akshay-diwadkar/first-four-hours

Skills for coding agents that settle the decisions before the edits, and leave a file you can open and check.
Install · See it work · Skills · Docs · Recipes · Changelog
> "Give me six hours to chop down a tree, and I will spend the first four sharpening the axe."
Your agent said done. Then came the missed requirement, the decision it quietly reopened, the caller it broke.
First Four Hours does the sharpening first: each skill does one job before the work moves on, and leaves a file you can open and check. On a small fix, that takes minutes.
## Install
You need Node.js for npx and Python 3.11 or later. From your project's root folder:
bash
npx skills add akshay-diwadkar/first-four-hours --skill ''
Then type /provision in your agent. It installs any Python packages the installed skills declare and sets up the suite's subagents on the models you choose.
Claude Code plugin, Cursor, or one skill only
On Claude Code 2.1.275 or later, install the suite as a plugin from a Claude Code session:
text
/plugin install first-four-hours --marketplace akshay-diwadkar/first-four-hours
For Cursor, see Install as a Cursor plugin. Install First Four Hours covers installing one skill, choosing the agent, installing for every project, and checking that the install worked.
## See it work
Name a skill and say what is at stake. For example (illustrative; FreshCart is a made-up online grocery service):
text
/blueprint : Add per-client rate limiting to the FreshCart public API before the holiday sale. Partner integrations must never be throttled, every response under the limit must stay exactly as it is today, and a cheaper model will build this overnight with nobody watching.
Blueprint reads the project without editing it and seals blueprint.md. Its Goal, shortened, is what you read and approve before anything is edited:
markdown
## Goal
Approach: add a per-client token bucket to the existing middleware chain, keyed on the API key when one is present and on the client address otherwise. Partner keys bypass the limit. A limited request gets 429 with a Retry-After header; every other response is unchanged.
Load-bearing invariants: responses under the limit are exactly what they are today; partner keys are never limited; the client address comes only from the app's trusted proxy helper.
Main risk: /search/suggest shares the search handler, so it is limited too.
Below it, the plan cites the files its facts come from, gives ordered instructions, and names the check for each outcome. /construct then carries the plan out; if the plan turns out wrong against the real code, Construct stops and reports instead of deciding on its own. Anatomy of a sealed blueprint walks through a whole plan.
## Built with itself
This repository is planned, built, and reviewed with its own skills, and its pull requests keep the record:
Epic #27: research to shipped features. One Ideate run researched how real examinations and good tutors work; Ticket split it into four issues, closed by four pull requests.
#108: the skill that didn't pay its way. Atlas gave agents a map of the codebase. Tested on Django and Temporal, with answers graded blind, the map cost about 18% more per question on one and the same on the other, with no gain in accuracy. So Atlas was deleted from the suite, through a sealed Blueprint that Construct carried out.