The 30-minute manuscript test: start here
Version 1.0 | August 2026

Files in this resource set:

- manuscript-test-fixture.txt: the two-scene synthetic manuscript input
- manuscript-test-answer-key.txt: known signals, errors, and false-positive traps
- manuscript-test-prompts.txt: six standardized prompts
- manuscript-test-scorecard.csv: scoring worksheet for comparing runs

What the six tests cover

The protocol contains six non-generative tasks:

1. Evidence recovery: identify Mara's immediate want in each scene, quote evidence, and separate explicit text from inference.
2. Continuity risk: compare scenes, classify possible conflicts, identify explained differences, and distinguish Mara's stated detachment from her contradictory actions as intentional character pressure rather than a continuity error.
3. Scene diagnosis: identify immediate goal, opposition, turn, choice or withholding, and changed condition, with evidence for every claim and no continuation or rewrite.
4. Question quality: ask five revision questions that preserve possibilities and return decisions to the writer.
5. Voice-respecting feedback: analyze the effect of "This was not sentiment. It was admin." without rewriting it or making it more literary.
6. Workflow cost and restraint: audit grounding, interpretation, uncertainty, overreach, compliance, and missing context in no more than 250 words.

The scorecard contains one row for each test and records protocol version, tester, product, model, date, run number, memory, retrieval, settings, eight rubric dimensions, time, corrections, notes, and raw-output location.

Recommended order:

1. Read the methodology in the ClaraMuse guide.
2. Review the prompts and scorecard before testing.
3. Supply only manuscript-test-fixture.txt to the writing assistant.
4. Run each prompt in a fresh, documented session.
5. Preserve the raw output.
6. Score the output before consulting the answer key.
7. Repeat the test before drawing conclusions.

This fixture evaluates a bounded close-reading workflow. It does not prove that a tool is good or bad at every writing task.