Evaluating Heuristic Tasks

You evaluate prose paragraphs against FPR style rules and produce a JSON file of suggestions. This skill is invoked by the editing-documents skill after the deterministic pass exports a _heuristic_tasks.json .

THIS IS A RIGID SKILL — Follow the protocol exactly.

Input

A JSON file (typically *_heuristic_tasks.json ) containing an array of objects:

[ { "paragraph_index": 42, "text": "The paragraph text from the document...", "language": "en", "prompt": "Full evaluation prompt with rules context..." } ]

Protocol

Step 1: Read the heuristic tasks file

Use the Read tool to load the JSON file. Note the total count of paragraphs.

Step 2: Evaluate each paragraph

For each task in the array:

• Read the prompt field — it contains the full evaluation context including active rules

• Evaluate the paragraph text against the rules described in the prompt

• For each style violation found, create a suggestion object

Step 3: Write the results file

Write a JSON file with ALL suggestions from all paragraphs:

[ { "paragraph_index": 42, "original": "exact minimum text to replace", "replacement": "suggested replacement", "rule_id": "NNQ-002", "confidence": 0.75, "rationale": "One-line explanation of why this change improves the text" } ]

Save to the same directory as the input file, with name *_heuristic_results.json (replace _tasks with _results in the filename).

Step 4: Report completion

Tell the user:

• How many paragraphs were evaluated

• How many suggestions were produced

• The path to the results file

• Instruct them (or the editing-documents skill) to run the apply step

CRITICAL CONSTRAINTS — Violations are failures

original field = minimum text. If the problem is one word, original is that one word. Never put the entire sentence in original .

Never upgrade certainty. Do NOT change: may→will, aims to→ensures, can help→guarantees, seeks to→achieves, is intended to→delivers. If you see cautious language, LEAVE IT.

Never touch protected terms. Protected terms are project-specific and listed in projects/{PROJECT}/protected-lexicon.yaml . The prompt context mentions key protected phrases. Do not suggest changes to any of them.

Confidence must be calibrated:

• 0.90-1.0: Clear grammar error, obvious typo, unambiguous terminology fix

• 0.75-0.89: Clarity improvement, sentence structure fix, AI pattern rewrite

• 0.60-0.74: Subjective improvement, minor style preference

• Below 0.60: Do not include — it will be ignored

Preserve FPR voice. If a change makes text shorter but flattens the institutional voice (removes systems framing, equity language, or transformation narrative), do NOT suggest it.

Empty array is valid. If a paragraph has no problems, produce no suggestions for that paragraph. Not every paragraph needs a change.

JSON format must be exact. Each suggestion is a flat object with exactly these 6 keys: paragraph_index , original , replacement , rule_id , confidence , rationale .

Replacement "" (empty string) means deletion. Use this for filler words or phrases that should be removed entirely.

References

• style-rules.md — FPR voice, clarity, AI-humanizer rules condensed

• examples.md — real input/output evaluation pairs showing correct behavior