Your team's playbook,
as a file the agent
actually executes
Prompts are read. Playbooks are executed. ContextJemo turns the way your team actually works into declarative files — AET tactics and kernel MD — that agents run, humans review, and git tracks. Cheap models become enough, because the structure carries the intelligence.
# A real tactic — steps are prompts, not code schema_version: "0.2" aet: id: saas-pricing-monitor version: 0.1.0 steps: - id: s2 name: 对比历史快照，检测定价结构变化 instruction: > 读取最新两个快照文件，逐一 diff… agent_requirements: [data-processing, analysis] depends_on: [s1] checkpoint: success_criteria: > diff 报告生成；changed_vendors 为空时标记"无变化"
Three properties that make
playbooks into assets
Semantic deposits
A tactic is a file, not a chat log. It lands in git, diffs like code, rolls back like code. A teammate can read what your agent was told to do — and why — without asking you.
#CHARTER-03.7Semantics replicate
One tactic, many agents. The same declarative file drives different models and different platforms. When you improve the playbook, every consumer inherits the fix — the improvement is not re-explained.
#BRIEF-1.3The layer becomes AI
The semantic layer is not static documentation. Agents read it, execute against it, and write back to it: update_kernel evolves the kernel; cross-model verification gates the result. The layer improves because it is used, not because someone maintains a wiki.
#SPEC-1.0Two artifacts.
Both plain text.
Everything in ContextJemo is a prompt — at three levels of structure. The kernel holds what must never change. The AET tactic holds what to do. The prompt template holds the smallest reusable unit.
Values & red lines
≤ 90 lines, hard budget. Axioms, self-cognition, bottom lines. A validator rejects it if it overruns — attention is a scarce resource, and restraint is a capability.
#CHARTER-03.3Steps & checkpoints
1–20 steps. Each step is a natural-language instruction with a machine-checkable success criterion. You declare what "done" means; the runtime decides how to get there.
#SPEC-2.1### 四、红线规则（Bottom Line） | 红线 | 触发动作 | |------|---------| | R1：无来源结论 | 结论不写入报告，改为「待核」标注 | | R2：营销内容当事实 | 降级为「参考」，注明来源性质 | | R3：超出知识截止日期 | 标注「数据可能过时」 | | R4：单一来源支撑关键结论 | 强制要求第二来源印证 | | R5：预算超支（>80%）| 停止深度检索
Black box makes you believe it.
Declarative makes you see it.
Mainstream agent orchestration is a black box: you hand over a task, and how it scheduled, how many tokens it burned, what each step produced — none of it is visible. ContextJemo takes the opposite road.
| Dimension | Black-box agent orchestration | ContextJemo declarative |
|---|---|---|
| Scheduling logic | Hidden in the engine, invisible | AET YAML, in plain sight — you define who and in what order |
| Each step's output | Invisible | Visible and costed, per step |
| Where experience accumulates | Scattered across chat logs | In git — diffable, shareable, compounding |
| Control | Handed to the engine | Kept by you — edit the YAML, you edit the behaviour |
| Iteration | Waiting on a vendor release | Yours — change one line, run once, see the effect |
| When it breaks | You don't know which step died | Machine-judged checkpoints point straight at it |
Say → make → try → use → grow
One Paper: a playbook's full life, from a sentence of intent to an asset that compounds. The demo is not a throwaway — it is the fuel the asset is refined with.
Say
Describe the outcome in one sentence. No interrogation, no three-question intake.
Make
The co-creation brain drafts the tactic immediately. Open questions are marked inline, not asked back.
Try
Rehearsal with visible outputs and visible cost. Missing inputs are labelled honestly, never invented.
Use
The tactic lands in the library as YAML — versioned, shareable, runnable by any agent.
Grow
Run data and human confirmations feed back. The asset sharpens with every confirmed run.
Files first.
Platform second.
The repository is Apache-2.0. The formats are documented. AET spec v1.0 is frozen — and the triggers for independently open-sourcing it are written down, not aspirational.
# Node ≥ 22 git clone https://github.com/xingkongxia1988/contextjemo cd contextjemo && pnpm install # conversational mode — just say what you want npm run chat # › /run → s1 ✓ 13.5s s2 ✓ 7.5s # › /save → ✓ saved ./tactics/your-tactic.aet.yaml
Bring one real workflow.
We'll show you its ledger.
No slides, no discovery workshop. Bring a task your team already does by hand — we'll turn it into a tactic and show you what it costs, step by step.