Skip to content

st fix

Jump to bottom
b2o2i edited this page May 10, 2026 · 6 revisions

st-fix — Improve a story using fact-check feedback

Rewrites the weak parts of a story using its fact-check results. Only sentences scored False or Partially False are touched — everything that checked out stays exactly as the AI wrote it.

Run after: st-fact Run before: st-post

Related: st-fact st-merge st-post Multi-Model

Multi-model (0.9.0+): when --agent is omitted, st-fix defaults the rewriter to the agent whose (make, model) matches the source story — an Opus-authored story is rewritten by Opus, not by the bare anthropic handler default. See Multi-Model.

How it works

  1. You run st-fact (or st-cross) and a fact-checker AI scores every sentence True, Partially True, Opinion, Partially False, or False.
  2. You run st-fix. It finds all False / Partially False sentences and sends them to an AI for rewriting.
  3. In the default iterate mode it checks each fix immediately: if the rewritten sentence scores better, it keeps the change; if not, it leaves the original alone. No change is ever made that makes things worse.
  4. The fixed story is saved back to the container, ready for st-post or another round of st-fact.

Quick start

# Simplest use — auto-selects the most fixable story + fact-check in the container
st-fix subject.json

# Specify story 1, fact-check 1 explicitly
st-fix -s 1 -f 1 subject.json

# Use a specific AI to rewrite
st-fix -s 1 -f 1 --agent openai subject.json

# After fixing, run st-prep automatically to refresh the title/hashtags
st-fix subject.json --prep

Modes

Mode When to use it What it does
iterate (default) Most cases Fixes one sentence at a time, checks each fix immediately. Tries every available AI per sentence before giving up — keeps only verified improvements.
patch Quick pass; large batches Bundles all false claims into a single prompt. Faster but no inline verification — the AI rewrites everything at once.
best-source After st-bang / st-cross Uses the other AI stories in the container as reference material. Good when a different AI got the same fact right. Requires at least 2 stories.
synthesize After st-cross (full pipeline) Passes all stories and their scores to one AI and asks for a single best-of-all rewrite. Highest quality but uses the most tokens. 
st-fix subject.json                            # iterate (default)
st-fix --mode patch subject.json               # one-pass patch
st-fix --mode best-source subject.json         # multi-story reference
st-fix --mode synthesize subject.json          # full synthesis

Options

Option Description
-s N / --story N Story to fix (default: auto-select the story with the most false claims)
-f N / --fact N Fact-check to use (default: auto-select; not used in synthesize mode)
--mode <mode> Fix strategy: iterate (default), patch, best-source, synthesize
-a / --agent <name> AI to use for rewriting (default: same AI that wrote the story)
--checker <name> AI to use for inline verification in iterate mode (default: original fact-checker)
--prep Run st-prep after fixing to refresh the title and hashtags
--cache Enable API cache (default: enabled)
--no-cache Bypass API cache
-v / --verbose Show a diff of each change
-q / --quiet Minimal output

What to expect

When you run st-fix subject.json without any flags, it:

  • Scans every story+fact-check pair in the container and ranks them by number of fixable claims.
  • Picks the combination with the most to fix and prints a summary before making any changes.
  • Runs claim-by-claim in iterate mode: you see each sentence, which AI rewrote it, and whether the fix improved, kept, or skipped it.

If a story has no False or Partially False claims, st-fix exits cleanly and lists any other fact-checks in the container that do have things to fix.

For developers

For implementation details — the data-structure design, how each mode works internally, the inline fact-check loop, and notes on the future claim-level assembly architecture — see st-fix: Implementation Details.

Clone this wiki locally