vemodalen-x

@vemodalen-x source repo

30 skills
0 followers
7 hours ago last updated

▌
Announcing Framework Releases · vemodalen-x bundle
Announce a governance framework's formal version release to the team chat as a Feishu (Lark) interactive card — framework name + code, old→new version, change-class summary (Added / Changed / ⚠️ BREAKING) extracted from its CHANGELOG, a one-line consumer-impact verdict (sync now? breaking?), repo + CHANGELOG links, maintainer, plus an optional 🏆 cumulative contribution leaderboard (instance-gated by include_leaderboard, computed from the same hub ledger as announcing-skills). Use after a framework release is tagged and push-verified. Identity-decoupled — group and the maintainer/contributor open_id maps are instance values; repo URL is resolved at runtime from the submodule remote.

0
installs
▌
Contributing Framework Changes · vemodalen-x bundle
Contribute a local improvement to a governance framework back to that framework's repo via a pull request. Use when a local framework or skill-home change should be proposed upstream. Works for any contributor — write-access holders push a branch to the repo directly; contributors without write access fork and open a cross-repo PR. The path is chosen at runtime by probing the actual permission; no identity, org, or account is hardcoded.

0
installs
▌
Rendering HTML Eval Reports · vemodalen-x bundle
Render pre-computed results into a single self-contained HTML report (CSS inlined, images base64-embedded, mandatory provenance header) in one of two report types — an EVALUATION report (per-class accuracy vs a threshold, a confusion matrix including the abstain/rejection class, a latency distribution, an error-case gallery with overlays) or a TRAINING-EXPERIMENT / ablation report (experiment-ladder table, inline training curves, ablation table, measured-vs-inferred labelling, limitations). Use when an eval run or a training-experiment sweep has produced results that should become a shareable local HTML artifact. Render-only — it consumes results, it does NOT run inference, train, or compute them; the eval type collects ALL error cases (samples only the correct predictions); it embeds only de-identified images per the instance privacy policy; the HTML is a local artifact, not committed to git. Distinct from publishing-deliverables (which publishes a doc to the Feishu/Lark wiki).

0
installs
▌
Designing Diagnostic Prompts · vemodalen-x bundle
Design structured diagnostic or tutoring prompts and agent workflows that interview the user, locate constraints, configure style/depth, produce a plan, and loop through assessment and feedback. Use when asked to create or improve a Human 3.0-style self-discovery prompt, a Mr. Ranedeer-style tutor prompt, a coaching/custom GPT prompt, an onboarding interview, a diagnostic agent, a personalized learning flow, or any multi-turn prompt that must ask questions before advising.

0
installs
▌
Packaging Device Sdk Releases · vemodalen-x bundle
Assemble an algorithm library into a deliverable on-device (mobile) SDK release package — the release version is assigned by `bumping-library-versions` (referenced, not restated), plus a standard package layout (minimal public headers / per-ABI libs / models each license-annotated / RELEASE_NOTES / USAGE doc / a compilable examples/ source / THIRD_PARTY), and two ship-along reports (quality + performance) that REUSE the eval-report and on-device-validation skills rather than re-spec them. Owns the package-specific parts: the memory = system-delta measurement method, license-boundary annotation, and a manifest + sha + unpack-reverify check (incl. verifying the version surfaces agree). Packaging is NOT releasing — shipping the package outward is a human/lead decision. Use when an algorithm lib must become a versioned, auditable phone SDK package. Identity-decoupled — lib name, version, platform, group are caller parameters.

0
installs
▌
Visualizing Processing Pipelines · vemodalen-x bundle
Generate a single self-contained HTML report that visualizes and explains a multi-step processing pipeline. Each stage gets a before/after drag-to-compare slider, a difference heatmap, inline base64 images, what/why/formula annotations, timing bars, and pass/fail metrics. Use when the user wants to visualize, explain, debug, document, or present an image / data / ML pipeline; build before/after comparison sliders; create an algorithm walkthrough or a parameter-tuning playground; or turn scattered intermediate results into one shareable file. Covers both a static exported .html and an interactive server with live parameter sliders.

0
installs
▌
Reviewing Cpp Code · vemodalen-x bundle
Review C/C++ source for coding-standard conformance and compiler-warning risks, with an embedded / DSP / image-processing lens (type safety, undefined behavior, resource management, stride/alignment hazards). When the project mounts a coding spec, reviews against that spec as the authority (read at runtime, cited by section); falls back to a generic checklist otherwise. Use when a C/C++ file or change should be checked before commit. Read-only analysis — reports findings, does not edit.

0
installs
▌
Naming Skills · vemodalen-x bundle
Validate and standardize a skill's name + description against the authoring naming rules — name ≤64 chars, lowercase letters/digits/hyphens only, no leading/trailing hyphen, name == parent folder, gerund (verb+ing) form; description ≤1024 chars, non-empty, says what-it-does + when-to-use with trigger keywords. Use when authoring, renaming, or publishing a skill, or auditing the home for naming conformance. Read-only validator — reports pass/fail per rule, does not rename.

0
installs
▌
Announcing Skills · vemodalen-x bundle
Announce newly-registered skill(s) to a team chat as a celebratory Feishu (Lark) interactive card — a 上新表 (version / category / summary / @contributor) plus an optional 🏆 cumulative contribution leaderboard read from the hub's contribution ledger (instance-gated by include_leaderboard). Use after a skill-hub release adds skills. Identity-decoupled — group, repo URL, and open_id map are instance values.

0
installs
▌
Publishing Skills · vemodalen-x bundle
Publish a standard skill package into this shared skill home — agree on its functional category, place it under skills/{category}/{name}/, activate it in skills/index.json, then maintain the README and verify binding. Use when adding, moving, or renaming a skill in VEMO_SKILLS. Placement + registration only — adoption stays a user-consent decision.

0
installs
▌
Reviewing Decisions · vemodalen-x bundle
Reliability ruleset for technical design decisions — a lightweight design-review discipline. Each key decision must carry 6 fields (options/evidence/trade-offs/assumptions/failure-modes/validation); a cross-model red-team pass critiques the doc before publish. Use when the research-solution agent finalizes the solution_document. Owned by Framework 1 (research). Generic; reviewer models/decisions are instance values.

0
installs
▌
Bumping Library Versions · vemodalen-x bundle
Bump a device/library's four-segment version after an acceptance build+run passes, and keep the version single-sourced. The four segments X.Y.Z.W are embedded in three agreeing surfaces — a source string constant, the init log line, and a getVersion() API. Bump rules: W (last) = bug-fix +1; Z (penultimate) = feature +1 and reset W to 0; both-in-one-release = Z +1 and reset W to 0; X.Y (first two) = major/platform-line, human-set only. Precondition: the acceptance build+run has already succeeded. After bumping, report old → new explicitly. Use when releasing a library and the version must advance. Identity-decoupled — the version-string field name + file are caller parameters; no project names embedded.

0
installs
▌
Syncing Frameworks · vemodalen-x bundle
Check whether any governance framework pinned as a submodule has advanced upstream (new release/commit), report the diff, and apply pin bumps only on a user version gate. Use at session start (continuity check) or when the user asks to check for framework updates. Submodule model — NOT file-copy.

0
installs
▌
Loading Model Checkpoints · vemodalen-x bundle
Load a PyTorch checkpoint into a model robustly when the state_dict nesting, key prefixes (model./module.), architecture, or input-channel count are uncertain. Use when a checkpoint silently loads onto random weights, load_state_dict reports missing/unexpected keys, a checkpoint comes from a different trainer (Lightning / DataParallel / raw torch), or you must infer the architecture and input channels before building the model. Picks the prefix variant by maximum key-overlap, prints missing/unexpected diagnostics, and flags the weights_only security caveat. Triggers: load a checkpoint, state_dict mismatch, missing/unexpected keys, strip module. prefix, wrong weights loaded, ckpt onto random weights.

0
installs
▌
Gating Tflite Op Envelopes · vemodalen-x bundle
Statically gate a TFLite model against one or more target runtime envelopes — parse the flatbuffer (no runtime load) for custom ops + declared min_runtime_version, and emit PASS / REJECT with the offending operators and version per envelope. Use when screening a candidate .tflite or .task bundle for a runtime before adopting it (e.g. a host runtime and a device runtime), or recording a model card's runtime-compatibility verdict.

0
installs
▌
Optimizing Cpp Performance · vemodalen-x bundle
Analyze C/C++ for cache-access, ARM NEON SIMD vectorization, and multithreading optimization opportunities, with an embedded / image-processing lens. Use when a hot-path C/C++ routine needs a concrete optimization plan. Read-only analysis — proposes optimizations with code, does not edit.

0
installs
▌
Quantizing On Device Models · vemodalen-x bundle
Quantize a model for on-device (mobile / NPU) deployment and verify it still meets quality — choose fp16 vs int8 (dynamic-range PTQ, full-int8 PTQ with a representative dataset, or QAT), pick per-channel weights and input symmetry, decide which layers stay float, and gate on an accuracy-vs-latency budget. Use when a fp16 model is too slow or large on device, when planning an INT8 conversion, when choosing a calibration set, or when a quantized model's accuracy regresses. Verifies parity in decision space (argmax / IoU), not raw logits. Triggers: quantize model, int8 / PTQ / QAT, representative dataset, calibration set, per-channel quantization, accuracy drop after quant, fp16 vs int8, shrink model for mobile.

0
installs
▌
Breaking Down Prds · vemodalen-x bundle
Decompose a PRD / requirements document into a structured, governable breakdown — functional & non-functional requirements, acceptance criteria, work-breakdown by stage/agent, requirement→stage traceability, and a coverage check. Use at project kickoff or a major new feature, when the orchestrator must turn a PRD into actionable, traceable work before research/build begins. Owned by Framework 0 (orchestration).

0
installs
▌
Challenging Assumptions · vemodalen-x bundle
A deterministic thinking partner that challenges assumptions and applies mental models to sharpen decisions, solve problems, and think more clearly. Use this skill whenever a user says "help me think through X", "challenge my thinking", "what am I missing", "apply mental models to this", "play devil's advocate", "stress test this idea", "poke holes in my plan", "help me decide between X and Y", "what are the second-order effects", "I'm stuck on a decision", names any specific model (SWOT, first principles, inversion, pre-mortem, etc.), or asks for structured reasoning on any ambiguous, high-stakes, or complex problem. Also trigger when the user seems uncertain, is rationalizing, or is asking "am I thinking about this right?" Even casual phrases like "what do you think about..." on non-trivial topics should trigger this skill.

0
installs
▌
Converting Pytorch To Tflite · vemodalen-x bundle
Convert a trained PyTorch (or ONNX) model into a numerically faithful mobile TFLite — fp16 or int8-hybrid — and fold the camera colour transform into the first conv so the on-device model consumes the raw camera format (YUV/BGR) directly. Use when exporting a checkpoint to TFLite for on-device deployment, when a converted model's output drifts from the PyTorch reference, or when the runtime colour space differs from the training colour space. Covers reparameterize/deploy-before-trace, multi-stem colour folding, and a PyTorch-vs-TFLite parity gate. Triggers: export to tflite, pytorch/onnx to tflite, convert model for mobile, tflite output mismatch, YUV/BGR colour fold, fp16 vs int8 export.

0
installs
▌
Polishing Chinese Prose · vemodalen-x bundle
The canonical Chinese-prose authority — checkable rules for Chinese output, in two bands. 翻译腔 band (R14–R20: active voice, no vague modifiers, sentence-splitting, unambiguous pronouns, consistent terminology, 的/地/得, consistent persona) + 文牍腔 band (R33–R39: verbs over noun-stacks, no self-coined abbreviations, ≤1 arrow-chain per paragraph, ≤1 parenthetical per sentence, the read-aloud test, term-preservation with human prose, no self-coined concept-terms / metaphor-as-jargon) + an EN→zh term table. Use when authoring/reviewing Chinese deliverables, reviewing the README_zh mirror's fluency, or when an agent replies to the user in Chinese. Other skills cite this by name as the prose authority; it does not restate their rules.

0
installs
▌
Structuring Solution Docs · vemodalen-x bundle
Structure rules for a technical solution / design document (the research-solution framework's solution_document deliverable). Style-as-rules for document STRUCTURE — arc42 skeleton + MADR per-decision blocks — turned into checkable rules. Use when the research-solution agent authors the solution/design doc after the survey + recommendation. Owned by Framework 1 (research). Generic; project values live in the instance.

0
installs
▌
Evaluating Segmentation Models · vemodalen-x bundle
Evaluate a segmentation or image-matting model with the right metrics instead of eyeballing outputs. Use when accepting a trained seg/matting model, comparing two checkpoints, or judging whether a converted or quantized model still meets quality — computing IoU/mIoU and boundary-F for masks, and SAD/MSE/Grad/Conn in the trimap unknown band for matting, reported per-class and at edges. Distinguishes globally-similar-but-edge-wrong outputs (low mean error but high max_abs at thin structures). Triggers: evaluate segmentation, matting metrics, IoU / boundary F, SAD MSE Grad Conn, per-class accuracy, compare checkpoints, is this mask good, alpha matte quality.

0
installs
▌
Validating On Device Inference · vemodalen-x bundle
Run a converted model on a real device and accept it — push the package, run, collect results + logs, then judge numerical consistency (host↔device) first and performance second. Consistency = elementwise |Δ| ≤ caller tol + argmax agreement N/N; low-precision budget judged on softmax distance + decision equivalence (NOT raw logit); canary headroom = margin ÷ observed device deviation. Performance = warmup/timing rounds separated, latency reported as a distribution (not one shot) with the platform it was measured on, delegate on/off each re-verified, power as an annotated proxy. A weaker-than-target test platform extrapolates conservatively (pass = directional PASS; fail = inconclusive not dead; a thin pass margin must be discounted and flagged "target must be measured"). Use when a model has cleared host-side gates and must be signed off on the target hardware. Methodology checklist — emits PASS/FAIL report rows; all device/model/threshold values are read from the caller, none are baked in.

0
installs
▌
Governing Project Fleets · vemodalen-x bundle
Operate VEMO's private local fleet control plane to discover, register, assess, and preview-first onboard Git projects across one PC. Use when the user asks to govern all PC projects, scan local repositories, choose solo/team/regulated profiles, generate fleet readiness reports, or roll out VEMO without overwriting project-owned files.

0
installs
▌
Visualizing Governance · vemodalen-x bundle
Render the governance system's operation as diagrams — a mermaid flowchart of how frameworks/gates/agents run, a stateDiagram of the task seven-gate lifecycle, and an interactive markmap HTML overview for onboarding. Use when a README needs its governance diagram, when onboarding material is produced, or at release time to regenerate embedded SVGs. Owned by Framework 0 (orchestration).

0
installs
▌
Publishing Deliverables · vemodalen-x bundle
Publish a project deliverable (proposal / research report / experiment report / design doc) to the project's Feishu (Lark) wiki as a clean technical-report-styled document — tables and flowcharts for structured info. Use when a stage produces a deliverable that should be filed to the team Feishu wiki. Owned by Framework 0 (orchestration); style is constrained by ml-paper-writing + the local style spec.

0
installs
▌
Selecting Mobile Gpu Convolutions · vemodalen-x bundle
Decide between standard and separable (depthwise+pointwise) convolutions for mobile-GPU (OpenCL / TFLite delegate) deployment using three measured heuristics — first-frame time tracks OpenCL kernel count (not FLOPs), warmup gain tracks arithmetic intensity, steady-state time tracks FLOPs ÷ GPU utilization. Use at model-design time to narrow conv choices and on-device to confirm. Read-only advisory — heuristics from one anonymized field benchmark; verify on the target runtime. Reversal conditions (large channels / kernel fusion) apply.

0
installs
▌
Authoring Skills With Evals · vemodalen-x bundle
Author, evaluate, and improve skills with an eval-driven loop instead of by intuition. Use when creating a new skill, revising an existing one, judging whether a skill's description actually triggers, or tuning that description for better trigger accuracy. Runs a with-skill-vs-baseline behavioral eval with variance, a trigger-rate eval, and a train/test-split description optimizer via the repo's skill-creator harness, modeled on Anthropic's official skill-creator. Triggers include create a skill, write a SKILL.md, run a skill eval, benchmark a skill, description not triggering, improve a skill description, skill quality, eval-driven authoring.

0
installs
▌
Attending Group Mentions · vemodalen-x bundle
Staff a team group chat — when someone @-mentions the bot, read what they asked and reply to that need. Pull new messages by cursor (per-instance cursor file), filter for @bot mentions since last handled, triage the request (report / data / status / question / over-authority decision), serve what is serviceable (send a file or link as the bot, answer numbers ONLY by quoting the named ledger/report — never fabricate), reply to the asker with a post @-tag, in instance-policy Chinese (via polishing-chinese-prose). Decision-class requests escalate ("已转负责人" + @user), never auto-answered. Use to run group-chat attendance / answer @bot mentions on demand. Identity-decoupled — group id, bot/user identity, cursor + wordlist paths, open_id map are instance values.

0
installs

30 published skills