The complete walkthroughs from that post

You are auditing a prompt rewrite. I will give you OLD and NEW versions of a system prompt. For EVERY substantive difference, output one row: | # | change (one line) | type | keep? | why | type is exactly one of: CONTRADICTION_FIX - old text told the model two incompatible things OUTPUT_SPEC - adds/clarifies format, schema, length, or refusal behaviour TIGHTENING - same meaning, fewer or clearer words CONSTRAINT_DROPPED - a rule, edge case, example, or domain fact present in OLD and absent in NEW NEW_ASSUMPTION - NEW asserts something OLD never said Rules: - Do not summarise. One row per change, in document order. - Default keep? to NO for CONSTRAINT_DROPPED and NEW_ASSUMPTION. - Ignore pure whitespace and heading-style changes. - After the table, list every behaviour the OLD prompt guaranteed that the NEW one no longer does. OLD: <<< [paste original prompt] >>> NEW: <<< [paste optimized prompt] >>>

You are a document data extractor. Extract ONLY what is literally printed in the attached page images. Return a single JSON object. For EVERY field, emit an object with exactly these keys: "value" - the normalized value (string, number) or null "quote" - the verbatim text you read, copied character-for-character from the page, including currency symbols and punctuation "page" - the 1-based index of the image the quote appears on "confidence" - "high" or "low" Rules: - If a field is not printed on the page, set value, quote and page to null. NEVER infer, compute, or guess a value. A missing field is a correct answer. - "quote" must be an exact substring of the text on the page. Do not paraphrase, reformat, expand abbreviations, or silently fix typos inside "quote". - Do all normalization (date reformatting, stripping currency symbols, parsing numbers) in "value" only. - If the page shows two plausible candidates, pick the one physically nearest its field label and set confidence to "low". - Never merge values from two different pages into one field. - Output JSON only. No prose, no explanation, no markdown fences. Fields: invoice_number, invoice_date, due_date, vendor_name, vendor_tax_id, bill_to_name, currency, subtotal, tax_total, grand_total Also return "line_items": an array where each element has the same {value, quote, page, confidence} shape for: description, quantity, unit_price, amount.

#!/usr/bin/env bash # refactor.sh — one file per worktree, merged only if tests pass. set -uo pipefail TEST_CMD="npm test" # <-- your real test command MIGRATION="convert every callback-style function to async/await" # <-- your real migration BASE="$(git rev-parse --abbrev-ref HEAD)" mkdir -p .refactor && : > .refactor/failed.txt while read -r FILE; do [ -z "$FILE" ] && continue SLUG="$(echo "$FILE" | tr '/.' '--')" WT="../wt-$SLUG" git worktree add -b "refactor/$SLUG" "$WT" "$BASE" || { echo "SKIP $FILE"; continue; } ( cd "$WT" || exit 1 # deps are untracked, so a fresh worktree has none — reuse the main checkout's: [ -d ../"$(basename "$OLDPWD")"/node_modules ] && ln -s "$OLDPWD/node_modules" node_modules 2>/dev/null claude -p "Refactor ONLY the file $FILE: $MIGRATION. Hard rules: 1. Edit no file other than $FILE. Do not edit tests, config, or lockfiles. 2. Do not change exported names, function signatures, or observable behaviour. 3. Preserve error semantics exactly: an error that used to reach a callback must still reach the caller by the path the caller already handles. 4. Add no new dependencies. 5. If the file cannot be migrated without breaking rules 1-4, change nothing and explain why. When the edit is done, run: $TEST_CMD If it fails, fix $FILE only until it passes. Finish with a one-paragraph summary of what changed and any behaviour you were unsure about." \ --permission-mode acceptEdits \ --allowedTools "Read,Edit,Grep,Glob,Bash" # the gate: our verdict, not the agent's $TEST_CMD ) if [ $? -eq 0 ]; then echo "PASS $FILE -> branch refactor/$SLUG (worktree $WT)" else echo "$FILE" >> .refactor/failed.txt echo "FAIL $FILE -> left at $WT for inspection" fi done < files.txt echo "---"; echo "failed files:"; cat .refactor/failed.txt

# summarize.py - bulk-summarize one email per line of emails.txt from transformers import pipeline pipe = pipeline( "text-generation", model="openai/gpt-oss-20b", dtype="auto", # older transformers (<4.56): use torch_dtype="auto" device_map="auto", ) with open("emails.txt") as f: emails = [line.strip() for line in f if line.strip()] for i, email in enumerate(emails, 1): messages = [ {"role": "system", "content": "Summarize the customer email in one sentence, then output SENTIMENT: positive|neutral|negative. Reasoning: low"}, {"role": "user", "content": email}, ] out = pipe(messages, max_new_tokens=160, do_sample=False) print(f"--- {i} ---") print(out[0]["generated_text"][-1]["content"])