Prompting guides
How to prompt the leading models — one guide per model, distilled from that vendor's own model-specific prompting page and refreshed when the guidance changes. The guides support cross-model work: building on, targeting, self-hosting, or evaluating another model. Each guide carries only the deltas that are distinctive to its model — the cross-cutting principles are below, stated once.
Filenames carry the model version (opus-4-8, gpt-5-5, glm-5-2) so a new generation adds a file rather than rewriting an old one in place.
The guides
Anthropic — the tiers the harness runs
| Model | Guide | When to reach for it |
|---|---|---|
| Claude Fable 5 / Mythos 5 | Guide | The hardest long-horizon, ambiguous, autonomous work — multi-hour runs, subagent orchestrations, unsolved problems |
| Claude Opus 4.8 | Guide | The heavy-lifting tier: reasoning, coding, agentic work where Fable 5 is overkill |
| Claude Sonnet 5 | Guide | Fast, cost-efficient tier: well-scoped coding and agentic tasks, high-volume or latency-sensitive workloads |
| Claude Haiku 4.5 | Guide | Fastest, lowest-cost tier: mechanical/bulk steps, sub-agent fan-out, scaffolding and conform-style edits |
The model pick — how to trade cost against capability — is governed by the ki-tokenomics skill, while ki-delegate turns banked planning reasoning into bounded, cold-agent-ready execution briefs.
Portable model types — one vocabulary across runtimes
The KI harness declares work by a portable model type (purpose), not a vendor's model name; each runtime resolves the type to a concrete model (ADR-KI-HARNESS-009, governed by ki-tokenomics). The type is the stable thing; the columns below are the volatile resolution.
| Type | Purpose | Claude Code | Codex (GPT-5.6) |
|---|---|---|---|
frontier |
Long-horizon, minimally-supervised autonomous execution — multi-hour runs, subagent orchestration | Fable 5 | Sol @ Ultra |
reasoning |
Hardest one-shot judgment — architecture, research, novel design | Opus 4.8 | Sol @ High/Max |
standard |
Well-scoped default — everyday coding, high-volume or latency-sensitive work | Sonnet 5 | Terra @ Medium |
fast |
Mechanical/bulk steps where full reasoning is wasted | Haiku 4.5 | Luna @ Light |
Model and effort are two independent axes on both vendors' real pickers — a single config value (reasoning = "opus, gpt-5.6-sol") can list both runtimes' models and each resolves the first it supports.
Other frontier and open-weight models — cross-model reference
| Model | Guide | When to reach for it |
|---|---|---|
| OpenAI GPT-5.5 | Guide | Building on or comparing against the OpenAI stack: coding, agentic, knowledge work |
| OpenAI GPT-5.6 (Sol / Terra / Luna) | Guide | The Codex CLI tier family (Sol/Terra/Luna × effort); the Codex-side resolution of the portable model types |
| Google Gemini 3 | Guide | The Gemini stack: long-context, multimodal, agentic work |
| GLM-5.2 (Z.ai) | Guide | Leading open-weight model for agentic/coding; self-hostable under MIT |
| DeepSeek V3.2 | Guide | Open reasoning/maths baseline; low-cost self-hosting |
| Llama 4 (Meta) | Guide | The most-deployed open-weight family; broad tooling and long context |
| Gemma 4 31B | Guide | Local/on-device, single-GPU; Google's open, runnable option |
| Qwen3 (small / coder) | Guide | Local all-rounder and local coding on consumer hardware |
| Ministral 3 (Mistral edge) | Guide | Local/edge on a laptop or single GPU; instruct + reasoning variants |
Principles
These hold across current reasoning models regardless of vendor; each guide notes where its model differs.
- Models take prompts literally. They do what is asked and won't generalise an instruction across items or infer unrequested work — to apply something broadly, say so, and state scope up front ("do X for every file", not "do X … first").
- Reasoning effort is the primary lever. A per-request effort / thinking-level setting (names vary:
effort,reasoning_effort,thinking_level) trades intelligence against latency and cost more than prompt wording does. Set it deliberately; re-measure cost when a model or its effort scale changes. - Communication is concise by default. Modern models lead with the outcome and may skip an interim status or a post-tool summary. If you need one, ask explicitly — and remove old scaffolding ("summarise every N tool calls") that assumed a chattier model.
- Don't force the model to echo its reasoning. Read structured
thinking/reasoning blocks for visibility; instructing the model to reproduce or explain its internal reasoning in the response is unnecessary and can trip a safety classifier on some models. - Give
max_tokensheadroom. Where the response cap also covers thinking, a tight limit truncates output (stop_reason: "max_tokens"). Size it generously on long agentic runs.
Anthropic-specific notes
These apply to the Claude guides only:
budget_tokensis gone; thinking is adaptive. Manual extended thinking (thinking: {type: "enabled", budget_tokens: N}) now 400s on Claude 5-class models. Set adaptive thinking viaeffort; whether adaptive thinking is on by default differs by model (off on Opus 4.8, on on Sonnet 5).- Refusal / fallback. Fable 5 runs safety classifiers and can return a
refusalstop reason; configure a fallback to Opus 4.8. See the per-model guides.
Refreshing
Each guide is distilled from the vendor pages its Sources table lists — tagged and dated. To update a guide, re-read those pages and re-date the "Last reviewed" column. The cross-cutting Anthropic reference is Prompting best practices.