prompt-engineering
Build CLIP zero-shot class prompts, tokenizer-safe templates, and normalized text classifier weights.
- License 许可
- MIT
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- prompt-engineering
- license
- MIT
- metadata
- { "disco-role": "operating" }
- description
- Build CLIP zero-shot class prompts, tokenizer-safe templates, and normalized text classifier weights.
- disable-model-invocation
- true
02 资源文件 4
- 参考文档references/prompting-workflows.md6.0 KB
- 参考文档references/tokenization.md4.3 KB
- 参考文档references/troubleshooting.md4.4 KB
- 脚本scripts/build_zeroshot_classifier.py12.3 KB
03 SKILL.md 正文
Prompt Engineering
Use this sub-skill when an agent needs to turn class labels into CLIP text prompts, tokenize them safely, build zero-shot classifier weights, or diagnose prompt/template failures.
Route
- Start with
references/prompting-workflows.mdfor class taxonomies, template design, prompt ensembling, normalized text features, and top-k zero-shot scoring. - Use
references/tokenization.mdbefore generating large prompt sets or whenclip.tokenizeraises context-length errors. - Use
references/troubleshooting.mdfor unsafe category design, malformed templates, device mismatches, unnormalized features, and multilingual/English limitations. - Use
scripts/build_zeroshot_classifier.pywhen you need a reusable helper for expanding templates, dry-run tokenization, or building normalized classifier weights from a loaded CLIP model.
Boundaries
- Model selection,
clip.load, image preprocessing, and image-text inference setup belong in../model-loading-inference/. - Linear probes, cached image feature evaluation, and dataset-scale benchmark loops belong in
../feature-evaluation/. - Dataset downloads mentioned in source docs are reference-only; do not make prompt workflows depend on downloading raw datasets by default.
Minimal Safe Flow
- Define a fixed English class taxonomy with non-overlapping labels and avoid sensitive or deployment-risk categories unless the user explicitly asks for a research-only audit.
- Choose templates containing exactly one
{}placeholder, such asa photo of a {}.or domain-specific variants likea photo of a {}, a type of bird. - Dry-run tokenize the expanded prompts with
clip.tokenize(..., context_length=77, truncate=False)before running model inference. - Encode prompt batches with a loaded CLIP model, normalize every prompt embedding, average templates per class, normalize the averaged class vector, then stack weights as
[text_feature_dim, num_classes]. - For image classification, normalize image features and compute
100.0 * image_features @ zeroshot_weights; treat softmax/top-k scores as relative to the candidate class list, not calibrated real-world confidence.
Helper Script
Dry-run prompt tokenization without loading or downloading model checkpoints:
python sub-skills/prompt-engineering/scripts/build_zeroshot_classifier.py \
--class-name cat --class-name "electric ray" \
--template "a photo of a {}." \
--dry-run-tokenize
Build classifier weights only when a model load is explicitly requested or a caller imports build_zeroshot_classifier(model, class_names, templates) with an already loaded model. See references/prompting-workflows.md for the importable contract.
04 同仓库更多 Skills
passnet-triton-opt
Make a single PassNet Triton kernel fast AND numerically faithful: performance model (when a replacement can win at all), block/grid/warp tuning, autotune policy, launch overhead, and per-op numeric recipes to pass the dtype baseline tolerances. Use when a pass matches and is correct but speedup below expected, or when correctness fails by small numeric margins.
passnet-solve
END-TO-END playbook for solving one PassNet sample: analyze the computation graphs, decide the optimization strategy (what to fuse, what to replace, what to leave alone), drive the iteration loop, and maximize the sample score. This is the ENTRY skill — invoke it first for any PassNet optimization task; it tells you when to use passnet-pattern-fusion, passnet-triton-opt and passnet-feedback.
passnet-skill
PassNet GPU kernel optimization via compiler passes. Design and implement Triton-based optimization passes, create pass files under ./pass_dir/, self-evaluate with pass_evaluator, and iterate to maximize GPU speedup.
passnet-pattern-fusion
Author PassNet pass files: write patterns that actually MATCH the FX graph, pick fusion regions (which ops to absorb into one Triton kernel), and structure multi-pass submissions with the mandatory shared-dispatch architecture. Use when creating or fixing pattern()/replacement_args()/replacement_func() files, when a pass "failed to match", or when deciding how to fuse multiple kernels/ops.
passnet-orchestrate
Round/batch planning and decision support for PassNet graph optimization. Use for multi-sample triage, worker allocation, eval-budget policy, and post-eval keep/revert decisions. For solving one concrete sample end to end, use passnet-solve as the entry skill; consult this skill only when you need broader planning or a specific decision gate.
passnet-feedback
Fast, deterministic feedback for PassNet work: pre-flight pattern/match verification WITHOUT burning a GPU evaluation, per-node bottleneck analysis of a sample's graphs, and eval-log parsing into per-variant status + estimated score + failure classification. Use BEFORE every GPU evaluation (check_pattern), at the START of a sample (analyze_graph), and AFTER every evaluation (parse_eval_log).