菜单
Bootstrap a new E2SA project or run. Creates projects/<project>/runs/<run_id>/ with the canonical skeleton (RESEARCH_PLAN.md, REPORT.md, run.yaml, notebooks/, data/, figures/). Always fires first when a new run begins; the human asking for analysis directly does not get to skip it.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Scaffold and register a new E2SA skill so the steps are never half-done. Use when the human or agent wants to create a new capability — "add a skill", "create a new skill", "scaffold a skill", "make this reusable as a skill", "distill X into a skill". Creates the SKILL.md with correct frontmatter, registers it in the AGENTS.md capability catalog (the manual registry), seeds a Changelog, and stops for human review before commit. Human-gated.
Improve an existing E2SA skill from accumulated evidence, human-gated. Use when a skill missed a step or needs updating, when a lesson/feedback in a skill's domain has piled up, or during a periodic skill review — "refine the X skill", "improve this skill", "the X skill should have caught Y", "review the skills". Gathers the signal (lessons, run-journal outcomes, feedback), proposes a concrete SKILL.md diff, and STOPS for human approval; it never self-applies. After approval it edits, appends a Changelog line, and commits.
Implement a new E2SA data-source adapter (the code that fetches a source and parses it into Observation records). Use when writing the adapter for a source that has already been onboarded/inspected — "build the adapter for X", "write the X adapter", "implement the fetch/parse for X", or the coding step after e2sa-add-data-source. Walks the BaseAdapter contract (list_available / fetch / parse_to_schema), the FetchResult single-file-vs-whole-package shape, the Observation + Provenance schema, canonical units + depth + serves, ADAPTER_REGISTRY registration, the fixture + test, and the parse gotchas already learned (sentinels, CRS tiers, obs_id uniqueness, schema drift, idempotency).
Implement a new E2SA data-center connector (the per-data-center access layer that owns auth + search + whole-package fetch under Option C). Use when wrapping a data center not yet connected — "add a connector for X", "wrap the PANGAEA/Earthdata/Zenodo API", or when migrating an existing self-fetching adapter onto a connector. Walks the BaseConnector contract, the probe-the-API-first step, the two search patterns (server-side filter vs enrich + bbox coverage-ratio), the two fetch shapes (BagIt-zip vs file-by-file), the reusable conventions (urllib + User-Agent, raw layout, token-free reads, on-disk fast-path, manual fallback), CONNECTOR_REGISTRY registration, and the mocked-urlopen + E2E_LIVE test split. Distilled from the arctic_data_center and ess_dive connectors.
Add a new data source (data center, archive, or dataset) into an E2SA project. Use when the human or agent has found a dataset to bring in, is inspecting a new archive, or says "add a data source", "add this dataset", "I found data at X", or after downloading data from a center not yet documented. Walks the full procedure: inventory check, download whole package, inspect (format/metadata-standard/CRS/license/citation), write the project-named source card, update the sources index and data-center registry, capture a finding, and credential setup. Always uses the latest dataset version and records a full citation.
Audit an E2SA adapter's data output against the REAL source and propose source-grounded corrections (applied only on human approval). Use after building or changing an adapter, after a parse, when staged data looks off, or for a periodic adapter review — "audit/validate the X adapter", "check the data X produces", "is above_stdm correct", "review the adapters for data-quality". Runs the e2sa/qc checks (serves-subset, value-range, subsurface-depth, citation-not-synthesized, self-describing) on a real parse, surfaces the value distributions, cross-checks variables/units/citation against the source, and PROPOSES a source-grounded fix for each finding (never invents) that is applied only on human approval (propose-first, like e2sa-refine-skill). NOT for run-level shippability (run skeleton / notebooks / secrets / a run_id) — that is `e2sa-validate`. Distilled from the 2026-06-23 above_stdm fabrications (reflection 20260623s).
| name | e2sa_start |
| description | Bootstrap a new E2SA project or run. Creates projects/<project>/runs/<run_id>/ with the canonical skeleton (RESEARCH_PLAN.md, REPORT.md, run.yaml, notebooks/, data/, figures/). Always fires first when a new run begins; the human asking for analysis directly does not get to skip it. |
| allowed-tools | ["Bash","Read","Write","Edit"] |
Bootstrap entry point for any new E2SA run. Mandatory first step. The workflow guard refuses to advance past S0 Intake until this skill has run.
/e2sa_start slash command.runs/<run_id>/ exists for the requested run.<project> and <run_id>). run_id names the analysis, not the data. It is lowercase snake_case describing the work and unique per run (alaska_thaw_db_eda, ice_content_baseline, blocked_cv_v2) — it is not a dataset_id slug. A run often spans several datasets, and you will want multiple runs on the same data (an EDA, a baseline, a validation), so name it by purpose with a verb/kind suffix (_eda, _baseline, _validation), never after a single dataset's slug.e2sa init <project> <run_id>. The CLI does all filesystem work; this skill does not duplicate logic.memory/knowledge/ and the cross-workspace index (handled by e2sa-intake, not by this skill, but flag the handoff).projects/<project>/runs/<run_id>/
├── run.yaml # manifest: project, run_id, created, author, status, model versions
├── RESEARCH_PLAN.md # declarative scope (agent + human author together)
├── REPORT.md # findings (agent authors after results land)
├── notebooks/ # analysis as Jupyter notebooks (with saved outputs)
├── data/ # run-specific intermediate data (gitignored)
└── figures/ # run-specific figures
Bootstrap is non-skippable. If a human prompt asks for analysis directly (skipping the skeleton), the orchestrator must call e2sa_start first, then proceed.
After scaffolding, if the new run's intended analysis depends on site data that is not present under projects/<project>/data/processed/sites/<site>/ and does not have a completed MANIFEST.md in projects/<project>/tasks/, fire e2sa-file-task once per missing site (or once per missing work-type for a multi-task site). Set related_run in the new <task_id>.md's frontmatter to this run's id. The decision rule for file vs inline-do vs ask lives in memory/knowledge/methods/20260521-when-to-file-a-data-prep-task.md. Do NOT execute the filed task in the same fire; surface to the human and wait.
Print the path of the new run subdirectory. Print which template files were seeded. Do NOT begin analysis in this skill; that is the next stage's job.
docs/design/09_skill_evolution.md); prior history is in git.run_id naming convention in step 1 (run_id = the analysis/purpose, lowercase snake_case with a verb/kind suffix; NOT a dataset_id slug). Prompted by keeping alaska_thaw_db_eda rather than renaming it to the dataset slug.