Model Prompting
ASecurityPer-model prompting profiles for Claude Code skills and agents - how Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 4.5 read instructions, where each one fails and the wording that fixes it, plus the rules for a file that runs on several models. Use when choosing `model:` or `effort:` for a skill or agent, writing or tuning one for a named model, porting one to another model, or when an agent misbehaves on one model only (stops early, overreaches its scope, under-reports findings, skips parallel c...
- 6 stars
- 0 votes
- 0 copies
- 1 view
- Added October 7, 2026
Works with
Security analysis
100/100Pro scans all 6 files and shows the line behind each finding
npx -y skills add p2p2sp/p2p2.claude --skill model-prompting --agent claude-codeAre you the author of Model Prompting?
Add the live security badge to your README. It updates with every re-scan.
[](https://www.skillsdirectory.com/skills/p2p2sp-model-prompting)---
name: model-prompting
description: Per-model prompting profiles for Claude Code skills and agents - how Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 4.5 read instructions, where each one fails and the wording that fixes it, plus the rules for a file that runs on several models. Use when choosing `model:` or `effort:` for a skill or agent, writing or tuning one for a named model, porting one to another model, or when an agent misbehaves on one model only (stops early, overreaches its scope, under-reports findings, skips parallel calls, fabricates to fill a format). Not the general authoring doctrine.
---
# model-prompting
Knowledge as of 2026-10-01. A model not listed here is not covered: say so, never extrapolate from a neighbour.
## Input
- The skill or agent file being written or tuned, or a role to staff.
- Optional: the target model or models, the observed misbehaviour.
## Workflow
1. Resolve the target models: the file's `model:` frontmatter, the model its caller dispatches it with, or the user's request. `inherit`, an absent `model:`, or a skill that runs on the session model means several models.
2. Read `${CLAUDE_SKILL_DIR}/references/cross-model.md` on every run: only it holds the rules several models share and the frontmatter mechanics.
3. One target model: also read its profile. Several or unknown models, or choosing a model or effort for a role: add a model's profile only to handle a quirk that model shows.
- Fable 5.1: `${CLAUDE_SKILL_DIR}/references/fable-5-1.md`
- Opus 5.5: `${CLAUDE_SKILL_DIR}/references/opus-5-5.md`
- Sonnet 5.5: `${CLAUDE_SKILL_DIR}/references/sonnet-5-5.md`
- Haiku 4.5: `${CLAUDE_SKILL_DIR}/references/haiku-4-5.md`
4. Apply only the mitigations the file's task exposes it to: an unattended-loop fix goes to an unattended agent, a recall fix to a reviewer. Never paste a whole profile into a file.
5. Write each mitigation in the target file's own voice, merged into the rule it modifies when one exists. Quoted snippets are measured wording: shorten one only by whole clauses, and keep verbatim a sentence the profile marks so.
## Output
- The edits to the target file, or the model and effort recommendation for the role.
- Per addition, one line naming the model behaviour it addresses.
Files in this skill
- SKILL.md
- references/cross-model.md
- references/fable-5-1.md
- references/haiku-4-5.md
- references/opus-5-5.md
- references/sonnet-5-5.md
Attribution
Comments
Loading comments…
Newsletter
Get the weekly list of safe new skills, and the ones we pulled.