prompt-router-starter

A design pattern you can use, not a complete solution. A prompt injection, triggered from a UserPromptSubmit hook, decides how the reply should look before the model answers: it scores the prompt's shape against readable rules and injects one output style beside the prompt. The same hook carries a numbered requirements ledger, and a Stop hook runs the proofs and refuses a false done. This is a demo of the pattern, not a working solution: the styles are placeholders, the mechanism is the point. Python 3.10+, stdlib only, nothing to install. Claude Code only. Tests: python -m pytest tests -q.

Setup

git clone https://github.com/Glitch-Cat-Club/prompt-router-starter
claude                    start it in the folder you just cloned
type a question           the reply comes back in the skin the rules picked

What you'll see

The three hooks

router.py    UserPromptSubmit: scores the prompt, injects the skin and the ledger
gate.py      Stop: reads the contract table back, runs the proofs, refuses a false done
restore.py   SessionStart: hands the contract back to a session after /clear

How it routes

Six of the shape signals in algorithm/rules.json. Each returns 0.0 to 1.0, each skin weights the ones it cares about, and the top total over the 0.35 threshold wins.

Nothing clears the threshold and the default skin answers. Two skins within 0.10 of each other and both get named, so the model picks from the conversation.

Skins

bluf      one decision, the answer first and the reasons under it
sitrep    where the work stands, the whole board on one screen
contract  the requirements you just heard, back as a table to approve
bars      the build running, one row per requirement, gates as they catch
board     the requirements as a kanban: open, building, proven
receipt   the contract closed off, printed like a till receipt
minimal   an acknowledgement or a rating, answered in one line
banter    an insult, answered with one dry line

These eight are demo styles, not optimised outputs: they prove the mechanism. Write your own with your AI, one file in skins/ (front matter, then the shape), one line in algorithm/rules.json.

The contract

Say everything you want in one go and it comes back as a numbered table, one proof a line. Strike or reword any line by number, then say go: the ledger sits in .router/, so a /clear never loses it. Claim done and the Stop gate runs the proofs before the turn ends.

The gate runs allowlisted commands only (pytest, python, python3, git, grep, rg, ls, test, node, npm test, curl to localhost); anything else is stamped manual and never runs. A red line holds the turn open with one sentence naming it. A proof written <command> → 0 passes when the command prints nothing, because grep-style tools exit 1 on no match. demo/site/ is a prop project, so every proof in the demo is a real command.

Force it

/bluf /sitrep /receipt /board /minimal   the skin by name
/e1 /e2 /e3 /e4 /e5                      the depth, E1 shortest to E5 deepest
sitrep receipt board go                  the bare word, alone, as the whole prompt
default sitrep                           the house floor for this project
default off                              clear it

The slash tokens go at the END of the prompt and beat every score.

The haiku switch

Set "classifier": "haiku" in algorithm/rules.json and the skin comes from claude -p --model haiku instead. Off by default: one Haiku call a prompt on your subscription, API-key variables stripped from the child, back to the rules on any failure.

Tune it

python -m router.report          from the repo root

Top-1 accuracy over tests/golden.jsonl, precision and recall per skin, the confusion table, the ten worst misses with their top-2 scores. The rules here score 0.9595 on the 148-line set. Demo-tuned rules: the shape is the lesson, tune the words to yours.

Windows

Python on PATH. Git Bash or PowerShell for the proofs, whichever you have.

Hooks run on first launch

The hooks in .claude/settings.json are a project setting, so they run the first time you start Claude Code here. The three files above live in .claude/hooks. Read them first.

Licence

MIT. The block letters use the ANSI Shadow FIGlet font, vendored from xero/figlet-fonts.