PromptTailor

Model-aware prompt rewriting for Claude Code. Write a rough request — get it rewritten the way your current Claude model works best.

한국어 README

Why

Claude Fable 5, Opus 5, Sonnet 5, and Haiku respond best to different prompt styles — Fable wants goals and constraints in prose (no step lists), Opus over-verifies if you tell it to double-check, Haiku wants small numbered steps. PromptTailor keeps these differences as data (model profiles), detects which model you're running, and rewrites your rough request to match — also routing by task intent (fix / build / research / refactor / docs).

Your input language is preserved: English in → English out, Korean in → Korean out.

Example

$ prompt-tailor "fix the login bug asap, users keep getting logged out" --model fable-5

Users are repeatedly logging out unexpectedly. Before fixing, investigate: exact reproduction steps (when and under what conditions does this happen?), when this started, relevant error logs or console messages, and the login/session management code structure.

Once you've identified the reproduction path and root cause, apply the minimum fix to prevent unintended logouts. Scope: session and login logic only — do not modify other features.

Validation: confirm the issue no longer reproduces through direct testing, or verify that related tests pass.

Notice what happened: vague urgency ("asap") became an investigation directive, a scope boundary, and a validation criterion — and nothing was invented. Unknowns become investigation steps; any added specifics are tagged as assumptions.

Install

As a Claude Code plugin (recommended):

/plugin marketplace add Createyouracccount/PromptTailor
/plugin install prompt-tailor@prompt-tailor

This gives you the /pm command with no path setup.

As a CLI / MCP server:

pip install prompt-tailor    # installs `prompt-tailor` and `prompt-tailor-mcp`

Requirements: Python 3.10+, the claude CLI installed and logged in (no separate API key). Verified on macOS/Linux; Windows untested.

Usage

prompt-tailor "rough request" --model fable-5    # rewrite for a target model
prompt-tailor "rough request" --json             # JSON output
prompt-tailor "rough request" --concise          # faster, condensed meta-prompt

Inside Claude Code — /pm rough request: rewrites for your session's detected model, shows a one-line change summary, then executes the rewritten request. In auto mode, add the permission rule printed by claude-code/install.sh so prompts containing risky-looking words (e.g. "docker prune") aren't false-positive blocked — the backend only rewrites text.

Hook auto mode (opt-in) — rewrite every prompt automatically via a UserPromptSubmit hook. Run bash claude-code/install.sh for the settings snippet. Escape hatch: include #raw in a prompt to pass it through untouched. Prompts under 6 tokens or over 800 chars are skipped; if a rewrite doesn't finish within 28s it fails open (your original prompt goes through).

Cursor / any MCP client — a built-in stdio MCP server exposes refine_prompt(raw, target_model, concise):

// ~/.cursor/mcp.json
{ "mcpServers": { "prompt-tailor": { "command": "prompt-tailor-mcp" } } }

claude mcp add prompt-tailor -- prompt-tailor-mcp   # register in Claude Code

Same input, different models (real outputs)

Input: 로그인 버그 고쳐줘 ("fix the login bug" — Korean in, Korean out):

Full texts in eval/results.json.

Measured cost (n=2, 2026-08-15)

Raw data: runs/cost_measurement.json, method: eval/measure_cost.py.

Evidence — including the negative result

Every claim is backed by ledgered experiments (blind pairwise LLM judging; LOOP_LOG.md):

• Prompt quality (golden set of 20 vague requests): 20/20 judged better than the original (clarity 5.0, fidelity 4.8, actionability 5.0) — EVAL.md
• Model profiles produce structurally different rewrites: 5/5; intent routing beat profile-only meta 4–1–1
• Task outcome pilot (n=3): the raw prompt won 3–0. On already-clear, self-contained codegen tasks run headless, rewriting hurt: it inflated scope, its investigation directives stalled a run, and its verification demands made the executor fabricate test results — eval/ab_task_outcome_results.json

What this means: the measured benefit is on vague, underspecified requests — the golden set's territory. Already-clear requests should be left alone, so since v0.2.0 the rewriter runs a clarity gate first: if your request is already specific it returns it untouched (action: keep; the hook then injects nothing). Gate accuracy on a balanced 40-prompt benchmark: 39/40 (98%) — vague recall 20/20, clear recall 19/20. Dataset, runner, and the one miss are documented in BENCHMARK.md.

What we do NOT guarantee

• Higher task success is not proven. Prompt-quality wins are judge-based; the only outcome data so far is the 3-task pilot above, which the raw prompt won.
• The rewriter occasionally adds specifics without an [assumption] tag (fidelity 4.8, not 5.0), and can misclassify intent.
• All experiments are small-n, single-LLM-judge, and were run in this repo's environment.

Development

python3 -m unittest discover tests   # 37 offline tests, no LLM calls
python3 eval/run_eval.py             # golden-set evaluation (spawns claude)

Project docs (Korean): PLAN.md · ARCHITECTURE.md · RESEARCH.md · gate criteria in GATES.md.

License

MIT