Always-on rules + banned words/phrases for this project, honored every reply. Plain lines ban exact phrases. Lines starting with re: are regex patterns — they catch a whole structure (e.g. re:\bnot\s+[^.!?\n]{1,90}?,?\s+but\s+ flags "not X, but Y" shapes). Critique the AI in chat, then ⋯ → “Teach slop catcher” to make it stick.
Highlight banned phrases/patterns in output
Teach The Workshop your personal “please stop writing like this” rules in normal language. Exact bans catch one phrase; pattern bans catch similar shapes; habit and character rules steer the writer.
Auto-suggest rules from my corrections
These run before the writer and feed hidden steering into the main reply. Each one can fail safely; generation keeps going.
Turn-sense (judge talk vs write)
In Auto mode, a quick cheap-model read decides whether you're asking to write a scene or just talk — instead of keyword-guessing. Adds a small beat before ambiguous replies, and falls back to keyword detection if it can't reach the model. Off = keywords only.
Length tiers steer pacing and density instead of acting like hard scene caps. API max tokens remain as a safety rail.
Retry once if the reply breaks
Show hidden passes (plan + inner voice)
Reveals what the planning and inner-voice passes actually produced for the last reply — the quickest way to check they are working and seeing your secrets.
After replies, the extractor saves verbatim details & planted threads. Backs up before each write.
Review new memory before saving
When on, new facts/secrets wait in a little review pile instead of becoming canon immediately.
Turns older than this get folded into the summary. Raise it for detail-dense chats.
Local vector search supplements lore and saved facts. It never overrides pinned facts, AU rules, or keyword lore.
No semantic index yet.
If your writing model is already a Thinking variant, set this to Off — otherwise reasoning eats your reply's length and prose comes out short.
Auto-expand short "long" replies
Follow text while it generates
Craft directives (fight flat prose)
When length is Long and a reply ends early and short, automatically write one seamless continuation to extend the scene. (Skipped when "end with options" is on, since the options footer can't be written past.)
Pushes richer, more layered prose from the first line — sensory detail, interiority, fuller paragraphs. Your slop bans keep it honest instead of purple. This is the natural way to get longer, denser scenes in one pass.
Replies under this word count get expanded — up to 3 seamless continuations — until they clear the floor. Needs auto-expand on and "end with options" off (it can't write past the options footer). When a floor is set, it applies on any length. Higher floors can invite padding; pick the highest number that still reads clean.
End scenes with numbered options
Web search (appends :online)
Checking…
Cache-friendly prompt order
Puts everything that changes each turn at the end of the request, so the unchanging part and your message history can be served from the provider’s cache instead of billed in full every time. Biggest saving on long chats. Turn off only if replies feel worse.
Cheaper model for conversation turns
When a turn is talk rather than prose — questions about the writing, planning, reacting — answer with your extraction model instead of your writing model. Scenes always get the full model. Big saving if you talk to your co-author as much as you write with it.
Tap a key to make it active. On a rate-limit (429) the app auto-rotates to the next saved key.
Added models show up in both dropdowns above. Verify slugs on openrouter.ai/models.
Fallback order: 0905 → Thinking → K2.6. Verify slugs on openrouter.ai/models.
Full precision (avoid quantization)
Requests providers serving at bf16/fp16/fp8+ only. Costs a touch more; steadier prose.
Locks routing to one provider so prompt caching stays warm — or list several comma-separated (e.g. DeepInfra, Together) to allow only those, in that order, with no fallback. Blank = OpenRouter picks. Names from openrouter.ai/models.
Omits temperature, top-p, frequency and presence from every prose request so the model uses its own built-in defaults (like the official app). The four sliders below are ignored while this is on.
0.10
0.00
0.90
1.00
Lower top-p = safer word choices; 1.00 = no nucleus cut. Tune this OR temperature, rarely both hard.
Omits the top_p parameter from the request completely (not just 1.00), in case a provider mishandles the field.
Session cost$0.0000
Adds up provider cost reported by OpenRouter for writing replies this session.