Agent Workflow Blueprint

Rendered documentation

Asset architecture migration notes

Version 3.0.1

Browse documentation Download Markdown source

Version 3.0.1

Asset IDs, source paths, source versions, and specialist-control IDs remain stable. Composition now separates review context from agent instructions.

  • Website and CLI copy output excludes the source title, applicability, metadata, dependencies, and references.
  • Goal, “Use this when”, “Do not use this when”, curated research citations, and implementation references remain visible in the website dialog outside the code region.
  • research-sources.json resolves every research_pattern_labels entry to an exact paper or standard plus the repository-level translation.
  • scripts/compose_assets.py --with-references is available for review artefacts that intentionally append citations and implementation sources.
  • Multi-asset packs preserve explicit Prompt, Skill, and Contract module boundaries; single-asset output starts directly with the executable body.

Version 3.0.0

All 100 stable asset IDs, paths, titles, governance profiles, and specialist-control IDs remain available. The change strengthens prompt sources and machine-readable quality metadata without removing skills, contracts, starter packs, or static-site behavior.

What changed

  • All 40 prompts now include explicit use and non-use boundaries, four or more required inputs, five to ten domain-specific instructions, decision gates, concrete evidence, failure-and-recovery pairs, rejection conditions, a domain output record, and a worked example.
  • The common handoff schema lives once in GOV-HANDOFF-01; each prompt contributes only its domain-specific record.
  • The manifest records risk_level, required_inputs, expected_outputs, and evaluation_cases for every asset.
  • Every prompt maps to at least one fictional evaluation case. The suite covers 40 cases and compares five conditions, including frozen v2.1 and v3 prompt inputs.
  • Validation enforces prompt source size, composed size, specialist-content ratio, input alignment, conditional gates, concrete evidence, recoveries, rejection conditions, output schema, example quality, evaluation coverage, and existing normalized-control rules.

Adoption

Existing v2.1 links and asset IDs continue to resolve. Recompose copied assets to receive the v3 prompt content:

python3 -S scripts/compose_assets.py --asset <asset-id>

Do not replace the frozen files in evals/baselines/v2.1-composed/; they are comparison inputs. Deterministic v3 validation does not establish behavioural improvement. Use the recorded-output process in docs/engineering/evaluation-methodology.md for that evidence.

Version 2.1.0

All 100 stable asset IDs, paths, and titles remain available. Source assets are now concise specialist modules rather than copies of shared policy.

What changed

  • Shared governance has stable GOV-* identifiers in GOVERNANCE-KERNEL.md.
  • Specialist instructions used by multiple assets have one authoritative SPC-* entry in SPECIALIST-CONTROLS.md.
  • Each manifest entry declares its governance profile and ordered specialist-control references.
  • compose_assets.py deterministically renders each selected module, the applicable kernel rules, and each referenced specialist control once in copy-ready order.
  • The website uses one dialog containing a short introduction, the full standalone prompt, skill, or contract as literal Markdown code, a compact copy control, and linked references.
  • Every source module retains at least three asset-specific instructions, procedure steps, or hard gates plus a distinct evidence section.
  • Validation rejects hollow registry delegation, duplicate specialist bodies or evidence sections, repeated inline specialist controls, near-duplicate controls at the configured threshold, orphaned controls, source/manifest drift, unresolved composition, and any unclassified repeated prose block.

Adoption

Use the compact copy control beside the website’s prompt/skill/contract code or run:

python3 -S scripts/compose_assets.py --asset <asset-id>

Repeat --asset to compose a prompt, skill, and contract together. Source modules remain readable and stable, but require their declared kernel and registry dependencies when used outside the repository.

Version 2.0.0

The 100 asset paths and titles remain stable. No prompt, skill, or contract identifier was removed.

What changed

  • Shared scope, evidence, uncertainty, traceability, failure-disclosure, escalation, and handoff rules moved to GOVERNANCE-KERNEL.md.
  • Each asset now concentrates on its specialist role, mission or objective, task-specific procedure, evidence, and rejection or failure conditions.
  • assets.json now records a stable ID, semantic version, purpose, intended model context, dependencies, and evidence expectations.
  • asset.schema.json defines the manifest entry contract.
  • Validation is quality-based: metadata integrity, stable paths, dependencies, required specialist sections, evidence language, rejection language, and duplicate-block limits replace the previous twenty-numbered-point size gate.

Adoption

Existing links continue to work. Version 2.0 consumers should migrate to complete v2.1 composition so shared specialist controls are not omitted.

Deprecated patterns

  • Repeating the full research stack and generic output contract in every asset is deprecated.
  • Unsupported certainty wording is deprecated; use research-informed or source-mapped wording.
  • Asset quality is no longer inferred from line count or numbered-rule count.

No asset is currently renamed or removed.