Source: raw/Claude_5.5_Model_Mistakes_you_may_still_be_making_23_Tips_and_Fixes.md. Cross-checked against ai-research/claude-docs-prompting-claude-opus-5-5-2026-10-27.md (Anthropic’s Opus 5.5 prompting guide) and existing wiki articles linked inline.

Creator: Robo Nuggets (Jay) URL: https://www.youtube.com/watch?v=gvESqM6s0kM Duration: 20:32 Platform: YouTube Published: 2026-10-12

Jay’s checklist of 23 habits that cost usage or answer quality on the Claude 5.5 models, built from Anthropic’s guides, comments by Anthropic engineers and power-user tips. Most items repeat first-party advice already in this wiki, so the value here is the single list, ordered from easy to advanced, with the 5.5-specific cache and effort details added. Each row says what backs the fix: Anthropic’s own docs, this wiki’s release notes, or Jay alone. His paid-community pitch and PDF offer are left out.

Key Takeaways

  • Most waste is re-reading. Every prompt re-sends the whole conversation. On the 5.5 models the prompt cache cuts that re-read to about 5% of the normal input price (Opus 5.5 cache reads are 4 input), but you still pay it and it still counts toward limits. Unrelated history also makes answers worse.
  • The cache is per model, and it goes cold. Switching model mid-chat re-reads everything at full price. Jay says the subscription cache goes cold after about an hour away, so the first message back is billed in full.
  • Effort is the thinking control; prompt phrasing isn’t. Leave effort at medium for most work. Delete old lines like “think carefully step by step” and all-caps commands. Anthropic found removing “think carefully” made replies start sooner with no clear quality loss.
  • Give reasons and specifics, not slogans. A rule with a one-clause reason works better than a bare rule. “Make it not look like AI” only swaps one default style for another; naming the exact patterns to avoid works.
  • Scope, isolate and clean up. Use /btw for side questions, subagents for exploration, /rewind after two failed corrections, and /compact with instructions at a natural break.
  • Route by task size. Sonnet for defined work, Opus for exploratory work, a router for everything else. Haiku 5.5 lists at about 1/40th of Opus 5.5’s per-token price.
  • Don’t trade safety for speed. Use auto mode instead of skipped permissions, keep dated backups or git, and vet plugins before installing.

Easy fixes (1–11)

#MistakeFixBacking
1Kitchen-sink session: one chat all day across unrelated tasksOne session per task; /clear between topics (keeps CLAUDE.md and skill descriptions loaded)First-party: Anthropic names this failure mode in its Claude Code best practices
2Switching models mid-chat to save moneyPick the model at the start and keep it. Jay relays an Anthropic engineer: 100K tokens into an Opus chat, switching to Haiku for an easy question costs more than letting Opus answerPer-model cache: wiki (Two-Model Workflow). Engineer quote: Jay
3Arguing instead of rewinding: five rounds of correctionsAfter two failed corrections, Esc Esc or /rewind to the point before it went wrongFirst-party (best practices says two corrections, then reset)
4Side questions in the main chat/btw <question>: the answer appears in a side panel and never enters historyFirst-party (best practices)
5Max effort for everythingMedium for normal work, high or xhigh when it matters, max for rare big jobs. Jay cites Artificial Analysis: Opus 5.5 at max scores about 4 points above high at over 3x the cost per taskMedium default: first-party guide. AA numbers: Jay, not verified here
6One-line vague prompts (“make me a website”)State the output, who it is for, the goal and the tone. For repeat work, give a harness: a design system or a skillDesign defaults: first-party guide
7Rules without reasons (“never use bullet points”)Add the reason in one clause (“it gets read aloud on a podcast”)Consistent with Anthropic’s rules-to-judgement shift
8”Make it not look like AI”Name the exact patterns (cream background, italic headline accents) or give a design systemFirst-party guide, same wording
9Taking the first draftFor short creative work (hooks, titles, emails), ask for three independent versions and have Claude grade them against the goal. Each version costs usage, so skip this on big buildsJay (his “burst” skill)
10Drip-feeding follow-upsWrite the full request once. If you spot a mistake in your prompt, rewind and resend rather than adding correctionsJay
11Old-school prompting (“think carefully”, “don’t be lazy”, ALL CAPS)Plain instructions; let effort control thinking. Ask Claude to audit CLAUDE.md and skills for lines written for older modelsFirst-party guide (the “think carefully” test)

Intermediate (12–19)

#MistakeFixBacking
12Walking away mid-session and typing “continue”Jay: the subscription cache goes cold after about an hour. His example: continuing a ~400,000-word session cost about 12 cold. Before a break, /clear if done or /compact while warm. To keep it warm, some users have /loop ping every 30 minutes (about $0.10 per ping on a big session, per Jay)Jay. The one-hour figure is consistent with Claude Code’s 1-hour TTL options in the CLI reference^[inferred]
13Stuffing CLAUDE.mdUnder 200 lines, used as a router that points to skills, docs and foldersFirst-party cap (Steering Claude Code, CLAUDE.md primer)
14Turning off all safety (skipped permissions, clicking yes to everything)Use auto mode, where a classifier reviews each action. Add a guard: Jay names DCG (“destructive command guard”), a free open-source add-on that blocks commands like folder deletes and force pushes, and a mod he calls Blast Radius that shows which files a command would delete with proceed/cancel buttonsAuto mode: first-party. Jay reports a user’s whole C drive deleted in skip-permissions mode and “at least four” similar Windows reports since September; not verified here. DCG and the mod’s name and maker are not verified^[ambiguous]
15No backup before big editsAsk Claude to copy the folder with today’s date first; better, set up gitJay
16Every connector on all the timeAudit MCP servers every so often: which you used in two weeks, what each costs, which have a CLI Claude could use instead. Claude Code now loads full tool details only when needed, but each server still adds a tool list to every chatDeferred tool loading: first-party (context-engineering post). CLI-over-MCP preference: Jay
17Infinite exploration (“look into why sales dropped”)Scope it (“only the Q2 reports in the sales folder”) or hand it to a subagent, which can run on Sonnet or Haiku and returns only a summaryFirst-party failure mode (best practices)
18Never checking usage/usage breaks spend down by skill, subagent, plugin and MCP server. Jay says it flags any habit taking 10% or more. A mod can show 5-hour and weekly usage beside the prompt box/usage categories: CLI reference (v2.1.149). 10% threshold: Jay
19Skipping alignment on big buildsLet Claude interview you first with numbered questions you answer by numberJay (his “align” skill)

Advanced (20–23)

#MistakeFixBacking
20Losing old chatsJay: Claude Code deletes local session history after 30 days by default. Set cleanupPeriodDays to 365 in settings, then reopen sessions with /resume. It does not recover sessions already deletedJay. Desktop and Cowork sessions stopped being deleted after 30 days in v2.1.248 (Week 35); the CLI default is not confirmed here
21Installing mods and plugins blindA plugin can run scripts on your machine and many add tokens to every message. Check what it installs and what it costs per turn, or have Claude assess fit, safety and needJay (his “personalize” skill). See Plugins and Marketplaces
22Letting auto-compaction pick what to keepJay relays an Anthropic engineer: the model is “at its least intelligent point” while compacting. Compact yourself at a natural break with instructions (/compact keep the client brief and the final price list), or add compact instructions to CLAUDE.md/compact <instructions>: first-party. Engineer quote: Jay
23No model routing (biggest model for file renames and quick emails)Sonnet for defined work, Opus for exploratory work. Jay routes every prompt through Jev: over three weeks, 9 in 10 prompts went to Sonnet or Haiku, and the routing cost him $0.13 in totalPrice ratio: Haiku 5.5 (0.50) vs Opus 5.5 (20) is 40x at list price. One user’s report that Fable spun up 10 side agents and emptied a 5-hour limit in 5 minutes: Jay, unverified. Jay numbers this item 22 a second time; it is the 23rd

Try It

  1. Run /usage now and note the top three categories. Fix the matching rows above first.
  2. Audit CLAUDE.md with this prompt: “Find lines written for older models (think carefully, ALL CAPS, rules with no reason) and anything that could move to a skill. Keep the file under 200 lines.”
  3. Audit connectors: “List my MCP servers, which ones I used in the last two weeks, and which have a CLI you could use instead.”
  4. Before your next break, run /compact with what to keep, or /clear if the task is done.
  5. Set a backup habit: ask for a dated copy of the folder before any large edit, or initialise git.
  6. Route one week of work on purpose: start defined tasks on Sonnet, exploration on Opus, and don’t switch mid-chat.

Open Questions

  • Cache TTL on subscriptions. Jay’s “about an hour” is not confirmed by a first-party source in this wiki.
  • Effort switches and the cache in Claude Code. Jay says changing effort mid-chat keeps the cache on all 5.5 models and Fable 5.1 (beta). Anthropic’s guide documents this for the API’s per-message effort change only. See Prompting Claude Opus 5.5.
  • DCG and the “Blast Radius” mod. The caption reads “Blast Rages”; neither the repo nor the mod’s Anthropic origin is verified.
  • The Artificial Analysis max-vs-high figures for Opus 5.5 need checking against AA’s own page.