prompt-fixer
Resolve intent, compile context-aware prompts and direct builders toward checked delivery with less wasted context, questioning and rework.
Safety scan
A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.
Badge for your README
Markdown
[](/listings/carlosantano--prompt-fixer)HTML
<a href="/listings/carlosantano--prompt-fixer"><img src="/api/listings/carlosantano--prompt-fixer/badge.svg" alt="findsafeskills safety rating" /></a>Install
Install instructions from the README
As written by the repo's author. findsafeskills has not verified these steps.
Clone, then install into **one existing project**. Python 3.12+ is needed only for the optional offline helper; no pip install is required.
git clone https://github.com/CarloSantano/prompt-fixer.git cd prompt-fixer python -B -m prompt_fixer install --host codex --project "/absolute/path/to/your-project" --dry-run python -B -m prompt_fixer install --host codex --project "/absolute/path/to/your-project" python -B -m prompt_fixer doctor --host codex --project "/absolute/path/to/your-project"
Replace the example path with your real project. On Windows use a quoted path such as `"C:\Users\you\Projects\my-app"`; `py -3.12` may replace `python` if that is how Python is installed. For Claude Code replace `--host codex` with `--host claude` in each command.
| Host | Installed project folder | Invocation after discovery | | --- | --- | --- | | Codex | `.agents/skills/prompt-fixer/` | `$prompt-fixer` plus your request | | Claude Code | `.claude/skills/prompt-fixer/` | `/prompt-fixer` plus your request |
Open or reload your project in the host and confirm `prompt-fixer` is in its skills list. Try:
$prompt-fixer build me an internal sales tool; choose the technical approach.
Use `/prompt-fixer` in Claude Code. Success means a visible comparison appears without building the underlying app. Say "execute that prompt" to proceed.
The installer refuses unmanaged or user-edited installs and changes no global settings. `doctor` verifies files and receipt, not discovery by a running host.
**Without Python:** copy the entire `skills/prompt-fixer/` folder, including its references, into the host's project folder above. Manual copying creates no installer receipt; manage later updates yourself.
What was scanned
The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.
plugin.json · Manifest · 29 lines
- {
- "$schema": "https://agent-plugins.org/schemas/1.0.0/plugin.schema.json",
- "author": {
- "name": "Carlo Santo v3.0"
- },
- "description": "Resolve intent, compile context-aware prompts and direct builders toward checked delivery with less wasted context, questioning and rework.",
- "extensions": {
- "com.openai": {
- "interface": {
- "capabilities": [],
- "category": "Productivity",
- "defaultPrompt": [
- "Paversk mano laisvai parašytą mintį tikslingu darbu: parink vaidmenį, kontekstą, klausimus ir sėkmės patikrą.",
- "Padėk išgryninti, ką iš tikrųjų noriu sukurti, ir pradėk nuo svarbiausio žingsnio.",
- "Peržiūrėk šią užduotį: ar teisingas tikslas, darbo būdas ir rezultato patikra?"
- ],
- "developerName": "Carlo Santo v3.0",
- "displayName": "Prompt Fixer",
- "longDescription": "Prompt Fixer padeda laisvai parašytą mintį paversti tikslingu promptu arba vykdomu darbu. Pirmiausia atskiria klausimą, prompto perrašymą ir vykdymą. Išsaugo kontekstą, reikalavimus ir ambiciją, parenka naudingą atsakomybę bei metodus. Vibecoderiams padeda rinktis tinkamą techninį kelią, pristatyti naudojamą rezultatą ir sustoti pasiekus sutartą kokybę. Taupo kontekstą ir dėmesį vengdamas kartojimosi; faktinės tokenų ar pajamų naudos nežada be matavimo. Pritaikomas verslo ir asmeninėms užduotims. Aiškus @Prompt Fixer su padrika užduotimi pateikia pagerintą promptą; vykdo paprašius. Automatinis parinkimas priklauso nuo aplinkos.",
- "shortDescription": "Nuo minties iki aiškaus darbo"
- }
- }
- },
- "name": "prompt-fixer",
- "version": "0.7.0",
- "repository": "https://github.com/CarloSantano/prompt-fixer",
- "homepage": "https://github.com/CarloSantano/prompt-fixer#readme",
- "license": "MIT"
- }
README.md · README · 150 lines
- # Prompt Fixer
- [](https://github.com/CarloSantano/prompt-fixer/actions/workflows/ci.yml)
- [](LICENSE)
- **Rough idea → visible prompt improvement → execute when you ask.**
- An instruction layer for your existing AI assistant. For builders who want to
- ship useful products, services and automations without learning prompt syntax.
- No extra server, API key or model call from this package. Your AI host's access,
- tools, permissions and usage costs still apply.
- [Lietuviškai: pradžia draugui](docs/QUICKSTART_LT.md) ·
- [Download the release](https://github.com/CarloSantano/prompt-fixer/releases/latest) ·
- [Evidence and limitations](qa/README.md)
- ## What you see
- 1. **Your request** — your actual words; long inputs use a marked excerpt.
- 2. **Improved prompt** — separately copyable instructions in your language.
- 3. **What changed** — one short account of concrete changes.
- A short status identifies rewriting versus execution. Ask for "only the final
- prompt" to skip the comparison. Specific user instructions remain in control.
- | Request | Behavior |
- | --- | --- |
- | Invoke with "build me a client-request app; choose the technology" | Show an improved prompt and stop before building. |
- | "Improve and execute now: add a CSV export to this project" | Inspect the available project and perform authorized work. |
- | "Does Prompt Fixer need a backend?" | Answer the question directly. |
- | Invoke without a task | Ask for the task. |
- Ordinary implicit use preserves the original mode. Automatic selection is
- host-dependent. This skill is not an interceptor or cross-chat memory service.
- ## Quick install: Codex or Claude Code
- Clone, then install into **one existing project**. Python 3.12+ is needed only
- for the optional offline helper; no pip install is required.
- ```sh
- git clone https://github.com/CarloSantano/prompt-fixer.git
- cd prompt-fixer
- python -B -m prompt_fixer install --host codex --project "/absolute/path/to/your-project" --dry-run
- python -B -m prompt_fixer install --host codex --project "/absolute/path/to/your-project"
- python -B -m prompt_fixer doctor --host codex --project "/absolute/path/to/your-project"
- ```
- Replace the example path with your real project. On Windows use a quoted path
- such as `"C:\Users\you\Projects\my-app"`; `py -3.12` may replace `python` if
- that is how Python is installed. For Claude Code replace `--host codex` with
- `--host claude` in each command.
- | Host | Installed project folder | Invocation after discovery |
- | --- | --- | --- |
- | Codex | `.agents/skills/prompt-fixer/` | `$prompt-fixer` plus your request |
- | Claude Code | `.claude/skills/prompt-fixer/` | `/prompt-fixer` plus your request |
- Open or reload your project in the host and confirm `prompt-fixer` is in its
- skills list. Try:
- ```text
- $prompt-fixer build me an internal sales tool; choose the technical approach.
- ```
- Use `/prompt-fixer` in Claude Code. Success means a visible comparison appears
- without building the underlying app. Say "execute that prompt" to proceed.
- The installer refuses unmanaged or user-edited installs and changes no global
- settings. `doctor` verifies files and receipt, not discovery by a running host.
- **Without Python:** copy the entire `skills/prompt-fixer/` folder, including its
- references, into the host's project folder above. Manual copying creates no
- installer receipt; manage later updates yourself.
- ## ChatGPT Work Mode
- If your account supports creating custom plugins, download the portable ZIP
- from [Releases](https://github.com/CarloSantano/prompt-fixer/releases/latest),
- attach it in Work Mode with Plugin Creator available, and ask it to create a
- **private personal plugin from the attached archive**. Select your own Prompt
- Fixer plugin and mention it with your request.
- This depends on your account's features. An ordinary chat does not automatically
- install a repository. The author's private plugin grants no access to another
- account. This is a portable source distribution, not a marketplace listing.
- ## Update or remove
- Run `git pull --ff-only` in this source checkout, then rerun installation with
- `--replace`. An unchanged managed install is backed up in the target project's
- `.prompt-fixer/backups/` before replacement.
- ```sh
- python -B -m prompt_fixer install --host codex --project "/absolute/path/to/your-project" --replace
- python -B -m prompt_fixer uninstall --host codex --project "/absolute/path/to/your-project"
- ```
- Removal moves unchanged managed files to a recoverable backup. Reconcile user
- edits first; the helper refuses to overwrite or remove them. No auto-update or
- global configuration is added.
- ## Offline checks
- ```sh
- python -B -m unittest discover -s tests -v
- python -B -m prompt_fixer check --evidence
- python -B qa/evalkit.py validate
- python -B qa/downstream_checks.py candidate ingest
- python -B qa/downstream_checks.py candidate approval
- python -B tools/build_release.py
- ```
- These commands make no API/model calls. Run outputs stay outside the source
- bundle. CI checks Windows and Linux. Historical response replay is separate from
- correspondence to current recorded evidence; neither executes a new AI model.
- Instruction or helper-code changes invalidate current evidence.
- Evidence includes a small same-model study and targeted fresh-context checks.
- It does **not** establish human validation, universal activation, production
- integrations, billed-token savings or improved revenue.
- To compare actual caller-supplied host/API usage for the whole path:
- ```sh
- python -B -m prompt_fixer measure examples/usage-ledger.json
- ```
- Count fixer, execution and repair steps. The empty template reports unknown
- usage; missing counts are never zero. Successful outcomes, matched conditions
- and the same execution model are required. Cache/reasoning counts are subsets,
- not extra totals. The helper cannot capture telemetry, verify outcomes or detect
- omitted steps independently. Never publish private task data in a usage ledger.
- ## Design
- Resolve intent, preserve constraints and selectively load references for roles,
- questions, architecture, methods and completion. Distinguish assumptions from
- requirements; ask questions that change a meaningful decision. Personal work
- needs no business interview. Roles identify useful responsibilities, not
- fictional experience. Approvals retain the exact target/recipient and reviewed
- content/version. Prompt text alone does not enforce production security.
- Skill locations follow the [Codex documentation](https://learn.chatgpt.com/docs/build-skills)
- and [Claude Code documentation](https://code.claude.com/docs/en/skills), checked
- 2026-11-21. Verify actual discovery in your own session.
- See [contributing](CONTRIBUTING.md), [changes](CHANGELOG.md),
- [provenance](NOTICE.md), [public-distribution notes](docs/PUBLIC_DISTRIBUTION.md)
- and the [MIT license](LICENSE).
Version history
v0.7.011/22/2026
Ratings and comments
Sign in to rate this skill.
Sign inFree to use. If it’s helped, pass it on.