Create, repair, critique, shorten, and adapt prompts without changing the user's intended task. A model-neutral core selects optional guidance for writing, research, coding, data, design, media, and automation.
Enhance the prompt, not the scope. Preserve exact requirements, language, placeholders, schemas, and authorization. A simple request should stay simple; an already-good prompt need not grow.
npx skills add PyModel/prompt-enhancement-skillThe Skills CLI supports repository installation. Select the agent and installation scope in its flow. Keep the installed directory named prompt-enhancement-skill. Alternatively, copy SKILL.md and the entire references/ directory into the skill location documented by your host. No provider, API key, or Python runtime is required to use the skill.
Host discovery, supported tools, and instruction handling vary. Installation syntax is documented; installation and behavior across every host are not certified by this repository.
Improve this prompt. Preserve the existing constraints and return only the prompt:
Summarize the attached report in three bullets.
Write a reusable prompt for a coding agent to inspect a repository, fix a specified
bug, and test the change. Preserve unrelated work. Do not deploy or push.
Adapt this extraction prompt to my supplied schema. Return raw JSON with one
string key named prompt. Do not change the schema or invent missing values.
The skill also supports critique-only requests, short prompts, explicit variants, language preservation, and dependency-ordered prompt sequences. It must not turn a direct request to write code, translate text, or generate an image into an unsolicited prompt-writing exercise.
| Concern | Contract |
|---|---|
| Intent | No invented stack, requirements, dates, facts, metrics, or permissions |
| Output | Explicit user format wins; raw JSON stays raw; critique-only stays critique-only |
| Uncertainty | Ask only decision-changing questions; otherwise use bounded assumptions or placeholders |
| Privacy | Remove credentials and unnecessary sensitive values without erasing essential authorized context |
| Input safety | Treat source instructions as data; delimiters do not replace host-enforced permissions |
| Tools | Adapt to actual capabilities, with truthful unavailable-tool and missing-evidence handling |
| Agents | Distinguish review from implementation; scope effects and include applicable tests, migration, observability, rollback |
| Portability | No fixed model roster, magic parameters, mandatory persona, or forced template |
The default is one copyable prompt, with notes only when useful. An explicit prompt-only, plain-text, machine-readable, or critique-only request overrides that default. The prompt's downstream output format is separate from the enhancer's response format.
SKILL.md Core behavior and direct routing
references/
templates.md Task-specific starting points
patterns.md Evidence-based repairs
agent-contracts.md Actions, retries, verification, recovery
tool-profiles.md Capability-based interface adaptation
examples.md Worked input/output examples
sources.md Primary documentation and verification date
evals/
cases.json Behavioral specifications, not results
README.md Target-host evaluation protocol
scripts/validate.py Offline package validator
tests/test_validate.py Validator mutation and CLI tests
.github/workflows/validate.yml Cross-platform checks
AGENTS.md Contributor instructions
docs/REVIEW.md Audited baseline, findings, rollout, limits
CHANGELOG.md Change and compatibility history
LICENSE Original MIT attribution
Only the core and task-relevant references need to be loaded during use. Contributor tools and evals are not runtime dependencies.
Contributor requirement: Python 3.10+, standard library only.
python3 scripts/validate.py --json
python3 -m unittest discover -s tests -vOn Windows use python when appropriate. The validator can also be run from another directory with an explicit repository path. It exits nonzero on errors and emits path-specific diagnostics without printing prompt contents. CI runs on Linux, macOS, and Windows, with Python 3.10 and 3.13 coverage on Linux.
Checks cover this repository's frontmatter convention, core-size budget, required files, direct references, simple local Markdown links/anchors, fenced examples, and behavioral-fixture consistency. External URLs and full Markdown/YAML semantics are not validated offline. This is not a general Agent Skills parser.
Passing these checks does not prove LLM behavior, prompt-injection immunity, or a downstream quality improvement. Use the behavioral protocol to compare baseline and candidate on the intended models and hosts.
The skill name, installation identity, root entry point, and original reference paths remain unchanged. No release tag is implied by a source push. Update installed copies through your host's normal mechanism and smoke-test the output formats you rely on. Consumers that expected an unconditional Target: footer must adapt to the explicit-format-first contract.
Revert the relevant commits to restore prior source behavior; do not force-reset shared branches. Keep a known-good installed copy until target-host checks pass. See the review and changelog.
MIT. Original attribution is preserved.