Context Generate
Generates the three context artifacts a project consumes:
- Skills - router-based:
SKILL.md router + atomic testable actions, evaluations declared before implementation.
- Agents - single-file agent definitions following the framework's agent template.
- Rules - framework rule files that govern editor/agent behavior.
Skill-generation actions
| # |
Action |
Role |
Input |
| 01 |
capture-intent |
Clarify output + decide generate vs modify |
user intent |
| 02 |
design-evals |
Write minimal scenarios.json |
expected output |
| 03 |
decompose-actions |
List actions + their tests |
evals + output |
| 04 |
draft-skill |
Write SKILL.md router |
decomposed actions |
| 05 |
write-actions |
Write each action file |
validated SKILL.md |
| 06 |
validate |
Spawn one agent per action, run its ## Test, aggregate into a pass/fail report |
complete skill |
Files: actions/skills/01-capture-intent.md … actions/skills/06-validate.md.
Other generation actions
actions/agents/01-generate-agent.md - generate a project agent file from assets/agents/agent-template.md.actions/rules/01-generate-rules.md - generate framework rules from assets/rules/rule-template.md.
Default skill flow
01 → 02 → 03 → 04 → 05 → 06. After each action, run its ## Test before moving to the next. Action 02 self-skips when 01 outputs invocation_mode = manual.
Modify flow
01 (detects modify) → 03 (re-decompose if structure changes) → 05 (targeted edit) → 06 (re-validate).
Runtime tracking
Materialize the flow as a task list at skill entry; a task closes only when its ## Test passes.
Rules
- R1 - SKILL.md is a pure router: description + action table + transversal rules. Zero business logic.
- R2 - One skill = one domain (tool OR activity). Tool → singular noun (
slack); activity → action verb (review). See references/naming-conventions.md.
- R3 - References one-level deep. Never chain reference → reference.
- R4 - SKILL.md ≤ 500 lines. If exceeded, split into references.
- R5 - Description must include: what, explicit triggers, "Do NOT use for..." clause.
- R6 - Zero duplication. Templates live in
assets/; actions point to them via @<path>.
- R7 -
references/ = documents to READ (conventions, cheatsheets). assets/ = files to COPY or INJECT (templates, ID tables).
- R8 - Every action has a
## Test: one sentence describing how to verify its intent - a command to run, a concrete check on the produced artifact, or an observable side-effect (API/MCP/state).
- R9 - Auto-trigger skills (
disable-model-invocation: false, default) ship evals/scenarios.json = JSON array of at least 3 {prompt, expect_action}. Manual-only skills skip.
- R10 - Generated skills are written in English only (frontmatter, body, actions, references, assets). Holds regardless of conversation language.
References
references/naming-conventions.md - tool vs activity naming, hard constraintsreferences/skill-structure.md - skill anatomyreferences/agents-coordination.md - multi-agent coordination patternsreferences/rule-structure.md - rule file anatomyreferences/rule-writing.md - rule authoring conventionsreferences/ide-mapping.md - syntax and file location reference for AI files (agents, commands, rules, skills)
Assets (templates to copy)
assets/skills/skill-template.md - SKILL.md skeletonassets/skills/action-template.md - action file skeletonassets/skills/evals-template.md - scenarios.json minimal schemaassets/agents/agent-template.md - agent file skeletonassets/rules/rule-template.md - rule file skeleton
1---2name: 03-context-generate-23description: Context Generate4---56# Context Generate78Generates the three context artifacts a project consumes:910- **Skills** - router-based: `SKILL.md` router + atomic testable actions, evaluations declared before implementation.11- **Agents** - single-file agent definitions following the framework's agent template.12- **Rules** - framework rule files that govern editor/agent behavior.1314## Skill-generation actions1516| # | Action | Role | Input |17| --- | ------------------- | --------------------------------------------- | ------------------ |18| 01 | `capture-intent` | Clarify output + decide generate vs modify | user intent |19| 02 | `design-evals` | Write minimal `scenarios.json` | expected output |20| 03 | `decompose-actions` | List actions + their tests | evals + output |21| 04 | `draft-skill` | Write SKILL.md router | decomposed actions |22| 05 | `write-actions` | Write each action file | validated SKILL.md |23| 06 | `validate` | Spawn one agent per action, run its `## Test`, aggregate into a pass/fail report | complete skill |2425Files: `actions/skills/01-capture-intent.md` … `actions/skills/06-validate.md`.2627## Other generation actions2829- `actions/agents/01-generate-agent.md` - generate a project agent file from `assets/agents/agent-template.md`.30- `actions/rules/01-generate-rules.md` - generate framework rules from `assets/rules/rule-template.md`.3132## Default skill flow3334`01 → 02 → 03 → 04 → 05 → 06`. After each action, run its `## Test` before moving to the next. Action 02 self-skips when `01` outputs `invocation_mode = manual`.3536## Modify flow3738`01` (detects modify) → `03` (re-decompose if structure changes) → `05` (targeted edit) → `06` (re-validate).3940## Runtime tracking4142Materialize the flow as a task list at skill entry; a task closes only when its `## Test` passes.4344## Rules4546- **R1** - SKILL.md is a pure router: description + action table + transversal rules. Zero business logic.47- **R2** - One skill = one domain (tool OR activity). Tool → singular noun (`slack`); activity → action verb (`review`). See `references/naming-conventions.md`.48- **R3** - References one-level deep. Never chain reference → reference.49- **R4** - SKILL.md ≤ 500 lines. If exceeded, split into references.50- **R5** - Description must include: what, explicit triggers, "Do NOT use for..." clause.51- **R6** - Zero duplication. Templates live in `assets/`; actions point to them via `@<path>`.52- **R7** - `references/` = documents to READ (conventions, cheatsheets). `assets/` = files to COPY or INJECT (templates, ID tables).53- **R8** - Every action has a `## Test`: one sentence describing how to verify its intent - a command to run, a concrete check on the produced artifact, or an observable side-effect (API/MCP/state).54- **R9** - Auto-trigger skills (`disable-model-invocation: false`, default) ship `evals/scenarios.json` = JSON array of at least 3 `{prompt, expect_action}`. Manual-only skills skip.55- **R10** - Generated skills are written in **English only** (frontmatter, body, actions, references, assets). Holds regardless of conversation language.5657## References5859- `references/naming-conventions.md` - tool vs activity naming, hard constraints60- `references/skill-structure.md` - skill anatomy61- `references/agents-coordination.md` - multi-agent coordination patterns62- `references/rule-structure.md` - rule file anatomy63- `references/rule-writing.md` - rule authoring conventions64- `references/ide-mapping.md` - syntax and file location reference for AI files (agents, commands, rules, skills)6566## Assets (templates to copy)6768- `assets/skills/skill-template.md` - SKILL.md skeleton69- `assets/skills/action-template.md` - action file skeleton70- `assets/skills/evals-template.md` - `scenarios.json` minimal schema71- `assets/agents/agent-template.md` - agent file skeleton72- `assets/rules/rule-template.md` - rule file skeleton
ai-driven-dev