runfish
Get in touch
AI Docs The runfish AI toolbox

The tools I reach for, grouped by job.

A curated directory of the AI tools I use daily — chat, transcription, OCR, agentic automation, data access, and web-app development. Everything here earns its place by doing real work, not by being new.

And a note on the part that is still genuinely hard: finding the information you already have. Search itself needs rethinking — jump to why conventional search falls short.

Chat / interactive AI

Transcription

OCR (image text extraction)

Agentic AI / automation

Coding tools

A coding agent is only as good as the sentence you hand it. Prompts for coding is the running collection I paste from: closing a task so nothing is left half-done, cutting through agreeable answers, spotting hand-rolled code, and the weekly cleanup pass.

A good sentence still needs something to point at. The tools only pay off once the agent knows what it is building, and that is a document problem rather than a model one: a chain of six specs where each feeds the next, and one operational file the AI actually reads at the start of every session. The full guide follows.

Spec-Driven Development Guide Open in Notion

Spec-Driven Development Guide

.claude/agent/spec-buddy Getting Started 4 specs buddies you might need DOCs Document Hierarchy

The Document Dependency Chain (6 steps)

In the repo (docs/specs/): Charter, PRD, ADD, WBS, Roadmap, CLAUDE.md, CHANGELOG

→ Version-controlled. AI reads these.

CLAUDE.md : the only doc AI loads automatically. It links to specs and says what to work on right now. 

Charter (why) → PRD (what) → ADD (how) → WBS (work packages) → Roadmap (when) → CLAUDE.md (now)

Each document feeds the next. The Charter establishes why, the PRD defines what, the ADD defines how, the WBS breaks it into work packages, and CLAUDE.md is the living operational doc that AI reads automatically.

The PRP Flow

PRD → ADD → WBS → PRP (context for one work package) → Claude Code session

When to Update

At milestone boundaries. Not mid-implementation. Bump the version, log the change, move on.

Sub-Pages

See the pages below for the full methodology.

The division of labor: AI Drafts, Human Decides

AI is great at Human must decide
Drafting initial specs from context Priority levels (P0/P1/P2)
Generating code from specs Architecture trade-offs
Identifying patterns in codebase Scope boundaries (in/out)
Writing tests from acceptance criteria Success criteria thresholds
Suggesting alternatives Which alternative to pick
A Claude Code session rewriting the PRD, with Rewrite ADD, WBS and Roadmap queued behind it and Rewrite Project Charter already struck through.

CLAUDE.md as the Operational Spec

CLAUDE.md is special because it’s the only document AI reads automatically at the start of every session. This makes it the bridge between your planning artifacts and your AI collaborator.

What belongs in CLAUDE.md:

  • Links to spec documents (AI can read them when needed)
  • Current milestone and status
  • What to work on next
  • Key patterns and conventions

What does NOT belong in CLAUDE.md:

  • Full requirements (that’s the PRD)
  • Architecture details (that’s the ADD)
  • Session-specific notes
Reproduced from the Notion page · open the original

Data access

The information-search issue

Today, important information is often hidden in text, making it hard to find when we need it. It is unfortunate that millions of hours are spent gathering information, yet we’re unable to recover it on demand. When we’re left with an incomplete picture, the result is inefficiency, project delays, or the spread of false information.

To unlock the information hidden in text, available search methods are insufficient — new search tools must be developed. The new approach must focus on searching information, not documents. In this context, knowledge graphs hold significant potential: