Skip to content

generic-automation-and-it/smooth-ai-stockanalysis / .agents/skills/ai-brain-dump

ai-brain-dump

Verified

Start and run a listen-first braindump session for tickets, issues, ADRs, worktasks, PR descriptions, requirements, designs, or implementation plans. Use when the user says they want to braindump, think out loud, capture rough requirements, or provide context incrementally before asking the agent to synthesize, update a ticket, create docs, or perform implementation work.

FileKindSize
AGENTS.mdmarkdown1.2 KB
README.mdmarkdown4.3 KB
SKILL.mdskill md8.1 KB
agents/openai.yamlconfig263 B
# ai-brain-dump — Intent & Token-Usage Review

> Companion notes to [`SKILL.md`](./SKILL.md). Explains what this skill is for and why
> "listen-first" actually saves tokens — including where it doesn't.

## Intent review

The skill is a **listen-first capture mode**. Its whole posture is **deferral**: acknowledge
the session, hold context, and **do nothing** — no file reads, no tool calls, no clarifying
questions, no spec generation — until the user explicitly says "synthesize / update / create."

It has 4 phases:

1. **Initialize**
2. **Listen**
3. **Compare-or-Clarify**
4. **Synthesize-on-request**

…plus a "capture model" of mental buckets it surfaces **only when asked**. The intent is
coherent and the guardrails reinforce it well (e.g. *"if the user says just listen, obey even
if you notice issues"*).

The default with **no switches** is still pure silent listen-first. Questioning and grounding are
strictly opt-in via the switches documented below (and in [`SKILL.md`](./SKILL.md)).

## Does "only listen, summarize at the end" reduce token usage?

**Yes, meaningfully — but understand the mechanism.**

### Where it genuinely saves

- **No tool calls during listening — this is the big one.** The skill forbids browsing/reading
  code while capturing. Tool results (file dumps, greps) are the heaviest token sink because
  they're injected into context **and re-sent on every subsequent turn**. Suppressing them
  during a long capture phase is the largest lever.
- **Short responses + no full spec per message** — cuts output tokens per turn vs. an eager
  agent that drafts a complete requirements doc each time.
- **"Don't expose the whole capture every turn"** — avoids O(n²) growth where the assistant
  restates the entire accumulating capture on every message. Synthesis happens once at the end
  instead of N partial rewrites.
- **No early clarifying loop** — avoids question/answer churn that inflates turns.

### Where it does *not* save (important caveat)

- **It does not compress input context.** Every braindump message stays in the conversation and
  is re-billed as input on each turn — so input tokens still grow linearly and re-accumulate.
  The "mental running capture" isn't real hidden state; the model reconstructs it from history
  each turn. The savings are on **output and tool-result bloat**, not on conversation
  accumulation.
- **Phase 3 (Compare/Clarify) explicitly re-enables tools.** If a user triggers it often, the
  tool-suppression savings erode.
- **Synthesis is one concentrated large pass** — cheaper overall than many partial specs, but
  not free.

### Net

Compared to a default "helpful" agent that eagerly reads files, asks questions, and
re-summarizes each turn, this skill should **reduce total tokens** — primarily by **deferring
all tool use** and **brevity during capture**, not by shrinking context.

## Modes & cost trade-off

The listen-first default can be relaxed with opt-in switches. They sit at very different price
points — the cheap ones add conversational turns; the expensive ones re-enable the tool-result
bloat the default was built to avoid.

| Switch | Cost impact | Why |
|--------|-------------|-----|
| _(none)_ | **Baseline** | Pure silent listen-first. No questions, no tools. |
| `--oktoask` (tool-free questions) | **Small** | A few extra output tokens (the question) + occasional extra turns (re-billing the growing prefix at cache-discounted rates). The big token sink — tool-result bloat — stays OFF. |
| `--thinking` (liberal questioning) | **Moderate** | More questions → more round-trips → more turns re-billing context. Bounded by conversation length. |
| `--oktoreaddocs` | **Large** | Reintroduces exactly what the skill avoids: file/code dumps injected into context and re-sent on every subsequent turn. The heaviest lever. |
| `--oktowebsearch` | **Large** | Web-search result payloads are big and likewise re-billed each turn. |
| Conclusion liberal Q&A | **Small, one-time** | A single bounded burst before synthesis; often *net-negative* cost because it prevents a wrong-synthesis + rework loop. |

**Bottom line:** `--oktoask` and `--thinking` cost little and can pay for themselves by catching
misunderstandings before synthesis. `--oktoreaddocs` and `--oktowebsearch` re-enable the skill's
*primary* savings as a cost — use them deliberately, not by default.
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