A skill that tells coding agents to skip the paperwork
A 150-line Markdown file asks Claude Code, Codex and other agents to ship the deliverable first and treat receipts and hashes as optional.
Illustration: a finished assembly sits in the light on the workbench while the stack of blank forms behind it stays untouched.
The Forward Implementation First skill makes an agent sort every planned action into implementation, focused validation or administrative bookkeeping, and then do only the first two unless the user asked for the third.
At a glance
- Three classes before every action: semantic implementation, focused validation, administrative bookkeeping.
- Rule: do classes 1 and 2; skip class 3 unless the user asks or the artifact is part of the product.
- A pipeline stage may be replayed only for a real reason — missing or stale metadata does not count.
- About 150 lines of model-agnostic Markdown under the MIT license, installed by copying into ~/.claude/skills/ or ~/.codex/skills/.
- At retrieval on August 31, 2026, GitHub showed 137 stars and 4 forks.
Forward Implementation First is a small agent skill with a blunt premise: build the thing before you build the paperwork about the thing. Every planned action has to be labeled first, and only two of the three labels earn the agent permission to act.
The three labels
The first is semantic implementation — producers, consumers, schemas, the actual deliverable. The second is focused validation: checking behavior, checking schemas, measuring samples. The third is administrative bookkeeping — hashes, locks, receipts, dashboards, metadata — and the skill tells the agent to leave it alone unless the user asked for it or the artifact is itself part of the product.
A companion rule governs re-runs. According to the skill, a stage may be replayed or rolled back only for a real reason, such as changed inputs, malformed outputs or an updated dependency. Missing or stale metadata is explicitly excluded from that list.
The failure it was written for
The repository describes the incident behind the rules. After metadata drifted, an agent spent hours regenerating markers for stages whose output had never changed. It then declined to continue by hand, on the grounds that the pipeline could not issue a receipt for a stage a person had run. The author puts the cost at days of compute spent producing identical output.
How it ships
The skill is roughly 150 lines of model-agnostic Markdown released under the MIT license. It follows the Agent Skills convention and is installed by copying SKILL.md into ~/.claude/skills/, ~/.codex/skills/ or ~/.agents/skills/, which is how it reaches Claude Code, Codex and other agents. The stated targets are long-running pipelines, batch jobs and multi-stage workflows. At retrieval on August 31, 2026, GitHub showed 137 stars and 4 forks.
What we could not verify
Only one source was available for this report: the repository itself. Every claim about the skill's effect and its backstory therefore comes from the author's own description and has no independent confirmation. There is no published measurement of how much runtime or cost the rules save in someone else's pipeline, and no third-party account of running it. Star and fork counts move constantly and hold only for the moment named above.
FAQ
What does the Forward Implementation First skill do?
It has a coding agent classify each planned action as implementation, validation or administrative bookkeeping, then skip the bookkeeping unless the user requested it or it is part of the product.
Does it work with Codex as well as Claude Code?
The repository lists Claude Code, Codex and other agents that follow the Agent Skills convention; installation is a file copy into ~/.claude/skills/, ~/.codex/skills/ or ~/.agents/skills/.
Is Forward Implementation First free to use?
Yes. It is a Markdown file of about 150 lines under the MIT license, so it can be used and modified freely, and it installs no additional software.
