Ten rules make coding-agent replies easier to start
i-have-adhd is a short instruction set with one job: make the first useful action hard to miss. Its 10 rules put commands, paths, or snippets first; number multi-step tasks; preserve progress; use concrete time estimates; and remove greetings and recaps. The skill stays active for a session until the user asks for normal mode.
Rules about suppressing tangents, limiting lists, ending on one next action, and restating state can change what the model selects and omits. That may help during a repair. It can hurt a security audit or evidence review where every finding must remain visible even when the answer becomes longer.
What happened when we ran it
Our sandbox installed commit cbe69fb in 6 seconds. The install added 0 packages and occupied 1 MB on disk. Npm audit reported 0 known vulnerabilities across critical, high, moderate, and low severities. The checkout had 57 files and about 1,829 lines of source, with 4 CI workflow files and a tests directory.
The repository had no build script or target, so our lab skipped the build. It exposed no test script or target to the lab runner, so tests were skipped. The contributor guide does name Python unit tests, compatibility checks, plugin validation, and an evaluation harness. Those documented commands are not a passing lab result.
Installation is broad, while activation differs by client
The guide covers more than 10 assistant environments and packaging routes. Claude Code and Codex have marketplace commands. Copilot reads the Agent Skills directory shape. Gemini CLI offers an on-demand command and an always-on extension. The repository also carries adapters or manifests for Cursor, OpenCode, Pi, OMP, Qwen, Kimi, Hermes, and Antigravity.
Codex disables model invocation, so the user must type $i-have-adhd unless the rules are copied into AGENTS.md. Claude Code can invoke the skill for one session or read a startup flag. Gemini's extension is always on after installation, while its command is opt-in. Test the exact route rather than assuming installation changed the next reply.
The five-item cap can remove findings instead of grouping them
Rule 9 says lists should stop at 5 items and suggests splitting longer material into sections. Issue 96 reports that models sometimes dropped relevant answers. The reporter removed the rule locally and said the rest still helped. Fourteen comments and an August 26 update make this an active design debate.
A five-step repair path is easier to finish than a 12-step wall of text. A vulnerability inventory or migration checklist becomes false if item 6 disappears. The skill says task requirements win, but models do not always resolve instruction conflicts consistently. Completeness-sensitive teams should rewrite the cap as a presentation rule that groups every finding.
Matter-of-fact errors can become unsupported diagnoses
Rule 8 asks the agent to state the error location, cause, and fix. Issue 99 reports a 3-trial evaluation where the skill condition scored lower on a partial-success case because the model invented a missing authentication header as the cause. The issue calls the result underpowered. The mechanism is still plausible: demanding a cause can pressure a model when evidence only shows a symptom.
A local edit can ask the agent to state the known failure and one diagnostic check when the cause is unconfirmed. The skill already says genuine uncertainty should remain and warns that deleting a meaningful hedge creates false confidence. Put that exception beside rule 8.
Always-on hooks deserve a visible verification step
Issue 129 reports that the Claude Code macOS desktop app did not export the plugin-root variable expected by version 0.2.0's startup hook. The hook exited with status 0 and no warning, leaving the user to believe always-on mode was active. The same installation worked in terminal Claude Code.
Issue 112 reports another failure on Grok 4.6: the active skill stopped tool calls, while disabling it restored action. That is one report on one model. It still justifies a small acceptance test after each model or client update: ask the agent to inspect a disposable file, make a safe edit, verify it, and report the result.
The evaluation harness is useful, but no release result is shipped
The repository includes cases, a rubric, runner definitions, blinded scoring, budget limits, retries, and resumable output. It asks evaluators to pin model and CLI versions and keep user plugins out of the baseline. Model choice and personal instructions can otherwise overwhelm the effect being measured.
No results file appears in the 57-file checkout, and issue 98 says one case could not pass because tools were disabled while the rubric required a repository edit. GitHub recorded the last push on August 26, 2026. The project had 25,562 stars and 22 open issues and pull requests when fetched, with another compatibility pull request opened on August 29. Each model and integration still needs its own measured trial.