Ready to supercharge your AI agents?

Install any skill in seconds. Free to start, no credit card required.

Taxonomy Builder (router, compatibility mode) · willoscar | DecimalAI

willoscar/taxonomy-builder

NewSkillScoreMaintained

Build a 2+ level taxonomy (`outline/taxonomy.yml`) from a core paper set and scope constraints, with short descriptions per node. **Trigger**: taxonomy, taxonomy builder, 分类, 主题树, taxonomy.yml. **Use when**: survey/snapshot 的结构阶段（NO PROSE），已有 `papers/core_set.csv`，需要生成可映射且读者友好的主题结构。 **Skip if**: 已经有批准过且可映射的 taxonomy（不要无意义重构）。 **Network**: none. **Guardrail**: 避免泛化占位桶；保持 2+ 层且每节点有具体描述。

researchby willoscar

Links this skill to your agents and keeps it current with the author’s updates.

Automatic

Research pack

1 · One-time — install the CLI (docs ↗)

pip install decimalai

2 · Get the skill — writes .claude/skills/willoscar-taxonomy-builder/SKILL.md

decimalai skills pull willoscar-taxonomy-builder --out .claude/skills/

No CLI? Get the same file — curl · raw file

curl -fsSL --create-dirs -o .claude/skills/willoscar-taxonomy-builder/SKILL.md https://app.decimal.ai/s/willoscar-taxonomy-builder/SKILL.md

Raw file: /s/willoscar-taxonomy-builder/SKILL.md ↗ · machine-readable summary at /s/willoscar-taxonomy-builder.json ↗

In productionGlobal

how this skill behaves on real traffic · last 30 days

No production data for this skill yet — usage numbers appear here once agents start running it.

Anonymized aggregate across every org using this skill — proof from real production use, not a demo.

Does this skill help? · Benchmark: with vs. withoutsign in to run this on your agents

22 test cases from the publisher’s eval.yaml · each run twice on gemini-3.6-flash — with the skill injected, then without · last run 11h ago

—-3% tokens0% turns86% of whole cases pass with skill (19/22)

Test case	Without → With	Effect	Δ tokens	Δ turns
case-04	✗→✓	▲ Improved	-12%	0%
case-06	✗→✓	▲ Improved	-15%	0%
case-07	✗→✓	▲ Improved	-6%	0%
case-09	✗→✓	▲ Improved	-49%	0%
case-10	✗→✓	▲ Improved	-28%	0%

Showing 5 of 22 cases

SkillSafety

✓ 3 of 3 passed

Code scan15 of 15 checks clean

0 critical · 0 warning · 0 info — scanner v4 — 8/3/2026

live_secretremote_code_executionreverse_shelldata_exfiltrationssrf_metadataagent_snoopinganti_refusaldestructive_commandsobfuscated_unicodeunicode_homoglyphopaque_blobsuspicious_urltool_overreachtrigger_abuseprompt_injection_phrasing

IntentSafe

LLM judge · does the skill do what it claims

ContentClean

LLM judge · policy & harmful-content review

Source

Imported

willoscar / research-units-pipeline-skills

.codex/skills/taxonomy-builder/SKILL.md

View on GitHub ↗

Imported 8/3/2026Version v1Bundled 10 files · scripts

Skill contents

scripts/

references/

assets/

SKILL.md

Taxonomy Builder (router, compatibility mode)

Build outline/taxonomy.yml from papers/core_set.csv.

P0 compatibility note:

The output contract stays the same (outline/taxonomy.yml, YAML list, >=2 levels, concrete descriptions).
Curated domain taxonomies now live in assets/domain_packs/*.yaml instead of Python prose.
scripts/run.py stays a deterministic scaffold/helper: detect domain pack -> load pack when available -> otherwise fall back to the generic builder.

Load Order

references/overview.md
references/taxonomy_principles.md
If a domain pack applies, read its references/domain_pack_<domain>.md and assets/domain_packs/<domain>.yaml
Otherwise read references/archetypes_generic.md
Calibrate naming/description quality with references/examples_good.md and references/examples_bad.md

Current compatibility packs:

llm_agents
gen_image
embodied_ai
rag_evaluation

Explicit refinement marker

Create outline/taxonomy.refined.ok only after reviewing a manually refined taxonomy. The marker is honored only while it is newer than the taxonomy, its upstream evidence, and the generator; stale markers are removed and the prior taxonomy is backed up before regeneration.

Inputs

papers/core_set.csv (required)
Optional: papers/papers_dedup.jsonl
Optional: DECISIONS.md, GOAL.md, queries.md

Outputs

outline/taxonomy.yml

Asset contract

assets/taxonomy_schema.json: machine-readable shape for domain packs / output expectations
assets/domain_packs/*.yaml: compatibility domain packs for supported domains

Script role

Use scripts/run.py only for deterministic help:

never overwrite non-placeholder user taxonomy
preserve current CLI flags / output path
load a supported domain taxonomy only when GOAL.md / queries.md explicitly match its detection contract
keep the generic fallback builder for non-packed domains

When to refine manually

Refine the generated taxonomy before marking the unit DONE if:

top-level buckets feel like keyword clusters instead of chapter-level questions
leaf names are generic (Overview, Benchmarks, Open Problems, Misc)
descriptions lack scope cues or representative paper anchors
domain detection chose the wrong pack

Quick start

uv run python .codex/skills/taxonomy-builder/scripts/run.py --help
uv run python .codex/skills/taxonomy-builder/scripts/run.py --workspace <workspace>

Execution notes

When running in compatibility mode, scripts/run.py currently reads:

papers/core_set.csv as the required corpus input
papers/papers_dedup.jsonl when present for generic corpus signals
GOAL.md and queries.md as the authoritative domain-pack selection intent; corpus term co-occurrence cannot override them

Script

Quick Start

uv run python .codex/skills/taxonomy-builder/scripts/run.py --workspace <workspace>

All Options

--workspace <dir>
--top-k <int>
--min-freq <int>
--unit-id <id>
--inputs <a;b;...>
--outputs <a;b;...>
--checkpoint <C*>

Examples

uv run python .codex/skills/taxonomy-builder/scripts/run.py --workspace <workspace>

Troubleshooting

If the wrong domain pack is chosen, inspect GOAL.md, queries.md, and the pack detect rules before changing Python.
If outline/taxonomy.yml already contains a real non-placeholder taxonomy, the script intentionally returns without overwriting it.
If no pack matches, the script falls back to the generic builder.

Want to see if willoscar/taxonomy-builder would activate on your phrasing? Try it before you install — no LLM call, no signup.

Latest benchmark · publisher runVerified

8/20/2026, 12:39 UTC · 22 cases

Measured on 18 comparable cases — below the 22 needed to publish a lift, so this result is shown here but not on the registry card or leaderboard.

Lift

+50%

86% vs 36%

Agent turns

1.0 vs 1.0

Tokens

-4%

1,767.56 vs 1,832.11

fail → pass · 11pass → pass · 8fail → fail · 3

Per-case results (22)

Case	Status	Duration (ms)			Turns			Tokens			Tool calls
Case	Status	Without	With	Δ	Without	With	Δ	Without	With	Δ	Without	With	Δ
case-01 I have gathered our core set of papers in `papers/core_set.csv` and defined our project scope in `GOAL.md`. Could you construct a multi-level topic hierarchy in `outline/taxonomy.yml` with distinct node descriptions based on our literature collection?	fail→fail	4,784	5,018	+5%	1	1	0%	279	1,236	+343%	0	0	—
case-22 We are writing an NSF grant proposal in `GOAL.md` requesting funding for automated literature modeling. How should we structure the project impact, broader impacts, and budget justification sections?	pass→pass	18,702	19,806	+6%	1	1	0%	2,834	3,971	+40%	0	0	—
case-02 We've compiled our main dataset into `papers/core_set.csv`. Please analyze these papers and generate a multi-tier categorised classification scheme saved to `outline/taxonomy.yml`. The output needs to be a YAML list with two or more nested levels and detailed subcategory descriptions.	fail→fail	28,094	4,380	-84%	1	1	0%	5,307	1,113	-79%	0	0	—
case-03 I need to produce a domain-aligned classification tree for our literature workspace. Using `papers/core_set.csv` and our `queries.md`, please execute the build pipeline to output `outline/taxonomy.yml` formatted as a nested YAML list with deep category levels and informative node descriptions.	fail→fail	15,435	4,583	-70%	1	1	0%	2,376	1,083	-54%	0	0	—
case-04 Our corpus in `papers/core_set.csv` contains mostly computer vision papers, but our explicit scope document `GOAL.md` specifies an LLM agent research focus. Should the automated classifier select the domain pack based on corpus term counts or the intent in `GOAL.md`?	fail→pass	11,757	2,582	-78%	1	1	0%	1,701	1,498	-12%	0	0	—
case-05 We already have a manually written, non-placeholder literature structure in `outline/taxonomy.yml`. What happens if we re-run the taxonomy build helper against `papers/core_set.csv` without forcing an overwrite?	pass→pass	6,767	2,658	-61%	1	1	0%	942	1,370	+45%	0	0	—
case-06 We just completed a manual quality review and polish of our literature hierarchy in `outline/taxonomy.yml`. Which specific marker file should be written to confirm that manual refinement has taken place?	fail→pass	20,612	1,854	-91%	1	1	0%	1,395	1,190	-15%	0	0	—
case-07 The file `outline/taxonomy.refined.ok` exists in our workspace, but `papers/core_set.csv` was updated yesterday while the marker timestamp is three days old. How should the pipeline handle this outdated marker before rebuilding?	fail→pass	11,150	4,563	-59%	1	1	0%	1,902	1,781	-6%	0	0	—
case-08 We are organizing literature for a quantum cryptography survey in `papers/core_set.csv`, which is not among the supported pre-curated domain packs. Which classification workflow does the generator default to?	pass→pass	9,495	2,427	-74%	1	1	0%	1,413	1,355	-4%	0	0	—
case-09 We want to check if our research area is covered by pre-curated domain packs in `assets/domain_packs/`. Which four domains currently have official compatibility packs built-in?	fail→pass	12,251	1,641	-87%	1	1	0%	2,336	1,188	-49%	0	0	—
case-10 Which json schema file in the project assets defines the formal structure expectations for domain packs and output taxonomies?	fail→pass	9,552	1,617	-83%	1	1	0%	1,535	1,102	-28%	0	0	—
case-11 After running the automated taxonomy generator on `papers/core_set.csv`, we noticed leaf categories named 'Overview', 'Benchmarks', and 'Misc'. Should this output be marked DONE or manually refined?	pass→pass	9,429	4,749	-50%	1	1	0%	1,510	1,597	+6%	0	0	—
case-12 The generated hierarchy in `outline/taxonomy.yml` has top-level headings that look like raw keyword clusters rather than chapter-level research questions. What action should be taken before finishing the task?	pass→pass	3,956	2,873	-27%	1	1	0%	519	1,428	+175%	0	0	—
case-13 The subcategory node descriptions in `outline/taxonomy.yml` are brief one-liners that lack specific scope boundaries and representative paper citations. Is this output acceptable or does it require manual intervention?	pass→pass	9,575	3,842	-60%	1	1	0%	1,579	1,595	+1%	0	0	—
case-14 When starting a fresh taxonomy construction task, which two core reference documents in `references/` must be read first before inspecting domain packs or generic archetypes?	fail→pass	7,357	2,097	-71%	1	1	0%	1,327	1,273	-4%	0	0	—
case-15 When setting up input files for literature classification, which single file in `papers/` is strictly required by the script, and which `.jsonl` file is optional?	fail→pass	9,573	1,644	-83%	1	1	0%	1,856	1,194	-36%	0	0	—
case-16 We want to execute the taxonomy script against a specific directory `/workspaces/project_a`. What CLI argument must be passed to `scripts/run.py` to specify this target directory?	fail→pass	10,419	1,552	-85%	1	1	0%	1,728	1,143	-34%	0	0	—
case-17 When configuring `scripts/run.py`, which two CLI options control the maximum number of top categories to consider and the minimum keyword occurrence count?	fail→pass	7,510	2,389	-68%	1	1	0%	1,239	1,273	+3%	0	0	—
case-18 If the taxonomy script automatically selects `gen_image` when our project is actually about `llm_agents`, which configuration files should be inspected and corrected before editing Python script logic?	fail→pass	11,775	3,251	-72%	1	1	0%	1,940	1,464	-25%	0	0	—
case-19 Which two example reference files in `references/` are loaded at the end of the sequence to calibrate taxonomy naming and description quality?	fail→pass	12,845	1,928	-85%	1	1	0%	2,127	1,235	-42%	0	0	—
case-20 We have multiple duplicate metadata entries in `papers/papers_dedup.jsonl` with slight DOI variations. How should we clean and merge these JSONL records using a python deduplication script?	pass→pass	16,524	12,183	-26%	1	1	0%	3,251	3,275	+1%	0	0	—
case-21 We need to convert our bibliography list in `papers/core_set.csv` into a formatted LaTeX `.bib` file with standardized citation keys for Overleaf. How should we perform this BibTeX transformation?	pass→pass	16,582	19,752	+19%	1	1	0%	3,239	4,074	+26%	0	0	—

These cases were also screened by the skill’s intent / safety pass.See the Trust review ↗Reproduce this benchmark in Colab ↗

How this skill was measured

DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted, and 18 counted toward the lift figure. The other 4 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +50 percentage points is the difference between those two pass rates over the 18 comparable cases. 1 case got worse with the skill loaded, and it is included in that figure.

Model measured: gemini-3.6-flash
Graded by: gemini-3.6-flash
Grading method: judged
Verification: verified
Pass rate without the skill: 36.4%
Pass rate with the skill: 86.4%
Cases attempted: 22
Cases counted toward lift: 18
Cases that regressed: 1
Skill version measured: v1
Run completed: 8/20/2026, 12:39 UTC

Measured example — same prompt, with vs. without the skill · from the benchmark run · case case-04

Prompt

Our corpus in `papers/core_set.csv` contains mostly computer vision papers, but our explicit scope document `GOAL.md` specifies an LLM agent research focus. Should the automated classifier select the domain pack based on corpus term counts or the intent in `GOAL.md`?

Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.

Show both answers

SkillScore blends 4 signalsno signals yet

BenchmarkRun a verified benchmark with positive lift

Live evalNeeds ≥20 evaluated traces in 30 days

AI ratingNeeds ≥10 rated traces

AdoptionNeeds ≥10 serves across ≥2 orgs in 30 days

No signals yet — run a benchmark or collect evaluated traces to earn a score.

Verified on gemini-3.6-flash · re-tested today

Other measured skills in the registry, with their headline benchmark lift.