Growing Assistant / 成长助手
这不是一个“帮你做某件事”的技能,是一个“帮你长出一个助手”的技能。 它先帮你搭好底座,再陪你跑通第一个工作流,最后把你的纠正变成规则和评测。
What this is
A meta-skill: it does not replace your resume/email/weekly-review skill. It builds the workspace, memory, rules, and feedback loop that make those skills yours and better over time.
It is deliberately small:
- one local workspace,
- one first workflow,
- one correction loop,
- one effort switch.
No database server. No platform. No multi-agent ceremony.
Hard rules
- One workflow at a time. Do not scaffold five workflows before one works end-to-end.
- File-first. Chat is not memory. Anything that must survive a new session goes into a file.
- Every fact cites a file. Use
path#headingorpath:line. No source, no claim. - Unknown is allowed. Write
not in docsor待补; never fill gaps with plausible guesses. - Corrections become rules. “I’ll remember” is not a mechanism. Write a rule + an eval case.
- Privacy is a hard gate. Secrets, customer names, and company-internal data go in
private/; public output must not contain them. - Stop at gates. G0–G3 exist to prevent a confident half-product.
- Build one loop, not a platform. If a feature does not make the first loop more verifiable, cut it.
Modes
| Mode | When | Output |
|---|---|---|
bootstrap | New user / empty workspace | assistant/ scaffold + PROFILE.md + first WORKFLOW.md |
run | User gives a real task | output/<date>-<task>/report.md + evidence.md + handoff.md |
grow | User corrects a result | new rule in RULES.md + new case in evals/ + re-run |
export | Workflow is stable | standalone skills/<name>/SKILL.md + a static case page |
If the user asks “build my assistant”, start at bootstrap. If they give a task and a workspace exists, go to run. If they say “that’s wrong / next time do X”, go to grow.
The 5-step loop
1. Profile (5–8 questions)
Ask one at a time, with a recommended default. Stop when you can fill:
- who the user is and what they do,
- 3 current goals,
- one repeated task that wastes time,
- what “done” looks like for that task,
- what data/tools are allowed,
- what must never be shared,
- preferred language/tone,
- effort preference (fast vs thorough).
Write PROFILE.md. Do not ask for what is already in the repo.
2. Workspace
Create the file contract below. Use the script:
python scripts/init_assistant.py --root ./assistant --name "My Assistant"
The workspace is the state. If it is not on disk, it does not exist.
3. First workflow
Pick one task that is frequent, painful, and verifiable. Write WORKFLOW.md:
- trigger,
- inputs,
- steps,
- output contract,
- acceptance checks,
- stop conditions.
Run it once. Produce a real artifact, not a chat answer. If it cannot produce an artifact, pick another workflow.
4. Correction
Ask: “What exactly is wrong — fact, spec, or preference?”
- Fact → fix the source file / citation.
- Spec → change
WORKFLOW.md. - Preference → write a rule in
RULES.md.
Then add one eval case to evals/. Re-run the same task. The second run must show the correction taking effect.
5. Effort
| Effort | Behaviour | Use when |
|---|---|---|
low | answer directly; one pass | simple, reversible |
medium | draft → self-check → revise once | normal work |
high | solve → check → revise; keep best version; record evidence | high stakes / long tasks |
Record the effort level in the output. High effort without extra checks is just slower low effort.
File contract
assistant/
├─ PROFILE.md # who, goals, constraints, preferences
├─ RULES.md # durable rules, one per correction
├─ WORKFLOW.md # the first workflow
├─ memory/ # dated notes, decisions, open loops
├─ docs/ # source material (read-only)
├─ workflows/ # additional workflows, one file each
├─ evals/ # good/bad cases, one file each
├─ output/ # dated task outputs
├─ skills/ # distilled standalone skills
└─ private/ # secrets/customer data; gitignored
Gates
| Gate | Pass condition |
|---|---|
| G0 Profile | PROFILE.md has who/goals/constraints/privacy/preferences |
| G1 Workflow | WORKFLOW.md has inputs/output/steps/acceptance/stop |
| G2 First run | artifact + evidence + no uncited factual claim |
| G3 Growth | one new rule + one new eval case; second run passes |
Do not skip a gate because the user is in a hurry. A skipped gate is the most expensive shortcut.
References
references/01-bootstrap.md— questions + profile templatereferences/02-workspace.md— file contract and namingreferences/03-run.md— run contract and acceptancereferences/04-grow.md— correction → rule → evalreferences/05-export.md— distill a workflow into a standalone skill + case page
Definition of done (v0.1)
A new user can:
- install the skill,
- answer 5–8 questions,
- get a workspace with profile/rules/workflow,
- run one real task with a cited artifact,
- correct one thing,
- see the correction persist in a new session.
If any step needs the skill author to explain it, the skill is not done.