Official · AnthropicApplied
State output and constraints clearly; request above-and-beyond behavior explicitly rather than relying on inference.
Open-source Agent Skill · Claude Code & Codex
Official guidance.
Zero prompt folklore.
PromptingCraft.com turns a development conversation into a prompt a fresh coding agent can execute. It keeps your requirements and permissions intact, consults the AI developers’ own documentation, and critiques the draft before you copy it.
Interactive walkthrough
One task, six stages.
A fixed, pre-written walkthrough of the published “Authorized implementation” example. No model runs on this page.
What the developer asked, after a longer discussion
Write a Claude Code handoff: add CSV export to the existing report view; reuse current columns, no backend change. Edits, tests and an MR are authorized; merge and deployment aren't.
Target agent named in the request: Claude Code. Executor model: not stated.
Requirements are classified and anchored to the user’s own words
- EXPLICIT_REQUIREDAdd CSV export to the existing report view
add CSV export to the existing report view
- EXPLICIT_REQUIREDUse the current columns
reuse current columns
- FORBIDDENBackend changes
no backend change
- EXPLICIT_REQUIREDAuthority granted: edits, tests, an MR
Edits, tests and an MR are authorized
- FORBIDDENMerge and deployment
merge and deployment aren't
- UNRESOLVEDEmpty-result behaviour was never discussedleft open — not invented
Methodology Classification is PromptingCraft.com’s own method, not a provider recommendation.
Only guidance that applies to the established target is used
Official · AnthropicApplied
Choose the simplest workflow that solves the task and add orchestration only when its quality benefit justifies latency and cost.
Official · AnthropicNot applied
If unsolicited additions occur, instruct the model to finish requested work and checks, then report helpful extras as suggestions.
Why not: this record applies to “Claude Sonnet 5.5 only”. The executor model was never established, so model-specific tuning is not transferred.
A first draft — the kind that looks fine at a glance
Implement CSV export, then ship it.
The draft is compared against the contract, both ways
"Ship" invents merge/deploy authority. Missing column and backend constraints.
ship
grants merge and deployment the user withheld.- Dropped:
reuse current columns
. - Dropped:
no backend change
.
The execution-ready prompt, with new ideas kept optional
Add CSV export to the existing report view using its current columns. Make no backend changes. Edits, tests and an MR are authorized; merge and deployment are not. Verify the export represents the chosen columns as valid CSV.
Optional suggestion: consult nearby export patterns; retain existing empty-result conventions if applicable rather than choosing a new one in this handoff.- Every required item is present; nothing forbidden is granted.
- The unresolved empty-result choice stays an optional suggestion, not a gate.
What it protects
A prompt can quietly change the task.
Rewrites tend to add authority, drop constraints and promote suggestions into requirements. PromptingCraft.com checks for exactly that.
| Common drift | What PromptingCraft.com checks |
|---|---|
| “Ship it” appears where only an MR was allowed | Granted and withheld permissions are carried over exactly — no new merge, deploy or publish rights. |
| A constraint from the discussion disappears | Each required and forbidden item is traced from the original into the final prompt. |
| A nice idea becomes an acceptance gate | New ideas stay under “Optional suggestion”, never as completion criteria. |
| An open question gets silently decided | Unresolved choices remain open, or one focused question is asked. |
| A good prompt gets rewritten anyway | “No material revision is necessary” is a valid outcome. |
Provenance
Every recommendation points back to its source.
The knowledge ledger holds 17 official recommendations from OpenAI and Anthropic documentation, each with its applicable model, limits and retrieval date.
Curated 2026-10-12. Shown from commit d0babf8f. Dated snapshot — not live.
Browse all sources- OpenAI2026-10-12
State the desired result and relevant supplied context; separate instructions from examples/reference data.
Applies to: OpenAI text and reasoning models
- Anthropic2026-10-12
For large document inputs, place long documents first and the query later; identify document sources and ground answers in relevant passages.
Applies to: Current Claude families covered by the living page
- OpenAI2026-10-12
Use simple direct outcome instructions and explicit constraints; try zero-shot before adding examples; omit generic step-by-step reasoning requests.
Applies to: Documented OpenAI reasoning models; historical o-series particulars stay model-specific
Four kinds of input, never mixed
You can always tell where an instruction came from.
What you asked for
Your words outrank everything else. They are preserved, not paraphrased away.
Provider guidance
Only OpenAI and Anthropic documentation can justify model-specific prompting advice.
PromptingCraft.com’s method
Contract classes, provenance tracing and the critique rubric are our own and labelled as such.
Design & SEO research
Used to build this website. Never used as evidence for how to prompt a model.
Get started
Install in a minute. Use it in any discussion.
Install the skill into a project, then ask your assistant for a prompt — “Prepare a Codex prompt for this task.”
Codex uses $promptcraft. Standalone archives, plugin packaging and removal are covered in the documentation.
git clone https://gitlab.com/revanalex/promptingcraft-skill.git
cd promptingcraft-skill
python3 scripts/install_skill.py --host claude --project-dir /path/to/your/project
# then, in Claude Code:
/promptcraftHonest limits
It writes prompts, not code.
PromptingCraft.com composes the handoff. It never executes the development task or grants permissions.
No proven superiority.
We publish our evaluations, but do not claim prompts written this way always produce better results.
Guidance ages.
Sources are revalidated when your agent can fetch them. Offline, the ledger is labelled as dated.