The 9 Fixes — Prompt Pack — AI Creators Roundtable

Companion to Your Old AI Prompts Are Costing You Money. Source: Anthropic, Prompting Claude Fable 5.1. Deciding whether to pay for 5.1 in the first place? Care or Skip?

The behaviours described here are Anthropic's, from the page above. What they mean for your business is my read, not theirs.

Blocks marked ▸ are adapted from Anthropic's own wording — trimmed, switched to first person, and in fixes 7 and 8 generalised from developer examples to everyday work. The originals are on the source page. Everything else is mine.

Nothing here needs code. Paste these into your custom instructions, your project instructions, or the top of a chat.

The ten-minute setup

Open wherever you keep your standing instructions — Claude's custom instructions, a project's instructions, your CLAUDE.md, your saved prompt doc.
Do fix 2 first. Deleting is free and it's the fix with the biggest payoff. Budget nine minutes — most of it is reading what you already wrote.
Paste in the blocks from fixes 2, 3, 4, 5, 7, 8 and 9 — seven blocks. Those are standing rules; they belong in the file, not in individual chats.
Fix 1 is a setting. Fix 6 is a habit. Neither is text.

That's it. You do this once.

Fix 1 — Effort is your bill, not a quality slider

A setting, not a prompt. Five levels: low, medium, high (default), xhigh, max.

⚠️ You only get this dial if your setup exposes it — Claude Code, the API, most builder tools. In a plain chat window it isn't available to you yet.

Run at	For
low	captions, transcript cleanup, renaming, sorting, tagging, first-pass formatting
medium	drafts, outlines, summaries, most day-to-day work
high	the default. Strategy, positioning, pricing, anything you'll act on
xhigh / max	only where you've actually measured a quality gain

That table is my starting map. What Anthropic documents is narrower and worth quoting exactly: at medium, 5.1 "roughly match[es] Claude Fable 5 at lower cost"; at low it is "often competitive with Claude Opus and Claude Sonnet models on cost per task while scoring higher." They also warn that effort level names don't mean the same thing across models — so re-test even if you swept before.

Fix 2 — Delete your expired rules

Start here. Search your instructions for these and remove them:

"don't use bullet points" / "minimize formatting" / "avoid headers"
"don't use bold" / "no markdown"
"hold all findings for the final response" / "don't narrate while working"
anything telling it to be less structured or less talkative

All of these were written to hold down behavior 5.1 no longer has. Per Anthropic, it "uses bold less and is less likely to reach for headers, lists, or quotation marks" than earlier models.

Deleting is only half the documented fix. If you want something in that slot, don't say "less formatting" — say when formatting is right:

▸

Use lists and bullet points when asked to, or when the content is multifaceted
enough that they help with clarity. If I explicitly request minimal formatting,
format without bullet points, headers, lists, or bold emphasis, as requested.
In conversational, personal, or emotional exchanges, keep to plain prose.

Fix 3 — Ask for progress updates

Mostly relevant if you run agentic or long tool-using jobs (Claude Code, an agent, anything that goes off and works on its own).

▸

Before you start, say in a line what you're about to do; brief updates while
you work help me follow along. Close with a short recap that stands on its
own — what you found, what you did, and what's next — so a reader who only
sees the last message has the full picture.

Also: search your existing instructions for anything like "hold all findings for the final response" and delete it. That line was written for an older model and now makes the silence worse.

Fix 4 — Kill the mannered prose

The short version, which usually does it on its own:

Please remove all mannered prose.

The full version, for your standing instructions:

▸

Mannered prose substitutes metaphor and flourish for direct statement. Instead
of "a parameter worth varying," the mannered writer produces "a dial worth
turning." Instead of "this point still matters," they write "this point earns
its keep." The phrases exist to display the writer, not to convey the idea, and
readers can tell. That is why mannered prose irritates: it makes the reader work
harder so the writer can perform. It is also imprecise. Metaphors drag in
connotations the writer did not choose and cannot control. The fix is to say
what you mean. When a literal phrase is available, use it.

Fix 5 — Check the quoting before you publish ⚠️

The behavior, documented: 5.1 "is more likely than Claude Fable 5 to reproduce passages of the source text without marking them as quotations."

What that costs you is my read: if you repurpose articles into scripts, newsletters or posts, you can publish someone else's sentences under your name without knowing. If a publisher recognizes their own wording, that's a takedown request. With a client it's a shorter conversation than that.

Anthropic's own fix is to paste one complete worked example into your system prompt — the request, a correct response, and a sentence explaining why it's correct. The block below is the short version of what that example teaches; use it if you're not maintaining a system prompt:

When you summarize or synthesize a source, put it in your own words. Organize
the response around where sources agree and differ, not as a walk through each
article. Use at most one short quoted phrase, in quotation marks, with
attribution — reword every other claim.

The habit that actually protects you: anything that comes back sounding unusually polished, paste one distinctive sentence into a search engine before it goes out. Ten seconds.

Fix 6 — Let it zoom in

A habit, not a prompt. On dense images — analytics dashboards, spreadsheets, contracts, anything with a lot of small numbers — 5.1 does its best work when it can crop into a region and enlarge it rather than reading the whole frame at once.

Crop to the part you care about before you paste it, or
Ask it to work through one section at a time rather than "what do you think of this dashboard?", or
If your setup can run code on the image, let it crop and zoom itself. Anthropic is explicit that this is where the real gain is — the model doing its own cropping, not you uploading a bigger file.

Fix 7 — Make it look things up

▸

When a question centers on a name you don't confidently recognize, or one from
a fast-moving area like AI models, pricing, or creator tools, the name itself
is the thing to verify: search before answering, and include the name exactly
as I wrote it in at least one query. This holds even when you have some
background on it — partial background is exactly what makes an out-of-date
answer sound authoritative.

Also: where being current actually matters, raise the effort level for that one question rather than for the whole conversation.

Fix 8 — Tell it to stay in scope

▸

If, while working, you find a pre-existing problem or something the task
doesn't mention, don't fix, optimize, or extend it in this piece of work unless
what I asked for cannot work without it — report it as a follow-up in your
summary. Where the task is ambiguous, do the reading its wording most directly
supports, state that assumption in your summary, and don't build for the other
readings as well. This is about extras only: do everything I actually asked
for, completely.

Fix 9 — Make it actually finish

▸

You are operating autonomously. I am not watching in real time and cannot
answer questions mid-task, so asking "Want me to…?" or "Shall I…?" will block
the work. For reversible actions that follow from the original request, proceed
without asking. Stop only for destructive actions or genuine scope changes I
must decide. Offering follow-ups after the task is done is fine; asking
permission before doing the work is not.

Exception: when I am describing a problem, asking a question, or thinking out
loud rather than requesting a change, the deliverable is your assessment.
Report your findings and stop. Don't apply a fix until I ask for one.

Before ending your turn, check your last paragraph. If it is a plan, an
analysis, a question, a list of next steps, or a promise about work you have
not done ("I'll…", "let me know when…"), do that work now with tool calls.
That includes retrying after errors and gathering missing information
yourself. Do not stop because the context or session is long. End your turn
only when the task is complete or you are blocked on input only I can provide.

Don't trim that first sentence — Anthropic says it carries much of the effect on its own. (I've swapped "the user" for "I" so you can paste it as yourself; keep the rest as-is.)

The documented trade-off: this also makes it less likely to check in on genuinely ambiguous requests. Fair trade for me. Might not be for you.

The one-line version of all nine

Cheapest effort that still holds · delete your expired rules · ask for progress updates · kill the mannered prose · check the quoting before you publish · let it zoom in · make it search · tell it to stay in scope · tell it to finish.

What this doesn't do

None of this makes the output good. It makes it accurate, cheaper, and finished. The taste and the judgment about what's worth publishing are still yours — that's the deal.

AI handles the busywork. You keep the taste.