Skip to content

mindbox-cloud/agent-skills / plugins/mindbox/skills/flow-create

flow-create

Verified

Build a correct draft marketing scenario (flow) from a plain-language request: read the flows reference, resolve the entities the design needs to real ids, create and fill the blocks, wire the outputs, verify every block by reading it back, and hand over a link. Use when asked to create, build, assemble or set up a new flow, or to add blocks to an existing draft; when asked to move, transfer or migrate a mechanic or a flow from one project to another, whether one or a batch of them; when asked for several flows in one go; and when asked to rebuild here the logic a client is moving off an external marketing-automation platform — in whatever language the request comes in, and whatever the reader calls the thing: a flow is also a scenario or a journey, a send step is a mailing or a campaign, an entry filter is an audience or a segment. Don't use when: the flow already exists and only needs describing or checking — this skill builds, it does not report; the request is to launch, stop or delete something, which stays a human decision.

references/feedback-and-selfcheck.md

View on GitHub
# Feedback and the self-check

Read this at the first docs mismatch, at step 7c, and when the user accepts the result (step 8).

## The feedback message

`feedback` is on the same connection as the flow tools; take its name from your tool listing. Its one
free-text field carries all the structure (any other argument its schema asks for is this build's
project). **No such tool — skip silently:** it is telemetry, never a shortfall, never told to the user.
First line the literal marker, then one `key=value` per line; `skillVersion` is `SKILL.md`'s
`metadata.version`. Never a secret, personal data, a filter body or a customer's value.

## A documentation mismatch

**The moment you establish** that live behaviour contradicts the wiki or `SKILL.md` — a recovery that does
not reproduce, a field said to be refused that was stored, a misdescribed schema — send, unasked, unnarrated:

```
[skill-feedback:v1]
skill=flow-create
skillVersion=<metadata.version>
source=agent-observation
kind=docs-mismatch
doc=<the wiki id, or the SKILL.md or reference file § section, that is contradicted>
expected=<what the doc says, one line>
observed=<what actually happened, one line>
calls=<in order: tool name + a one-line argument summary, per call that establishes it>
```

At most **three per session**, one per distinct mismatch: feedback spends the client's tokens, so it
is capped, but one slot can be too few. It replaces no handling — the live contract still wins.

## Rating the build

**Once per conversation thread, after the user accepts** (`references/hand-over.md` → *Actions*, last
item; several flows: once all are accepted), one unobtrusive line in their language asks for a 1-to-10
rating, flows and mailings together. Never earlier; asked once — an unanswered ask is spent, not
repeatable, even when a later build starts in the same thread. **A number** → send:

```
[skill-feedback:v1]
skill=flow-create
skillVersion=<metadata.version>
source=user-rating
score=<1-10>
comment=<their words verbatim; if long, a faithful one-line retelling; omit the line if none>
flows=<the flow ids built this session>
iterations=<the final metadata.iteration value>
```

**No number** (silence, refusal, words alone) sends nothing, is never pressed for, blocks nothing.

## The self-check

After step 7b's final clean validation, launch **one sub-agent per flow built this session** on the
read-only structural audit skill (it checks a flow for known anti-patterns; name from your skill
listing, "use that skill" as the first sentence, per-flow mode, the flow id and version). No such skill,
no check. It reads no wiki and lacks this session's context: **its findings are leads, not verdicts**.

- **Contradicts a decision the user made explicitly, or a documented platform behaviour** → drop it.
- **Names a real risk the user has not decided on** → into the hand-over, plainly, in their language:
  *Actions* if it needs their decision, *Suggestions* if optional. Never a separate report, a check
  number or the audit's vocabulary.

**Nothing survives → say nothing about the check.** No re-runs; a failed or unusable sub-agent is skipped
silently (a mismatch it exposed still goes to feedback). Read-only: no write budget, no reopened validation.

# References

- `SKILL.md` → step 7c, step 8, `## Where to look`; `references/hand-over.md` → *The three parts*
brainstormingobra/superpowersYou MUST use this before any creative work - creating features, building components, adding functionality, modifying behavior, or planning anything new, in software or out of it (a talk, a business, a renovation).Stars: 322kUpdated Claude CodeCodexVerifiedInstall
brand-guidelinesanthropics/skillsApplies Anthropic's official brand colors and typography to any sort of artifact that may benefit from having Anthropic's look-and-feel. Use it when brand colors or style guidelines, visual formatting, or company design standards apply.Stars: 198kUpdated Claude CodeCodexVerifiedInstall
frontend-designanthropics/skillsGuidance for distinctive, intentional visual design when building new UI or reshaping an existing one. Helps with aesthetic direction, typography, and making choices that don't read as templated defaults.Stars: 198kUpdated Claude CodeCodexVerifiedInstall
mcp-builderanthropics/skillsGuide for creating high-quality MCP (Model Context Protocol) servers that enable LLMs to interact with external services through well-designed tools. Use when building MCP servers to integrate external APIs or services, whether in Python (FastMCP) or Node/TypeScript (MCP SDK).Stars: 198kUpdated Claude CodeCodexVerifiedInstall