ai-workspace

Files (24)

• SKILL.md
• audit-rubric.md · references/
• creation-rubric.md · references/
• domain-knowledge-rubric.md · references/
• AGENTS.md · templates/
• accumulated-actions-process.md · templates/
• action-process.md · templates/
• api-guide.md · templates/
• daily-triage-process.md · templates/
• data-contract.md · templates/
• deck-outline.md · templates/
• distillation-process.md · templates/
• domain-knowledge-contract.md · templates/
• insight-process.md · templates/
• memory-process.md · templates/
• object-process.md · templates/
• report-artifact.md · templates/
• reporting-process.md · templates/
• runtime-process.md · templates/
• scenario-process.md · templates/
• source-process.md · templates/
• summary-process.md · templates/
• tension-process.md · templates/
• validation-flow.md · templates/

SKILL.md (first lines)

# AI Workspace

Create or review AI workspaces operated by agent hosts and humans.

An AI workspace makes behavior visible through files:

- `AGENTS.md` or host-equivalent root instructions as the workspace contract;
- `data/` for local evidence, dated snapshots, and layered durable knowledge
  state;
- `process/` for event handlers that create or update state;
- `my-work/` for dated working notes, reports, briefings, and generated
  deliverables when the workspace supports recurring analysis;
- optional `output/` for simple generated views when the repo already uses that
  convention or no dated work area is needed;
- optional `skills/` for host-discoverable reusable workflows;
- optional `scripts/` only for deterministic logic that host tools cannot do
  well;
- a simple validation flow that proves the main event path works.

The default execution model borrows AppRun's event/data/handler shape:

```txt
event + current state -> event handler -> new or updated state
```

Events can come from user requests, cron/timer triggers, scheduled reviews, file
or data changes, API/webhook/input updates, validation runs, audit runs, or
manual decisions. Classify the event, load the smallest relevant
event handler defined under `process/`, read the necessary current state, perform
documented effects, then create or update state in `data/`, render dated work
under `my-work/` or simple output under `output/`, or respond directly.

Prefer `AGENTS.md + data/ + process/ + my-work/` for domain workspaces that use
local records, recurring evidence, domain objects, or briefing outputs. Use
`output/` only for lighter workspaces or existing repo conventions.

Add skills only when host skill discovery, reuse, packaging, or a triggerable
entry point is useful.

## Mode

Classify the request first:

- Create: design, scaffold, patch, or improve a workspace.
- Review/audit: inspect and report gaps. Stay read-only.
- Validate: run or describe checks and say what passed.

Mixed request order:

1. Review current state.
2. Explain gaps.
3. Edit only the requested parts.

## Architecture Choice

Choose the smallest structure that preserves the workspace contract.

Root/process workspace:

```txt
AGENTS.md
data/
process/
my-work/
```

Use this when root instructions, durable data, event handlers, and generated
outputs are enough.

Evidence-backed knowledge workspace:

```txt
AGENTS.md
data/{partition}/{yyyy}/{mm}/{dd}/<object-type>/<id>/<object-id>-data.json
data/{partition}/{yyyy}/{mm}/{dd}/<object-type>/<id>/<object-id>-source.md
data/{partition}/{yyyy}/{mm}/{dd}/<object-type>/<id>/<object-id>-summary.md
data/{partition}/daily-triage/{yyyy}/{mm}/{dd}/accumulated-actions-{yyyy-mm-dd}.json
data/{partition}/daily-triage/{yyyy}/{mm}/{dd}/actions-{yyyy-mm-dd}.md
process/source.md
process/distillation.md
process/summary.md
process/memory.md
process/tension.md
process/insight.md
process/action.md
process/accumulated-actions.md
process/<scenario>.md
my-work/{yyyy}/{mm}/{dd}/
my-work/{yyyy}/{mm}/{dd}/output/
```

Omit `{partition}` only when the domain has no team, tenant, business unit,
source system, workspace, or other durable partition.

When daily distillation is expected, the dated bucket is the canonical write
path for each run. Do not use `current/` as the only record of daily judgment;
use it only as an optional latest view that points to or copies the dated
artifact.

Use this when the domain has local exports, record-like data, dated snapshots,
object-level evidence, recurring triage, scenario analysis, briefing decks, or
reviewable reports. This is the default when the user asks for a workspace
similar to `../crm-ai-workspace`.

API-backed workspace:

```txt
AGENTS.md
data/
process/api.yaml
process/api.md
process/<event-handler>.md
data/api-responses/
my-work/
```

Use this when event handlers call an external API as a documented effect.
Add `scripts/` only when `AGENTS.md` or `process/*.md` references them.

From yysun/awesome-agent-world, MIT.

More from yysun/awesome-agent-world