24d • 🛠️ Show Your Work
Tired of WisprFlow getting it wrong?
Started seeing lots of errors recently so I asked Codex to fix it!! Prompt can be adapted for whichever tool you're using.
Reusable prompt: Optimize Wispr Flow from Claude or Codex history
Improve my personal Wispr Flow transcription accuracy on this Mac or Windows computer. Wispr Flow is the only speech-to-text application in scope.
Use the tools built into the Claude or Codex environment running this prompt. Do not assume access to any organization-specific session index, MCP server, workspace, or external memory system.
1. Find the available session history
Determine whether you are running in Claude or Codex.
Use that product’s built-in tools for accessing the current user’s past sessions or conversation history.
Analyze only history available to the currently logged-in account.
Include past sessions that are retrievable through built-in tools even if they do not appear in the current conversation interface.
If session-history tools are unavailable, say so and continue with Wispr Flow’s accessible transcript history.
Do not search application databases, credential stores, hidden account files, or unrelated filesystem logs.
Do not ask the user to export sessions unless no built-in history access exists and the additional evidence is necessary.
2. Analyze the user’s language
From accessible Claude or Codex sessions, extract language written or dictated by the user, including:
Names of people, companies, brands, products, and projects.
Technical terms, acronyms, and specialist vocabulary.
Uncommon words and recurring phrases.
Words requiring unusual capitalization, spacing, punctuation, or hyphenation.
Singular and plural forms the user regularly employs.
Similar-sounding words likely to confuse speech recognition.
Exclude:
Assistant-generated language not also used by the user.
Code, logs, quoted documents, and pasted third-party writing unless the user’s own messages confirm the vocabulary.
Sensitive terms that are unnecessary for improving transcription.
Ordinary words Wispr Flow already recognizes reliably.
3. Inspect Wispr Flow
Using available application or computer-use tools, inspect:
Wispr Flow’s Home transcript history.
The personal Dictionary and existing correction rules.
General settings for microphone and language.
System settings, including Auto-add to Dictionary.
Style and cleanup settings.
Experimental settings, including Bulk Import.
Data and Privacy settings, including Context Awareness.
Look for:
Recurring incorrect words or names.
Consistent incorrect-to-correct substitutions.
Capitalization, punctuation, spacing, and hyphenation errors.
Over-aggressive cleanup that changes the user’s meaning or voice.
Conflicts or duplicates in the existing dictionary.
Incorrect language, regional variant, or microphone selection.
When Wispr Flow offers Undo AI edit, use it to compare a processed transcript with the original dictation when that comparison would clarify an error.
4. Classify each finding
Place each supported finding into one of these categories:
1. Vocabulary entry — an uncommon term Wispr Flow needs to recognize.
2. Replacement rule — a specific incorrect transcription that repeatedly needs to become a specific correction.
3. Style or cleanup setting — formatting or rewriting behavior that should be adjusted.
4. No change — evidence is weak, the term is too ordinary, or adding it could create conflicts.
Do not invent likely misspellings. A replacement rule must be supported by an observed error.
5. Create the import files
Create two UTF-8 CSV files:
1. wispr-flow-vocabulary.csv
One vocabulary entry per row.
No duplicates.
Preserve correct capitalization, punctuation, spacing, and hyphenation.
Include singular and plural forms only when both are genuinely used.
Keep every entry within Wispr Flow’s 60-character desktop limit.
2. wispr-flow-corrections.csv
Two columns in this order:
1. Incorrect transcription
2. Correct replacement
Include only evidence-backed corrections.
Each correct word may have only one replacement rule.
Wispr Flow desktop accepts one-column vocabulary CSVs and two-column correction CSVs. Bulk Import supports up to 1,000 entries and a 3 MB file.
If the import preview interprets a header row as vocabulary, remove the header before importing.
6. Recommend the desktop configuration
Use this baseline when evaluating the current setup:
Update Wispr Flow to the latest available version.
Select the user’s specific language and regional variant.
If the user primarily dictates in one language, select only that language instead of Auto-detect.
If multiple languages are necessary, keep the selection to two or three. Manually switching is more reliable for similar-sounding languages.
On Mac, prefer Built-in mic over Auto-detect when external devices frequently take control.
On Mac or Windows, prefer a built-in or wired microphone over AirPods or an unstable Bluetooth microphone.
Confirm that the selected microphone remains active and that input levels respond normally.
Enable Auto-add to Dictionary when the user wants Wispr Flow to learn distinctive words from spelling corrections.
Enable Context Awareness only when its app-aware accuracy benefits justify sharing limited nearby text from the active application.
On Mac, verify that Wispr Flow has the accessibility permission required for Context Awareness.
Configure Styles for personal messages, work messages, email, and other applications.
Reduce Auto Clean Up or adjust Polish rules if Wispr Flow removes meaningful language or changes the user’s voice.
Enable Settings → Experimental → Bulk Import before importing the CSV files.
Star the most important dictionary entries on desktop so Wispr Flow gives them higher recognition priority.
Do not weaken Privacy Mode, cloud-retention controls, or security settings to obtain more history.
7. Present the plan before modifying Wispr Flow
Show the user:
1. Which Claude or Codex history tools were available.
2. Approximately how many sessions and Wispr Flow transcripts were analyzed.
3. The principal transcription problems.
4. A preview of both CSV files.
5. The exact settings you recommend changing.
6. Any privacy implications.
7. Which actions can be completed through computer use.
Create and validate the CSV files without waiting for approval.
8. Offer to implement the changes
If computer use is available, offer to implement the recommendations in Wispr Flow.
Obtain explicit approval before:
Importing either CSV.
Adding, editing, deleting, or starring dictionary entries.
Changing Wispr Flow settings.
Running microphone or transcription tests through the user interface.
Treat permission to inspect as read-only permission, not permission to modify.
Account, subscription, team, privacy, cloud-retention, and security changes require separate explicit approval.
9. Implement and verify after approval
After receiving approval:
1. Enable Bulk Import if necessary.
2. Import and verify the vocabulary CSV.
3. Import and verify the correction CSV.
4. Resolve rejected rows instead of silently skipping them.
5. Preserve existing entries unless a duplicate is confirmed.
6. Apply only the approved configuration changes.
7. Test at least 10 representative phrases containing high-priority names, acronyms, specialized terms, plural forms, and similar-sounding words.
8. Record what was said, what Wispr Flow produced, and whether each test passed.
9. Correct evidence-backed failures and retest.
Finish by reporting:
What was implemented.
What was verified.
Remaining limitations or errors.
The saved locations of both reusable CSV files.
8
9 comments
Tired of WisprFlow getting it wrong?
skool.com/cliefnotes
What we give away free beats most paid courses. Build durable AI systems with a Marine vet and Edinburgh researcher. 40+ lessons, growing.
Leaderboard (30-day)
1
+1163
2
+929
3
+674
4
+598
5
+586
Powered by