templates

Imported from yysun/awesome-agent-world (skills/ai-workspace/templates/AGENTS.md). Install upstream with npx skills add yysun/awesome-agent-world --skill templates. Copyright stays with the author.

Agent Instructions

Remove optional sections that do not apply. Use the same human language as the user's workspace request. Auto-detect vocabulary for category, object type, and layer names.

Purpose

This workspace is for:

• ;
• .

It must not:

• ;
• .

Host Assumptions

The agent host can:

• read workspace files;
• write workspace files when requested;
• create parent folders;
• fetch web content when needed.
• run documented local scripts when a process handler requires deterministic export, normalization, validation, or artifact generation.

Operating Rules

This is an event-driven workspace:

event + current local state -> selected process handler -> new or updated state, output, or direct response

• Keep always-on behavior in this file.
• Use process/ for workspace-level workflow contracts.
• Do not reference folders that do not exist.
• Store durable outputs in documented locations.
• Keep local actions local unless a process handler and the user explicitly approve an external write.
• Say what was inspected, changed, and verified.

File Map

List only folders that exist in this workspace. Delete this section if no workspace folders are created.

• <folder>/: .

Recommended evidence-backed workspace map when local records or recurring analysis are part of the request:

• data/: durable local evidence, generated source snapshots, summaries, and optional deterministic action-queue state.
• process/: source, distillation, summary, section, scenario, triage, and reporting handlers.
• my-work/: dated work products, briefings, reports, analyses, and final deliverables.
• my-work/{yyyy}/{mm}/{dd}/output/: final deck, PDF, PPTX, or other generated outputs for that dated run.
• my-work/{yyyy}/{mm}/{dd}/output/scratch/: previews and intermediate render artifacts.
• scripts/: deterministic helpers only when referenced by this file or a process/*.md handler.

Domain Knowledge

Delete this section if domain setup is skipped.

Domain setup:

• <domain name | skipped>

If a domain is supplied, use these semantic layers:

• source: factual evidence snapshot with provenance and unknowns.
• memory: durable object knowledge.
• tension: unresolved pressure, risk, or opportunity.
• insight: current interpretation and consequence.
• action: local recommended next moves.

Localize layer filenames and section headings to the user's language. Keep the semantic meaning above.

For evidence-backed workspaces, default to the combined artifact chain:

source.md -> summary.md

summary.md contains Memory, optional Tensions, optional Insight, optional Proposed Actions, Confidence, and Review Notes. Proposed actions stay inside summary.md by default. Do not create standalone action.md unless the user explicitly asks for separate action artifacts or this workspace already uses that convention.

If proposed actions feed queue state, use process/accumulated-actions.md to build deterministic snapshots and readable action reports from summary.md. Daily triage reads those accumulated snapshots and action reports.

Core process files:

• process/source.md
• process/distillation.md
• process/summary.md
• process/<localized-memory-layer>.md
• process/<localized-tension-layer>.md
• process/<localized-insight-layer>.md
• process/<localized-action-layer>.md

Actions are local recommendations unless the user approves an external write. Semantic distillation from source.md to summary.md is agent-authored. Scripts may refresh evidence, generate factual sources, rebuild indexes, audit and split large distillation batches, validate contracts, build accumulated-action state, or post through explicit write gates; scripts must not draft, template, migrate, or bulk-transform judgment inside summary.md.

Large distillation runs must use a fixed manifest under my-work/{yyyy}/{mm}/{dd}/. Use deterministic scripts to audit targets, split the manifest into disjoint batches, and load assigned paths for workers. Workers may write only their assigned sibling summary.md files. The parent agent owns final audit, validation, index rebuilds, and accumulated-action rebuilds.

Artifact Paths

Delete this section if no data/ contract is created.

Dated evidence paths, when local evidence changes over time or daily distillation is expected:

data/{partition}/{yyyy}/{mm}/{dd}/<localized-object-type>/<object-id>/<singular-object-type>-<object-id>-data.json
data/{partition}/{yyyy}/{mm}/{dd}/<localized-object-type>/<object-id>/<singular-object-type>-<object-id>-source.md
data/{partition}/{yyyy}/{mm}/{dd}/<localized-object-type>/<object-id>/<singular-object-type>-<object-id>-summary.md

Omit {partition} only when there is no team, tenant, business unit, source system, or other durable partition. Document that decision in process/data.md.

When this workspace needs daily distillation, every run writes dated source.md and summary.md artifacts under the matching {yyyy}/{mm}/{dd} bucket. current/ files may be maintained as latest views, but they are not the only record of daily judgment.

Accumulated-action paths, when queue state is part of the workspace:

data/{partition}/daily-triage/{yyyy}/{mm}/{dd}/accumulated-actions-{yyyy-mm-dd}.json
data/{partition}/daily-triage/{yyyy}/{mm}/{dd}/removed-actions-{yyyy-mm-dd}.json
data/{partition}/daily-triage/{yyyy}/{mm}/{dd}/actions-{yyyy-mm-dd}.md

Simpler knowledge paths, when object-first layer files are enough:

data/<localized-object-type>/<object-id>/<yyyy>/<mm>/<dd>/<localized-layer>.md
data/<localized-object-type>/<object-id>/current/<localized-layer>.md

Use only current/ when latest state is enough. Document the date-tracking rule in process/data.md.

Reporting And Output

Delete this section if reports, decks, or exportable summaries are not part of the workspace.

Treat natural-language requests such as report the current status, make a deck, 汇报当前情况, or 输出给业务看的 deck as reporting requests.

Map each request to one scope before reading files:

• current-status
• single-object
• object-group
• period-review
• custom-question

Load process/<localized-reporting-layer>.md before producing report or deck artifacts.

Default dated output paths:

my-work/<yyyy>/<mm>/<dd>/<scope>.md
my-work/<yyyy>/<mm>/<dd>/output/<scope>.marp.md
my-work/<yyyy>/<mm>/<dd>/output/<scope>.pdf

Use the repo-native export chain: markdown, Marp HTML/PDF, PPTX, or a documented deterministic exporter.

Validation

Minimum validation flow:

<host-executable validation flow>

Do not claim validation unless the flow ran.

Required Domain Inputs

Delete this section only after every item has been answered or explicitly marked not applicable.

• Object types:
• Stable ID source for each object type:
• Evidence sources and freshness rule:
• Partition rule:
• Daily distillation rule:
• Scenario lenses:
• Batch distillation script and manifest rule:
• Refresh script:
• Source-generation script:
• Index script:
• Validation script:
• Accumulated-action script:
• Posting script and external write gates: