Skip to content

TIKAZI/TIKAZ-AI-Skills / suites/context-economy

context-economy

Flagged

Automatically prepare fidelity-first, bounded context when users paste text, attach files or folders, provide webpages or PDFs for Markdown conversion, continue long conversations, or supply documents containing images and complex tables. Route each input through Text, Hybrid, or Source; reduce exact or formatting-only repetition; protect facts and source anchors; queue only useful visual evidence; preserve uncertain originals; and benchmark efficiency and fidelity separately.

FileKindSize
MANIFEST.inother353 B
README.mdmarkdown17.1 KB
README.zh-CN.mdmarkdown8.6 KB
SKILL.mdskill md5.5 KB
VERSIONother10 B
__init__.pyscript118 B
__main__.pyscript77 B
adapters/defuddle/extract.mjsscript694 B
adapters/defuddle/package-lock.jsonmanifest12.8 KB
adapters/defuddle/package.jsonmanifest172 B
agents/openai.yamlconfig420 B
assets/conversation-state.template.mdmarkdown715 B
assets/workflow.svgother5.7 KB
benchmarks/fixtures/chinese.mdmarkdown168 B
benchmarks/fixtures/code.pyscript127 B
benchmarks/fixtures/conversation.mdmarkdown176 B
benchmarks/fixtures/data.jsonconfig84 B
benchmarks/fixtures/efficiency-long.mdmarkdown3.5 KB
benchmarks/fixtures/english.mdmarkdown167 B
benchmarks/fixtures/long-form.mdmarkdown439 B
benchmarks/fixtures/multimodal-decoration.mdmarkdown150 B
benchmarks/fixtures/multimodal-hybrid.mdmarkdown220 B
benchmarks/fixtures/multimodal-plain-image.mdmarkdown112 B
benchmarks/fixtures/multimodal-simple-table.mdmarkdown75 B
benchmarks/fixtures/multimodal-table.mdmarkdown257 B
benchmarks/fixtures/multimodal-text.mdmarkdown184 B
benchmarks/fixtures/multimodal-two-charts.mdmarkdown101 B
benchmarks/fixtures/prompt-repeat-a.txtmarkdown140 B
benchmarks/fixtures/prompt-repeat-b.txtmarkdown148 B
benchmarks/fixtures/prompt-structural-a.txtmarkdown238 B
benchmarks/fixtures/prompt-structural-b.txtmarkdown220 B
benchmarks/fixtures/service.logother155 B
benchmarks/fixtures/source-scan.pdfbinary193 B
benchmarks/fixtures/web-article-images.htmlother334 B
benchmarks/fixtures/web-article.htmlother389 B
benchmarks/fixtures/web-empty.htmlother135 B
benchmarks/manifest.jsonconfig11.9 KB
benchmarks/pdf/fixtures/illustrated-report.pdfbinary1.9 KB
benchmarks/pdf/fixtures/table-report.pdfbinary1.9 KB
benchmarks/pdf/fixtures/text-report.pdfbinary2.2 KB
benchmarks/pdf/generate_fixtures.pyscript3.1 KB
benchmarks/pdf/ground-truth.jsonconfig1.0 KB
benchmarks/pdf/results/README.mdmarkdown1.4 KB
benchmarks/pdf/results/converted/illustrated-report.mdmarkdown112 B
benchmarks/pdf/results/converted/table-report.mdmarkdown246 B
benchmarks/pdf/results/converted/text-report.mdmarkdown201 B
benchmarks/pdf/results/metrics.jsonconfig1.4 KB
benchmarks/pdf/run_fidelity.pyscript5.5 KB
benchmarks/results/README.mdmarkdown2.2 KB
benchmarks/results/cases.jsonconfig30.2 KB
benchmarks/results/metrics.jsonconfig4.1 KB
benchmarks/results/summary.jsonconfig1.1 KB
benchmarks/retrieval-challenge-v1.jsonconfig9.1 KB
benchmarks/retrieval-manifest.jsonconfig14.5 KB
benchmarks/retrieval-results/challenge-v1/cases.jsonconfig14.1 KB
benchmarks/retrieval-results/challenge-v1/evidence-card.mdmarkdown536 B
benchmarks/retrieval-results/challenge-v1/metrics.jsonconfig494 B
benchmarks/retrieval-results/challenge-v1/summary.jsonconfig1.1 KB
benchmarks/retrieval-results/development-v1/cases.jsonconfig19.6 KB
benchmarks/retrieval-results/development-v1/evidence-card.mdmarkdown603 B
benchmarks/retrieval-results/development-v1/metrics.jsonconfig442 B
benchmarks/retrieval-results/development-v1/summary.jsonconfig1.1 KB
pyproject.tomlmanifest2.1 KB
references/benchmark-method.mdmarkdown2.9 KB
references/context-health.mdmarkdown480 B
references/conversion-adapters.mdmarkdown2.5 KB
references/output-contract.mdmarkdown2.2 KB
references/routing.mdmarkdown1.4 KB
references/threat-model.mdmarkdown5.1 KB
scripts/__init__.pyscript45 B
scripts/tikaz_context.pyscript94.3 KB

references/benchmark-method.md

View on GitHub
# Benchmark method

Use the versioned public manifest in `../benchmarks/manifest.json`. Each case declares synthetic source files, a task, a complete-output budget, protected facts, and expected source anchors. Interpret profiles separately: `correctness` tests small-input behavior; `efficiency` tests selection from longer sources.

Report four independent measures: complete-pack budget compliance, estimated token savings, literal protected-fact recall, and expected-anchor correctness. Preserve raw per-case failures in `cases.json`; never hide a failed case in the aggregate.

For retrieval claims, use committed labeled manifests and report Recall@K, Precision@K, MRR, nDCG@K, dynamic selected-evidence recall/precision, facet coverage, selection count, and estimated selected-text reduction separately. Publish the development manifest and a separately frozen challenge manifest; never present either as general search accuracy. A reduction number without its matching selected-evidence recall is incomplete.

The estimator is deterministic but is not provider billing telemetry. Cases with no declared protected facts are reported as vacuously complete and must not be counted as evidence that all facts were preserved. The benchmark does not measure answer quality or semantic equivalence. Add externally scored answer quality only as a separate field with its evaluator and rubric disclosed.

The bundled efficiency cases reuse one synthetic source across six tasks. Report that limitation whenever publishing their aggregate. Do not compare the percentage directly with command-output filters, model-based compressors, or proxy systems unless all projects run the same corpus, task, tokenizer, budget unit, and downstream evaluation.

For multimodal claims, add independent files with human-labeled informative/decorative visuals, duplicate targets, simple/complex tables, required source anchors, and blind downstream questions. Hold the model, image detail, page rendering, prompt, and token telemetry constant. Report routing accuracy, informative-visual recall, decorative skip precision, table-cell fidelity, answer correctness, latency, and input cost separately.

Report prompt preparation by mode. `exact` is the conservative baseline. `structural` may normalize formatting for duplicate detection but must keep the first original wording. Include no-duplicate controls so a mode is not rewarded merely for receiving easy repeated inputs. Keep semantic compression disabled or separately labeled until a downstream equivalence rubric passes.

For PDF conversion, generate fixtures from declared ground truth, render and visually inspect the fixtures, then convert page by page. Score required text, numeric facts, table cells, and page anchors separately. Record the adapter and failed capability probes. Literal extraction must never be presented as layout fidelity, diagram understanding, OCR accuracy, or real-world corpus coverage.
slack-gif-creatoranthropics/skillsKnowledge and utilities for creating animated GIFs optimized for Slack. Provides constraints, validation tools, and animation concepts. Use when users request animated GIFs for Slack like "make me a GIF of X doing Y for Slack."Stars: 198kUpdated Claude CodeCodexVerifiedInstall
ponytailDietrichGebert/ponytailLazy senior dev mode for any coding task: the smallest change that fully solves it, and a reply a busy human understands in one read.Stars: 196kUpdated Claude CodeCodexVerifiedInstall
ponytail-debtDietrichGebert/ponytailHarvest every shortcut: comment into one debt ledger, so deferrals get tracked instead of forgotten. One-shot report.Stars: 196kUpdated Claude CodeCodexVerifiedInstall
ponytail-debtDietrichGebert/ponytailList every `shortcut:` comment (and older `ponytail:` ones) as a debt ledger. One-shot report, changes nothing. Use for "ponytail debt", "what did ponytail defer", "list the shortcuts", /ponytail-debt.Stars: 196kUpdated Claude CodeCodexVerifiedInstall