comol/prompt-enhancer

prompt-enhancer — improving prompts and task statements

Turns a short note or unstructured task into a detailed imperative specification. Suitable for 1C developers who need to formalize a quickly written thought into a clear task statement.

When to use

Trigger this skill on user phrases like:

• "improve the prompt", "make the prompt better / more precise / more detailed"
• "refine the task statement", "formalize this task", "polish the task"
• "structure the spec", "prepare a spec from a note"
• "expand the task description", "turn this into a detailed instruction"

The skill works with any subject area, but specifically accounts for 1C task specifics (terminology, attribute names, object names, form names).

Modes

| Mode | Trigger | What it does | |------|---------|--------------| | Inline | argument is arbitrary text | Improved prompt is printed in chat | | File | argument is a path to an existing file | Result is saved next to the source with -enhanced.md suffix | | Interactive | argument is empty | Ask the user for the prompt text or a file path |

Mode detection algorithm

The mode is selected deterministically, without heuristic on the argument text:

1. If the argument is empty — Interactive mode. Ask the user for the prompt text or a file path with a regular message.
2. Otherwise try reading the argument via the Read tool.
   • If Read returns file content — File mode.
   • If Read returns a "file does not exist / not found" error — that is the normal path, not a failure. Switch to Inline: use the entire argument as the prompt text.
3. Do not detect mode by the presence of .md in the text. A phrase like "prepare a spec about .md generation" must not be interpreted as a path. File existence is checked by Read, not by string parsing.

Transformation principles

The core of the skill. Re-check this list on every generation:

1. Source as the single source of truth. All terms, attribute names, tab names, object names, form names, subsystem names, numbers, enumerations are carried over verbatim. Do not replace Айс_КодИсточника77 with "service attribute".
2. Imperative instead of description. "needs to be displayed" becomes "Display"; "we should analyze" becomes "Analyze". The result is a plan of actions, not a story.
3. Explicit edge cases. Exception conditions are extracted into a separate sentence in the form "if ..., then ...", not hidden inside parentheses. An edge case must stand out.
4. Fixed expected output. Add a "For each item, specify" section with a numbered list of fields. State the result format (table, structured list, JSON) explicitly.
5. No new requirements. Do not add requirements that were not in the source: tests, documentation, roles, rollback, logging. Improvement is form, not content.
6. Numbering of analysis steps. Split user wishes into discrete verifiable operations 1..N. Each step — one operation.
7. Preserve intent, not literal text. If the source says "prepare a list" but then lists analytical tasks, analysis comes first, list compilation second. Restore the correct order.
8. Neutral business register. Remove conversational and emotional elements ("would be nice", "somehow", "we'd like"), keep neutral phrasing.
9. Short essence in the first line. Before the step-by-step plan — one sentence stating the task goal.
10. Formatting per repo rules. No long or short dashes (only the regular hyphen), no letter "ё" with dots, no time or ROI estimates.

Improved prompt template

Mandatory result blocks:

1. Goal — one sentence, what to do.
2. Analysis steps — numbered list of imperatives (Analyze, Study, Determine, Output).
3. Edge cases — explicit "if ..., then ..." conditions with concrete names and values from the source.
4. Output format — name of the format (table / structured list / JSON) and the list of fields the result must contain.
5. Additional clarifications (optional) — details, if the source mentioned deadlines, owners, dependencies.

Before / after example

Source prompt (short, unstructured):

The task — display all non-typical attributes of typical objects on the "Доп. реквизиты Айс" tab. Place the tab last. If only one attribute is added, Айс_КодИсточника77, do nothing. Prepare a list of objects, analyze element forms, determine where and how to add the tab.

Improved prompt (structured, detailed):

Analyze all typical configuration objects and prepare a list of those that contain non-typical attributes. Study the element forms of these objects. Add a new "Доп. реквизиты Айс" tab last in order to each form. Move all non-typical attributes onto the tab, except when the only non-typical attribute is Айс_КодИсточника77 (then no changes are made). For each object, specify:

1. Object name.
2. List of non-typical attributes.
3. Name of the element form used and current tab structure.
4. Recommended changes to tab placement and attribute movement, taking interface logic into account.

Format the output as a convenient table or structured list ready for implementation.

Additional cases — see examples/.

DO / DON'T

DO:

• Preserve names of objects, attributes, tabs, constants, enumerations verbatim.
• Turn requests into imperatives, add structure and numbering.
• Extract exceptions and edge cases into separate sentences.
• Specify the expected result format explicitly.
• Add a short essence on the first line before the step-by-step plan.

DON'T:

• Invent new requirements that were not in the source (roles, tests, rollback, printing, logging).
• Replace concrete terms with generalizations.
• Add technical solutions ("use БСП", "do this via HTTP service") unless mentioned in the source.
• Use long dash, en dash, double hyphen, or letter "ё" with dots.
• Estimate effort in hours, days, weeks, or money.

Saving the file (file mode)

• Result name: <source filename without extension>-enhanced.md. If the original is postanovka.md, the result is postanovka-enhanced.md in the same folder.
• Encoding: UTF-8.
• Do not overwrite, delete, or rename the original.
• If the result file already exists — overwrite without asking (this is a rerun of the skill).

Encoding handling

Source files are read as UTF-8. If Read returns obvious mojibake (unreadable Cyrillic like "ж©", "­", squares), inform the user that the file is likely in cp1251 or another encoding and ask to save it as UTF-8. Auto-detection of encoding is intentionally not implemented in the skill — a deliberate simplification of the first version.

Trigger phrases (for reliable detection)

The skill should recognize these user phrasings as triggers, even when the /prompt-enhancer command is not specified explicitly:

• improve / make the prompt better
• expand / refine / formalize the task statement (task, spec)
• structure the spec / task
• make the prompt more detailed / more precise / better
• prepare a detailed spec from a note / draft
• polish the task / statement
• turn this into a clear spec / detailed instruction

Limitations

• The skill rewrites form, not content. If the source prompt is logically contradictory or incomplete, the result will also be incomplete (see principle 5 — no new requirements).
• For very long source texts (2000+ words), processing may miss details. In that case, split the text into logical blocks and process them in parts.
• The skill assumes no domain knowledge beyond the source. If 1C-specifics need clarification, consult the relevant on-demand rules in this repo (e.g. content/rules/coding-standards.md, content/rules/dev-standards-architecture.md) and the 1c-metadata-manage skill (docs/query-writing.md, docs/query-optimization.md) for query and metadata details.