CogSecSkills - Library of 100 analytic skills for cognitive secur

项目说明

CogSecSkills

![DOI]()
![CI]()

Agentic tool-use skills for Cogniti and analytic tradecraft — a
library of 100 analytic skills (Structured Analytic Techniques, cognitive-
security defenses, OSINT integrity, counterintelligence, and critical review),
each authored once as a harness-neutral skill with default adapters for
Claude Code, Codex, and Hermes, and the same structural contract for any
additional configured harness.

Install

To connect an agent harness, point it at skills/<group>/<slug>/SKILL.md, run
workflow.md, and bind tools through the matching harness/<name>.md adapter.
The full command reference is in Usage; harness details are in
docs/harness-installation.md and
docs/harness-cookbook.md.

Read it / cite it

Manuscript: CogSecSkills.pdf and

HTML. The manuscript's
release manifest records its ,
en, and local gate results. The sources are v1.8.0 dated 2026-10-01.
Local acceptance and the
publication receipt record the rendered
artifacts, exact release source, hosted CI, and archi.
The prior v1.7.0 review preserves its own e
and limits; use each release's receipt for later artifact identities.

Software archive: concept DOI

10.5281/zenodo.21513316.
Published v1.8.0 has DOI
10.5281/zenodo.23090954 and a
GitHub release
with the manuscript, web bundle, source archive, and distributions.
The separate historical v1.0.0 manuscript archive has concept
10.5281/zenodo.20804585,
10.5281/zenodo.20804586.
Citation metadata is in CITATION.cff. The tag and retained
manuscript carry the software concept DOI; the DOI was assigned
after publication and is recorded by a separate citation follow-up on main.
Local source gates and li remain separate e.

New here? QUICKSTART.md · Docs map:

docs/README.md

The library is strictly defensi, and accountable. Its
educational upstream is the AGEINT curriculum. These
skills recognize, assess, and defend against cogniti — they do not
author manipulation.

The three coherent artifacts

| Artifact | Role | Where |
|----------|------|-------|
| Teach | the concepts (the *why*) | docs/ageint/ |
| Plan | catalogue of all 100 skill areas (the *what*) | registry/skills.yaml |
| Build | canonical definitions and rendered multiharness skill implementations (the *how*) | definitions/ and skills/ |

python -m cogsecskills keeps Plan and Build coherent: every on-disk
skill must be catalogued, and every implemented catalogue entry must exist.

Skill groups (workflow subfolders)

| Group | Focus | Count |
|-------|-------|-------|
| sat | Structured Analytic Techniques (Heuer & Pherson) | 34 |
| cogniti | Defending perception, reasoning, decision-making | 24 |
| critical_review | Ad + constructi (incl. project critical review) | 12 |
| osint_integrity | Open-source collection with pro discipline | 10 |
| counterintelligence | Denial/deception detection, process hardening | 8 |
| information_en | Narratives, influence ops, coordinated behavior | 7 |
| research_methods | Synthesis, e grading, calibrated estimation | 5 |

All 100 areas are fully implemented as conforming, multiharness skill folders —
each with a real, technique-accurate procedure, a fitting tool plan, and adapters
that bind every declared verb under the configured harness set. The default
configured set is Claude Code, Codex, and Hermes.
python -m cogsecskills report shows the li (implemented: 100).

All 100 skills now ha canonical definitions under definitions/.
python -m cogsecskills definitions --write renders the full skills/ tree
deterministically from that source layer, and definitions --check proves the
definitions and rendered files have not drifted.
python -m cogsecskills scenarios --check adds a deterministic defensive
readiness gate o safe-use and unsafe-redirect fixtures; it checks
routing, local skill contracts, expected response-shape metadata, and re
expected-answer fixtures without calling external model runtimes.
python -m cogsecskills examples --write generates one worked defensive
example per skill from examples/skill-worked-examples.yaml.
python -m cogsecskills evals --write generates offline local output-review
fixtures and reports from scenario expected answers; it does not call live
models.
python -m cogsecskills dashboard --write generates a 100-skill quality
dashboard for na and drift review across scenarios, offline evals,
worked examples, quality capsules, harnesses, references, and source paths,
with Markdown, static HTML, and JSON views.
python -m cogsecskills release-metadata --write generates the local release
claim matrix and metadata snapshot without publishing, tagging, or archiving.
Exact git re, branch, and dirty-state are observed at command
runtime rather than embedded in committed generated files.

Anatomy of a skill

A skill declares its capabilities as closed-set tool verbs (read, search,
write, exec, reason, web, delegate, ask). Each harness adapter binds
those verbs to concrete tools in a Markdown binding table. The checks
that every declared verb is present in every adapter, so "multiharness" is a
property the test suite *proves*, not a hope.

The persistent source of truth for authored skill substance is
definitions/<group>/<slug>.yaml; skill.yaml and the companion Markdown files
are the generated, harness-facing build outputs.

Usage

(Run from the project root. uv run selects the en created by
uv sync; it does not require shell acti.)

To connect a harness, point it at skills/<group>/<slug>/SKILL.md, use
workflow.md for the neutral procedure, and bind tools through the matching
harness/<name>.md adapter. The default configured set is claude, codex,
and hermes; add more harnesses in cogsecskills.yaml and regenerate adapters
with python -m cogsecskills definitions --write.

For bounded examples, see
examples/harness-smoke-transcripts.md,
examples/group-worked-examples.md, and
docs/skill-worked-examples.md.
For claim discipline, see docs/claim-boundaries.md.
For the generated quality dashboard, see
docs/quality-dashboard.md and
docs/quality-dashboard.html. style rules
for generated figures, the co, dashboard pages, and manuscript tables
live in DESIGN.md.

The 8 reference exemplars

(All 100 skills are implemented; these 8 are the hand-authored references.)

| Skill | Group |
|-------|-------|
| sat.analysis_of_competing_hypotheses | Structured Analytic Techniques |
| sat.key_assumptions_check | Structured Analytic Techniques |
| sat.de | Structured Analytic Techniques |
| cogniti.narrati | Cogniti |
| cogniti.source_credibility_e | Cogniti |
| critical_review.project_critical_review | Critical Review & Assurance |
| osint_integrity.claim_pro | OSINT & Source Integrity |
| research_methods.structured_literature_synthesis | Research & Synthesis Methods |

Testing

Tests live in per-concern packages under tests/ (core/, authoring/,
artifacts/, quality/, contract/, conformance/). Run the whole suite —
the same command CI runs:

The li test (tests/conformance/test_skill_library_conformance.py)
runs against the real skills/ tree: adding a malformed skill, or an
implemented registry entry with no build, fails the suite.

Pro

Canonical public repository: <

Local docs/manuscript/render may use the sibling docxology template, but
the deli is the skills system: registry, skills, AGEINT docs,
runner, tests, generated manuscript supplements, and figures that describe
those source surfaces.

Software release chain: concept DOI

10.5281/zenodo.21513316.
v1.8.0 archives the immutable release
source; all 941 file payloads match its Git commit. The earlier
software record contains a source ZIP
linked to the GitHub v1.7.0 tree, although its metadata says 1.0.0.

Historical manuscript archive: v1.0.0 on Zenodo — concept DOI

10.5281/zenodo.20804585 (all
manuscript ), DOI
10.5281/zenodo.20804586.
The historical v1.0.0 GitHub release
includes its manuscript PDF. Inspect each
GitHub release for its
actual assets; a tag alone does not establish a manuscript deposit.

Educational upstream: AGEINT

(concept DOI 10.5281/zenodo.20732274).

The skill catalogue draws on Heuer & Pherson, *Structured Analytic Techniques

for Intelligence Analysis*, and the cogniti literature cited per-primer.

See ISA.md for the project's ideal-state articulation and acceptance
criteria.

项目信息

开发语言：Python

Star 数：7

技术要点

该项目属于防护思路方向的开源资源，代码结构清晰，适合安全研究人员学习参考。项目围绕Library of 100 analytic skills for cognitive security and tradecraft — structured analytic techniques, OSINT integrity, counterintelligence, and information-environment analysis, each with a harness-neutral spec and adapters for Claude Code, Codex, and Hermes. Python 3.10+/uv CLI, archived on Zenodo.展开，提供了可运行的实现，对理解相关安全机制、工具原理与攻防技术具有较高的研究价值。

从技术实现角度看，此类项目的核心在于对底层原理的深入理解与工程化落地。开发者在研究过程中可以结合官方文档与社区讨论，梳理项目的初始化流程、核心模块职责与数据流转，逐步掌握其中的技术细节与设计思路。

使用与学习建议

1. 下载源码压缩包，解压到本地环境。

2. 根据项目类型准备对应的运行环境与依赖。

3. 按照 README 中的步骤完成配置与部署。

4. 在隔离的测试环境中运行验证，避免影响生产系统。

免责声明

本资源来源于开源社区，仅供安全研究、技术学习与合法授权范围内的攻防演练使用，请勿用于任何非法或侵害他人权益的用途。使用本项目产生的一切后果由使用者自行承担。