. Feed an OpenPose skeleton via ControlNet, use concrete action verbs, pull the camera back, and pick an anatomy-strong checkpoint.
A watermark-shaped smudge or fake signature in the corner is training-data residue, not a real watermark. Negative-prompt it (SD only), crop the corner, or inpaint it out.
Flat tables, sideways stairs, oversized heads. Open the prompt with a 5-7 word camera block (focal length + viewpoint + perspective style). Works in Midjourney, ChatGPT Images, Nano Banana, and ControlNet.
An AI agent recommended an abandoned npm/PyPI package. Triage it in 3 seconds with npm view, swap it for the modern replacement, and stop it recurring.
All green, ship it, prod breaks. The tests covered the happy path only and mocks shielded the real branches. Five causes and the fastest fix.
Clip cuts to black before the motion completes. Fix it by adding a settle beat to the prompt, setting an end frame, and budgeting 1-1.5s of ending time.
Two clips feel hand-cut with scissors? Match clip A's end frame to clip B's start frame, unify color, and add a short dissolve plus a J/L cut. Step-by-step fix.
Same shot, but the tint suddenly turns colder or redder. The model re-guesses color temperature between frames — pin it with a Kelvin spec and unify with a LUT or shot match in post.
Red hat in clip 1, blue in clip 2, gone in clip 3? Multi-clip generation has no shared anchor. Lock one tagged reference image and chain last frame to next first frame.
A chunk goes black mid-clip, or the file freezes. Usually it's a codec/player issue, not a bad render. Diagnose with ffprobe, then re-encode with ffmpeg.
30 minutes at 0%? Hard-refresh first (often it already finished), cancel truly-hung jobs, free your concurrency slot, submit off-peak UTC, or fall back to a lighter platform.
Claude Code swept your `.env` or API key into a commit. Rotate the key first (always), purge git history second, then block it for good with a deny rule, scanner, and push protection.
You asked for one function; the PR ships five "might be useful later" siblings nobody calls. Bound scope in the prompt and CLAUDE.md, auto-delete dead exports with knip --fix, and lock it in CI.
Claude Code turns tests green by adding .skip, deleting assertions, or loosening matchers. Block test edits with a PreToolUse hook, diff test files separately for skip markers, treat any test change as its own PR.
Firestore reads, Functions invocations, Hosting bandwidth — each free-tier metric has a hard line. Here's how to find which one tripped and stop the bleeding.
Firestore throws FAILED_PRECONDITION because a multi-field where + orderBy needs a composite index. The fastest fix, the CLI workflow, and how to confirm it's built.
First request takes 1–5s while warm ones are 100ms. Diagnose the real cause (heavy imports, top-level await, region) and cut boot time.
`remaining connection slots are reserved` and `too many clients already` — switch serverless to the port 6543 pooler, the right way.
The visible "Published / Updated" date does not match `datePublished` / `dateModified` in JSON-LD. Why this is more than a cosmetic bug.
Rich Results Test passes but Search Console flags Breadcrumbs, or the SERP trail looks wrong. Diagnose the real cause and align JSON-LD with the visible UI (June 2026).
You set canonical to a new URL, but weeks later Google still ranks the old one. Canonical is a hint, not a command — here's how to align every signal so Google switches.
A crawl flags hundreds of pages sharing one meta description. There's no ranking penalty, but it quietly costs you clicks. Here's the targeted fix, not a weekend rewrite.
FAQ rich results vanished or Rich Results Test errors on FAQPage? Google deprecated the feature in May 2026. What that means and how to keep your schema valid.
Set hreflang for en and zh but a validator flags missing x-default? What x-default does, the 6 configs that break it, and copy-ready fixes.
When `<meta name="robots">` and the `X-Robots-Tag` HTTP header conflict, Google merges them and takes the most restrictive value. How to diagnose it with curl and keep both in sync.
You added `<meta name="robots" content="noindex">` weeks ago but the page is still in Google. The six reasons, in hit-rate order, with the curl checks to tell them apart.
A template default, env leak, or CDN header left `noindex` on pages that should be public. How to find the real source, remove it, and force a clean re-crawl.
Your sitemap.xml lists URLs that render `<meta name="robots" content="noindex">`. Google indexes some and not others. Why it happens and how to make both signals agree.
A layout refactor or framework upgrade broke your JSON-LD site-wide. Map the Rich Results Test error to its cause and fix the single field that regressed.
Your X/Twitter link shows no image even though og:image is set. The 6 real causes, how to tell which one you have, and the fastest fix for each.
Your VideoObject JSON-LD passes Rich Results Test but Search Console shows no indexed video. The real cause is usually thumbnail, embed URL, or video prominence — here is how to find which.
Your WebSite JSON-LD declares a name / URL that doesn't match reality, so Google shows the wrong site name above your homepage. How to align it — and why the SearchAction half is now dead.
Your host says A, a tutorial says CNAME, and the apex won't resolve or email broke. A web-verified cheat sheet for apex vs www vs subdomains (Vercel, Netlify, GitHub Pages, Cloudflare).
Code is on the page but the ad spot is blank. The fastest fix, plus the 7 real causes ordered by how often they break, with exact checks for each.
Troubleshoot the common AdSense placement problems — clutter, double-stacked units, layout shift — when Auto Ads and manual placements conflict. For the strategic comparison, see the linked guide.
What "low value content" actually means in 2026, how to diagnose which bucket your site is in, and the recovery steps that get sites approved on reapply.
An AdSense slot renders blank with "no ad served." Diagnose with DevTools, then fix inventory, size, policy, or ads.txt causes.
An AdSense policy warning is a countdown. Decode the exact Policy center category, fix the cited URLs only, then Start review process — without burning your limited reviews.
adsbygoogle.js returns 200 but the ad slot stays empty. Read data-ad-status, rule out account review, and fix the wiring in minutes.
Your AdSense site is stuck on "Getting ready" or "Requires review." Here's how to tell a normal queue from a real problem, and what to fix.
PageSpeed tanked after adding AdSense? It's almost always 3-4 wiring issues, not the ads. Fastest fix: reserve ad slot height (CLS) and async the script. Full diagnosis + 2026-current code.
Your agent keeps trying the same fix or oscillates between two states. How to spot the loop in 30 seconds and kill it with one prompt.
The agent said 'rolled back' but the migration, the dist/ bundle, a Stripe webhook, and a PostHog flag are still live. Here's the six-domain checklist and the exact inverse commands that finish the job.
The cut between two AI clips feels jarring. Match motion, color, and pacing, or let the model render the transition with start/end keyframes.
AI added a dependency without committing the lockfile, or used a Node feature your CI image doesn't have. Reproduce it in 5 minutes and pin the gap shut.
After an AI run, npm run build fails. Locate the bad change in 10 minutes with diff triage, tsc --noEmit, and targeted git restore — no full reset.
An AI coding agent rewrote .env with placeholders or deleted it. Recover the real values, then hard-lock the file so it never happens again.
Six-finger hands still slip through in 2026. Fix the framing, run a hand-specific inpaint pass, and swap models — exact steps for Midjourney, SDXL, Flux, and Imagen.
Agent imports `src/utils/superhelper.ts` or a package that was never created. Find every bad import, then fix or replace it.
You asked for anime, the model leaks photoreal skin, eyes, and lighting. Anchor a specific anime reference, strip realism cues, and match the checkpoint's own tag system.
The background steals attention from your subject. Name a background style explicitly, contrast the color, crop tighter, then inpaint or one-shot-edit the rest away.
Same description, different face every run. Lock identity with a reference image (Omni Reference, gpt-image-2, IP-Adapter FaceID), a fixed seed, and unique markers.
Too many objects fighting for attention? Cut to one hero, add shallow depth of field, and use negative space. Fastest fix and per-tool steps for Midjourney, ChatGPT, Gemini, and SDXL.
Output is blurry, soft, or plasticky? The fix is almost always resolution, sampler steps, or a prompt with no detail anchors. Diagnose in 60 seconds, then fix.
The model ignores half your prompt. Usually it's token order, conflicting styles, or negatives written inline. Here's the fix path by model.
Subject looks flat, harsh, or amateur? Name the light setup with cinematography vocabulary (direction + quality + source + temperature). It beats adjectives every time.
Your AI series images look like 8 different artists made them. Fix it with a fixed style header, seed locking, sref/oref reference, palette limits, LoRA, and image-to-image chaining — updated for Midjourney V8.1 and gpt-image-2 (June 2026).
Got 5 unrelated images when you needed 5 cohesive ones? Pin a numeric Midjourney sref code (or one IP-Adapter reference), lock the seed, and freeze the model. Verified June 2026.
Starts as image A, ends as someone else. Lower motion, cap clip length, and lock identity with reference + text anchors to stop the drift.
You set 9:16 and got a square. Per-platform aspect syntax, supported-ratio lists, and UI overrides for Midjourney, ChatGPT, Gemini, SDXL and Flux, verified June 2026.
Mouth movements don't match your separately generated audio. Diagnose the cause and pick the right fix: re-render, post-sync, or compose around it.
Your branch plus AI edits conflict with main. Resolve it without losing the AI's real improvements, and stop mistaking format noise for logic conflicts.
An agent ran an install or removed deps and now your lockfile is at war with the team's. Throw it away and regenerate — here's the exact path per package manager.
Photorealistic prompts come out looking plastic and fake. Fix it with concrete lighting, skin-texture, lens, and negative-prompt cues — verified June 2026.
AI product shots read as fake when lighting is too symmetric, the contact shadow is missing, or reflections don't match the material. Here is the prompt-and-model fix path.
Your agent wrote `UserList2.tsx` next to the original instead of editing in place. Diff, merge, delete, and fix imports without breaking the build.
An AI agent "simplified" your code and quietly deleted a branch real users depend on. Bisect to the commit, cross-check prod logs, restore just the deleted hunks, and add a regression test.
Undo bad AI edits without losing the good ones: pick the smallest git undo, and recover lost commits with reflog before the 30-day clock runs out.
AI code won't type-check (TS2322, TS2339, TS2554)? The exact tsc loop, error-code table, and agent rules that fix it in one pass.
AI image text comes back as nonsense like RESPCT or OPEM. Pick a typography-strong model (GPT Image 2, Ideogram V3, Nano Banana 2) or add the text in post. Step-by-step fixes for June 2026.
Background strobes between shapes or styles. Fix it by simplifying the background, using one motion source, dropping Motion amount, or switching models.
Static-shot prompts come back wobbling like a handheld iPhone clip. AI video tools default to gentle handheld motion. Force a locked frame with the right static clause, the camera control set to Static, and the motion slider at zero (plus a negative prompt in Kling/Hailuo).
Same character, but their face morphs mid-clip. Lock one canonical reference, use the model's built-in character feature (Runway References, Kling Elements), keep clips short, and drop motion a notch.
Frame 1 is your character, frame 90 is someone else. Anchor identity with a bound reference image, keep face-critical beats short, and drop motion. Fixes for Runway, Kling 3.0, Pika, Veo, Sora.
Your output looks nothing like the input image. Raise image fidelity (Kling Relevance / CFG), cut the text prompt to motion-only, and feed a source at output resolution or larger with the subject filling 40%+ of frame.
AI video breaks the moment things move: extra fingers, morphing faces, vanishing objects. Seven web-verified tactics to make Veo, Kling, Runway and Sora clips far more consistent, updated June 2026.
Legs pump forward but the body slides backward, water pours up, hands phase through objects. Verb is ambiguous and physics aren't simulated. Specify direction, target, and start/end pose.
Limbs pop and subjects teleport between frames in Runway, Kling, Pika, or Veo. Fix it: one action per clip, anchor the start pose, render at native fps.
Everything moves at once and the clip feels chaotic. Drop the motion strength, strip extra verbs, pick exactly ONE moving layer (camera OR subject), and add a static-camera negative prompt. Tool-by-tool settings for Runway, Kling, Pika, and Luma as of June 2026.
Clips drag or rush. Generate 2-4s clips, cut on beats with markers in your NLE, and vary clip lengths so it stops reading as a slideshow.
Bottle stretches, packaging warps, logo bends in AI video. Treat product shots like macro photography: high-res reference, lowest motion, 5s clips, locked camera.
Same room prompt, two different sets? Anchor the scene with one canonical wide shot, a named location reference, identical lighting language, and a shared prop list — then color-match in post.
Your subject starts as one person and ends as another. Fix it with a native character reference (Kling Element, Runway References, Sora Cameo), specific identity markers, and shorter shots.
Your AI clip is a still with 5-pixel parallax. Add action verbs, name ONE camera move, and raise motion intensity — the three fixes for under-motion in Runway, Kling, Pika and Veo (June 2026).
Liquid pours sideways, cloth defies gravity, balls roll uphill. AI video models have no physics engine. Reframe the shot, switch to a stronger model, or composite real footage.
Subject dead-center, no depth, no leading lines. Direct composition like you direct lighting: thirds placement, three planes, camera angle, depth-of-field cues.
Pages work in dev but 404 in prod — almost always a build.format, trailingSlash, or output-mode mismatch with the host. Copy-paste fix path.
Sign-in redirects to localhost or the wrong domain, or the provider errors with redirect_uri_mismatch. Add the exact callback URI to the allowed list and drive it from env.
Moved to a new domain but pages still emit the old canonical — fix hard-coded URLs, CDN cache, sitemap, 301s, and Search Console.
Search Console indexes the wrong language because your EN/ZH canonicals point at each other. Fix: make every locale self-canonical and let hreflang declare the alternates.
Switched accounts but ChatGPT still shows the old one. Usually the native 2-account switcher cap, a cached session cookie, SSO auto-select, or a Business workspace lock.
A conversation shows on the web but is missing in the app (or vice versa)? The cause is almost always a mismatched account or workspace, not lost data. Here's how to find which one in under two minutes.
Login completes, ChatGPT redirects to chatgpt.com, then a white page. Fix it fast: hard refresh, unregister the service worker, clear cache, disable extensions, check JS policy.
Blocked third-party cookies break OpenAI's login handshake and trigger a Cloudflare loop. Exact per-browser toggles to sign in again, verified June 2026.
Numbers come back wrong, headers misaligned, dates flipped. Force the analysis tool, declare your locale, and verify row counts — accurate results in three steps.
You asked for analysis and got bullet points anyone could write without opening the file. The fix is prompting for evidence, not switching models.
You uploaded a file, asked a question, and the file is gone. Since the March 2026 Library, most uploads are recoverable — here's how to find them and stop it recurring.
ChatGPT rejects your file or says it can't analyze it. Identify the real type, then convert to a known-good format. Conversion beats fighting the upload.
ChatGPT file upload stuck at 0%, fails after the bar finishes, or uploads but says 'can't read'? Different problems, different fixes. Verified limits and 9 causes for June 2026.
ChatGPT made a file but the download 404s, hangs, or comes back empty (often The code interpreter tool session has expired). Diagnose and fix it fast.
Click a Custom GPT and get a blank screen, or its action returns 'request failed'? Nine times out of ten it's cache, account access, or the creator's action backend. Here's how to tell which.
Sidebar empty? 9 times out of 10 the chats aren't gone — wrong account, Temporary Chat, archived, or a stale tab. Walk this checklist before assuming loss.
Image won't generate? Safety filter, quota, prompt conflicts, or capacity throttle — diagnose in 30 seconds with a minimal probe, then fix the right branch.
ChatGPT describes your screenshot or chart wrong. Fastest fix: re-upload a sharp PNG (long edge 1500px+), crop to the region, then give a directed transcription prompt.
Uploaded a 200-page report, but the summary reads like it only saw 30 pages. ChatGPT runs retrieval, not full reading. Force depth with chunked, page-anchored queries.
Credentials and SSO succeed but the page bounces back to login or the Cloudflare check. 90% of the time it's cookies, a rotating VPN IP, or a stale session. Here's the exact order to fix it.
Chrome works, Safari fails (or any other pair). The account is fine — the difference lives in cookie policy, extensions, stale SSO state, or browser version.
ChatGPT says "got it" then forgets next session? It's usually a toggle, Temporary Chat, a write that never happened, or the new background memory. Check these and fix it.
Hit the "You've reached your limit" wall in ChatGPT? Here's what the rolling cap really means, the fastest way back to work, and how to stop hitting it.
Same email but the ChatGPT app and web act like different accounts. The three real causes are workspace, SSO method, and App Store subscription — here's how to tell them apart and fix it.
Uploaded 3 files for comparison but ChatGPT only reads one? Name every file and force a per-file table. Verified for GPT-5.5, June 2026.
ChatGPT showing 'network error' or 'There was an error generating a response' mid-message? It's one error surface for eight different problems. Triage in 30 seconds, then fix the exact cause.
Forgot password and nothing came? It's a different flow from signup verification — stricter rate limit, SSO-only and passkey-only accounts are silently rejected. Here's the fastest fix.
Upload a PDF and ChatGPT says it sees nothing or 'No text could be extracted' — almost always scan-vs-text, size, or encryption. Fastest fix inside.
You uploaded files to a ChatGPT Project but the model answers as if they don't exist. Fastest fix: name the file in your prompt and force a quote. Full diagnosis below.
You wrote detailed Project instructions and ChatGPT keeps ignoring them mid-conversation. It is almost always instruction weight, a conflict with your current message, the 1,500-char field limit, or Memory overriding the rule. Here is the fix.
Set instructions and files in a Project but a chat ignores them, or one chat 'remembers' another? Map the five context layers and fix the leak fast.
ChatGPT keeps logging you out mid-chat? The 3 main causes are cookies being wiped, multiple tabs racing the token, and a corporate proxy stripping headers. Verified June 2026 fixes.
A chatgpt.com/share link 404s or shows blank? It's almost always sharer-side — deleted, policy-removed, or never created. Here's how to tell which, fast.
ChatGPT replies crawling? It is almost always a long conversation, a heavy model, or network RTT — in that order. Start a new chat first; it fixes most cases.
A ChatGPT shared link is a public read-only snapshot, not a login. Here's what the URL does, why you can't reply, and how to share or revoke one safely.
ChatGPT shows a spinner or 'thinking' that never returns? It's not one issue, it's six. Run this 1-minute triage to find the right fix and stop the spinning.
It's almost always your egress IP, not your account. Switch network, wait for the cool-down, re-login. Full diagnosis below.
PDF uploaded fine but ChatGPT misses tables, skips pages, or invents numbers. The fix is almost always the extraction layer, not the model. Step-by-step diagnosis with current limits.
Signup link or login code not coming? Spam, a typo, an alias filter, corporate ATP, a VPN, or regional throttling — here's how to find which one and fix it fast.
Voice silent, stuck on Connecting, or the waveform button missing? It's almost always microphone permission, the wrong input device, a stale build, or the retired macOS-app voice.
ChatGPT search spins forever, returns no sources, or refuses to browse? Here's how to force it, plus the six reasons it silently fails.
You signed up with Google but try email + password (or the reverse). ChatGPT refuses and it looks like a wrong password. Here is how to find the real method.
Artifact download does nothing or drops a 0-byte file. Fix it: wait for the stream to finish, disable blockers, or have Claude paste the content.
Claude in Chrome shows connected but says it can't see the page? Fix per-site permission, SPA/Canvas pages, wrong active tab, and the Cowork conflict.
Claude Code stops every few seconds to ask permission or to ask "TypeScript or JavaScript?" Two different problems, two fixes: a permission mode for tool prompts, CLAUDE.md defaults + Proactive style for decision questions.
Agent forgets earlier decisions, re-reads CLAUDE.md, contradicts the plan. Auto-compact dropped the load-bearing details. Run /rewind, re-anchor, or restart with a tight handoff.
Claude Code re-edits the same file or ignores whole workspaces? Run /init, anchor it with CLAUDE.md, and start from the repo root. Verified June 2026.
You fed Claude Code a detailed audit and it did its own thing. The report is context, not a contract — convert it to numbered Tasks, execute one at a time, and diff progress against the original list.
Claude Code suggests "extract this into a microservice" when you're intentionally monolithic. Fix it with explicit do/don't rules in CLAUDE.md, path-scoped .claude/rules/, and a PreToolUse hook that hard-blocks drift.
The diff is polished code for the wrong feature. Catch the divergence at the plan stage with Plan Mode (Shift+Tab) and a restate-before-execute step, not after the code lands.
Two hours of uncommitted edits gone after the agent ran. Try /rewind first, then editor local history and reflog — then lock it down with commit-before-agent, deny rules, and worktrees.
You approved a 5-step plan; Claude did 7 steps, skipped one, and refactored adjacent code. Bind execution to the approved plan with file-anchored steps + post-step status checks.
Claude Code editing the wrong files, duplicating functions that exist, or missing key code? It's almost never the model — it's missing project context. Seven concrete ways to make Claude Code read your repo, verified for June 2026.
Asked for a one-file rename, got a 12-file diff? Enforce the boundary with a permissions deny rule, scope every prompt with an editable-file list, and park adjacent issues in FOLLOWUPS.md.
Claude Code halts a refactor with 'usage limit reached.' Tell apart the 5-hour window, the weekly cap, and the separate Opus cap, save state, and resume without redoing work. Verified for June 2026.
Claude Code did 60% of the task then went silent — no error, no completion. Diagnose which of three flavors (context compaction, tool hang, dropped stream) and recover without redoing work.
Claude Code repeats the same edit, oscillates between two files, or chases a flaky test forever. Six loop signatures, the one-prompt fix for each, and the exact Esc / rewind / clear / Stop-hook commands to recover. June 2026.
Claude Code changed a file you didn't intend (lookalike name or it followed an import chain). Revert with git, then lock the boundary with permissions.deny Edit rules in .claude/settings.json.
A Claude connector refuses to authorize or returns empty data. Diagnose the four locks: workspace policy, OAuth scope, target-app admin policy, and token state.
Generating files in Claude costs far more of your usage limit than chatting. Here is how it's counted and six ways to conserve, verified June 2026.
Claude's GitHub connector shows no repos or 'access check failed'? Fix it: install the GitHub App on the right account, select the repo, and clear org approval.
When Claude returns a confident but inaccurate answer, switching models rarely helps. Grounding, system-prompt rules, and a fresh chat fix most cases.
After 50+ turns Claude contradicts itself or forgets decisions. Why it happens, how to spot which cause you have, and the fastest reset that fixes it.
Claude keeps looping on 429 rate_limit_exceeded — read retry-after, cap concurrency, separate 429 from 529, and stay under your tier's RPM/ITPM/OTPM.
A Claude tool greys out or shows "unavailable in your workspace." Triage status, plan, and admin policy in the right order to get it back fast.
Asked Codex to audit your project and got a 50-bullet report mixing typos with architecture? Re-scope to one dimension, cap output, force file:line, use /review's custom-instructions preset.
Codex says "done" but leaves broken imports, half-converted types, and untouched call sites. Bind "done" to a verifier in AGENTS.md, then check the exit code yourself.
Codex creates `utils.v2.ts` next to `utils.ts` instead of editing in place. The real cause is an apply_patch Update→Add fallback. Force in-place edits with AGENTS.md rules, git, and the right sandbox mode.
Codex skips the build, misreads it, or trusts a truncated tail. Fix it with exit-code checks, machine-readable verifiers, and a Stop hook gate — not prose summaries.
The reported bug is gone, two adjacent features regressed. Cap blast radius, enumerate call sites, prefer caller-side guards over shared-util changes.
New files land in /src/ when your repo uses /app/; deps go to the root package.json in a monorepo; tests use a framework you don't have. Fix it with a root AGENTS.md, nested overrides, and canonical-example pointers — verified against Codex behavior, June 2026.
Codex uses a column, env var, or function signature it never read. Force a quote-before-write rule and run Codex read-only so it grounds every line in your real files.
Codex writes `getUserById` when your codebase uses `findUserById`. Codify each convention as rule + a canonical example file in AGENTS.md, then verify it loaded.
Codex fails with "Invalid Context" or "No such file or directory" when applying a patch. Tell the two apart, then refresh state or restart the session.
Codex review returns generic bullets like "consider error handling." Fix it with the built-in /review command, AGENTS.md review guidelines, and file:line-anchored prompts.
Functionally correct but reads like a stranger wrote it: async/await mixed with .then(), wrong import order, off comment style. Anchor Codex to a canonical file in AGENTS.md, then enforce the rest with lint.
"Test the happy path and error path" is useless filler. Force Codex to bind tests to the real function signature, the actual types, edge cases, and bug history.
Codex puts files in the wrong place or ignores your conventions. Constrain it with AGENTS.md, a sibling-file reference, and plan-first prompting.
EEA/UK/Switzerland need a Google-certified TCF v2.3 CMP; US states need a Do-Not-Sell link. The fastest compliant setup, plus how to verify it.
Browser blocks your fetch with a CORS error while curl works. Diagnose the real cause and fix it server-side, with verification.
You hit Apply, Cursor says it's done, but the file never changed — usually a stale snapshot, an outside edit, or a locked file.
Cursor shows a diff but Apply does nothing, fails silently, or says the apply model made no changes. Fastest fix: save all files, close other editor windows, then Reapply.
Composer is sharp in a small repo and loses the thread in a 50k-file monorepo. Force focus with @Files, MDC rules, and scoped tasks — not heroic re-indexing.
Composer rewrote 10+ files and broke working code. Revert fast with a Cursor checkpoint or git, then bound the prompt and mode so it stays surgical.
Open a team repo and Cursor's formatting or AI behavior changes? The repo's .vscode/settings.json is overriding your user settings. Here's how to find which layer won.
The Cursor context panel skips the file you know is relevant. Fix it in 5 seconds with @, then stop it recurring: check .cursorignore, resync the index, and rename generic files.
Composer dropped a 30-file, 2000-line diff and you can't review it honestly. Reject (or restore the checkpoint), then make Composer split it into reviewable PRs.
Composer says done and the touched files lint clean, but the build fails. The fast fix: turn on Auto-run so the agent runs typecheck + build itself before declaring success.
Cursor wrote a new formatDate but src/utils already has one. The existing helper never reached retrieval — here is how to make Cursor grep first and reuse it.
Editor shows old code after an AI edit, save says "file modified externally", or AI changes silently revert — the watcher / AI-write race, and how to fix it without losing work.
Cursor's index spinner runs for hours with no progress. Usually it's a bloated repo, a missing .cursorignore, a corporate HTTP/2 proxy, symlink loops, or permissions. Here's how to find which one.
Cursor only edits files it can see. Fix indexing scope, open the right workspace root, and pin files with @ references.
You @ one file and Cursor answers about a different same-named file. Fix the @-picker, stale index, and symlink causes with full-path references and a resync.
Cursor pulls old, vendored, or generated files into context and produces wrong patches. Fix the index, .cursorignore, and prompt pinning.
Tab and Composer use different naming / formatting / patterns than your repo — training prior beats retrieval; anchor explicitly.
Grey inline completions vanished in Cursor? Re-enable Tab, check sign-in/quota, kill Copilot's keybinding, switch to HTTP/1.1, or clear the cache. Step-by-step, verified June 2026.
Cursor's agent runs a shell command and gets it wrong — wrong working dir, wrong package manager, env not loaded, output truncated, or .zshrc hangs the run. Concrete fixes.
SSL pending for hours after adding a domain on Vercel/Cloudflare/Netlify? It's almost always DNS, a CAA record, or a blocked HTTP-01 challenge. Here's the order to check.
Vercel / Firebase / Cloudflare say the deploy shipped, but visitors still see yesterday's HTML. The 5 ranked causes and a 10-minute client-to-origin diagnostic.
Updated DNS but the site shows the old IP or won't load? Verify with dig, check the active nameservers, flush local + browser cache, and wait out the TTL. Step-by-step, as of June 2026.
You changed DNS and some users see the new IP, some the old. Here's why regional inconsistency is normal, plus a measurable fix and verification path.
FUNCTION_INVOCATION_TIMEOUT or 504 on edge? Move heavy work off edge, stream responses, and timeout every upstream fetch. Verified limits, June 2026.
Works locally, fails in prod. Env var not set on the host, or wrong client prefix (PUBLIC_ / VITE_ / NEXT_PUBLIC_). Diagnose and fix in ~10 minutes.
firebase deploy fails with 403 or "caller does not have permission"? Diagnose the identity chain — account, project, IAM roles, CI auth — and fix it fast.
A callable returns "not found" even though the console shows it. Diagnose region, name, and deploy state in minutes.
Fix firebase deploy --only hosting: CLI version, expired auth, wrong public path, or Spark quota.
Homepage works but /about or a refreshed /blog/post returns 404 on Firebase Hosting? It's almost always firebase.json rewrites. Verified fixes for SPA, SSG, Astro, and Next.js as of June 2026.
Your firebase.json rewrites for an SPA or Cloud Function don't trigger. Fix the order, region, and deploy target in minutes.
Tap-and-hold power doesn't bring up Gemini on Android — toggle, app version, region.
Same Google account, different chat lists on the app vs gemini.google.com. Account mismatch, the 2026 sync-overwrite bug, Keep Activity, and Workspace policy explain almost all of it.
Deep Research output is shallow or one-sided? Fix it with scope, source filters, and follow-up drills — plus the new Deep Research Max (Gemini 3.1 Pro).
Gemini summaries miss key numbers and skip risk sections? Tell it what to extract and how to structure it. Tested prompts, model-picker fix, and verification steps for June 2026.
Gemini Share button grayed or missing, shared g.co/gemini/share link 404s, or Export to Docs fails. Fix by tier, Workspace policy, and source-conversation state.
File attached but Gemini can't see it, or the summary is wildly off? Fix size caps, format, Drive access, and scanned PDFs.
Gemini says "Permission denied" or can't access your account? Re-connect the app, turn on Gemini Apps Activity + Gmail Smart features, and clear Workspace policy.
Gemini won't generate an image, has no Create image button, or says 'I can't generate that image' — fix the safety filter, region, tier, and Workspace gates.
Gemini spins forever or returns a blank card? Work through account, network, browser, and Workspace gating in order — most cases clear in under five minutes.
Gemini says "try again later" or "feature not available." Almost always quota or tier, not an outage. Here is how to tell which, and the fastest fix.
No Gemini side panel in Docs or Gmail? Check edition, the admin Feature access toggle, and Smart features. Fix order inside.
Ask in Chinese, Gemini answers in English (or vice versa). Lead your prompt with 'Reply in Chinese' for an instant fix, then align account and browser language so it sticks.
You committed over code you needed, or reset --hard the wrong commit. The fix: most of the time the objects are still in your repo. Start with `git log -p --follow -- <file>`, or `git reflog` if a whole HEAD vanished.
Google replaced your title tag with a shorter or different one in search results. Diagnose which signal triggered the rewrite and fix it — with current 2026 length budgets and the new AI-headline behavior.
Google retired the International Targeting report in 2022, so hreflang errors no longer show in Search Console. Here is how to find No return tags, invalid language code, and 404 target errors with a crawler instead, and how to fix each.
A 200-line Claude Code or Cursor diff isn't safe just because it compiles. Read it deletions-first, in the right order.
Works on localhost, breaks on Vercel / Render / Cloudflare Pages. Reproduce the prod build locally, diff env vars, pin the Node version, and fix Linux case-sensitivity — the systematic checklist.
Your 500-word super-prompt scores worse than a 100-word one. Here's why long prompts degrade output, plus 5 fixes that work as of June 2026.
Google shows its own snippet instead of your meta description. Find which of 7 triggers your page hits and rewrite to win the snippet back.
Claude Code + Codex (or two parallel sessions) edited the same function and produced a merge conflict. Isolate scopes with git worktrees, abort the broken merge, and resolve as a human — never let an agent merge.
Cursor and Claude Code clobbered the same file. Stop both, recover each intent, merge by hunk, and isolate agents with worktrees so it never repeats.
After a DNS change your MX disappeared and mail stopped arriving. Diagnose the exact case and restore the correct records for Google Workspace, Microsoft 365, Zoho, Fastmail, or Proton.
Slack, X, LinkedIn, or Facebook unfurls your URL with no preview image. The 7 real causes and the exact debug commands and cache-refresh tools.
Share to LinkedIn / X / Slack and no image shows. Fix the meta tag, image size, or cache in minutes.
Stripe, Twilio, and SendGrid all return 429 once traffic ramps. Read the rate-limit headers, add exponential back-off with jitter, cache idempotent GETs, dedupe concurrent calls, and batch writes — in that order.
Your vercel.json / _redirects 301 deployed but fires a 404 or the old page. Diagnose with curl -I across Vercel, Netlify, and Cloudflare Pages — file location, rule order, file shadowing, CDN cache, trailing slash, rewrites vs redirects.
You changed robots.txt but Google still crawls or indexes the same URLs. The fix is almost always a cache layer or a duplicate file — here is how to find which one.
Your /rss.xml or /feed.xml 404s: the endpoint isn't in the build, the filename mismatches the link tag, or a host rewrite is hiding it. Per-stack fixes for Astro, Next.js, Netlify, Vercel.
/sitemap.xml or /sitemap-index.xml returns 404, or Search Console says 'Couldn't fetch'. Five ranked causes with checks you run against dist/.
Submitted a sitemap to Search Console weeks ago and still 0 pages indexed — or only a handful? Sitemap submission is not indexing. Here are the 6 real causes, how to tell which one you're hitting, and the exact fix for each.
You bound a custom domain on Vercel / Firebase / Netlify but SSL stays "pending." Diagnose DNS, CAA, and Cloudflare-proxy causes and fix it fast.
Images, CSS, and fonts 404 in production but work locally. Five web-verified causes and fixes: missing dist/ copy, base path, case mismatch, content hash, CDN cache.
Production URL loads to an empty page with no UI error. Open DevTools before navigating, read the first red console line, and map it to the fix — base path, hydration, CSP, build target, or a stale Service Worker.
Fix S3, Firebase Storage, and Supabase Storage upload 403s — IAM permission, signed-URL expiry, bucket policy, or disabled-ACL gotcha.
Rich Results Test shows your JSON-LD has warnings or invalid fields. Tell errors from warnings and fix each one.
Your apex domain works but blog.example.com gives NXDOMAIN or an SSL error. The fix is almost always a missing DNS row or a domain you forgot to attach at the host. Diagnose it in five buckets.
Suno Extend continues from the audio at your cut point plus the style text, not the song's identity. The cut-point, style-lock, and Get Whole Song fixes that smooth the seam.
Sad lyrics, upbeat backing? Put mood, tempo and key in the Style of Music field, add Exclude Styles, and raise Style Influence.
You wrote lyrics, Suno sang something else. The fix is Custom Mode plus bracket structure tags and matching lyric length to song length.
Suno slam-stops mid-phrase or does a 3-second fade. Fastest fix: end your lyrics with an `[Outro]` instruction plus an `[End]` tag on the very last line, then use the in-app Fade Out tool. Full causes, templates, and verification inside.
Suno v5.5 can generate up to 8 minutes in one shot, but Extend, Replace Section, and Get Whole Song are how you build a real song structure. Exact buttons, prompts, and verification steps.
Suno song stays at one energy with no chorus payoff? Add structure tags in the Lyrics field plus dynamics words in Style. Templates, ratios, and the v5.5 editor fix.
Your "synthwave, retro, 80s synth" prompt gives a piano ballad. Why Suno drops style words, and the fixes that stick (v5.5, June 2026).
Suno repeats the chorus byte-for-byte because your lyrics repeat. Break the loop at three layers — lyrics, melody hints, structure tags — or patch one section with Replace.
Suno mangles Mandarin polyphones, tones, and mixed-language lines. Fastest fix: generate on v5.5, respell problem words in your lyrics before generating, and keep lines short. Full diagnosis + fix path.
Suno vocals sound like cheap TTS — stiff, no breath, no soul. Strip synthetic style words, request natural breath, and fix line phrasing to get human vocals.
You wrote lo-fi hip-hop in Style and got generic pop. Fix it by putting the genre first, adding a BPM, and using the Exclude Styles field to strip conflicting instruments.
Supabase URL or anon/publishable key is undefined after deploy — fix the host env config, framework prefix, stale build, and the 2026 key migration.
Your Supabase table has rows but the frontend gets [] with no error. Diagnose which RLS bucket you're in and add the right SELECT policy in minutes.
/foo 308s to /foo/ then 308s back to /foo. One rule across framework, host, and links kills the redirect chain. Astro/Vercel/Next.js config included.
Build is green but production 500s. Diagnose FUNCTION_INVOCATION_FAILED, missing env vars, Edge runtime, and timeouts with real vercel logs commands.
Build sits at "Building" with no new logs. Cancel it, read the last log line, then fix the real cause: a hung fetch in static generation, an install deadlock, OOM, or a postbuild hook.
Stripe, GitHub, or Shopify says the webhook was delivered but your server never sees it. Read the delivery log, find the status code, and fix the real cause.
Apply too early and collect a 'low-value content' rejection; too late and you lose months of revenue. Run this 6-signal readiness check first.
www.example.com works but example.com returns an error (or vice versa). Fix the missing DNS record, then add a 301 redirect to one canonical version.
AI faces come out warped, melted, or twisted? Usually it is not the prompt. Here is the diagnosis-first 3-step fix: tighten framing, raise resolution, run a face-restore pass. Updated June 2026.
ChatGPT shows a blank page, hangs on a spinner, or returns access denied? Work this 6-step triage in order: status page, incognito, extensions, DNS, VPN node, device.
ChatGPT login failures fall into four buckets: email-password bounce, Google/Microsoft SSO error, the new 6-digit OTP code never arriving, and a suspicious-activity lock. Find your bucket, get the shortest fix.
Claude says 'usage limit reached'? Here's how the limit actually counts (5-hour window + weekly cap, weighted by model), exactly when each resets, and six ways to stretch it.
Search Console says 'Crawled — currently not indexed'? Google fetched the page and chose not to index it. Here are the real causes in order, the fastest fix, and how to confirm it worked.
Cursor stuck on Indexing or re-indexing in a loop? It's usually a build/cache directory your file watcher keeps touching. Here's the .cursorignore template, the watcherExclude fix, and how to confirm the index is healthy.
noindex blocks indexing; robots.txt blocks crawling. They are not interchangeable. Here's a decision table, the conflict that silently leaves pages indexed with no snippet, and how to confirm the fix in Search Console.
Suno verse sounds fine but the chorus has no lift? The fix lives in the first 4 bars. Use structure metatags plus three musical levers — key change, rhythmic contrast, hook landing — verified for Suno v5.5 (June 2026).
When a Vercel deploy goes red, the build log already names what broke — if you know where to look. Three log positions, plus the six most common failure modes and exact fixes.
Site live for a week and still not on Google? Start with URL Inspection, then work the 10 real causes: robots.txt, noindex meta, X-Robots-Tag header, bad canonicals, JS-only rendering, the new-domain wait, and thin content.