Piekwerk
@piekwerk.bsky.social
Version-pinned AI coding agent config kits for 12 stacks. One baseline rendered into AGENTS.md, CLAUDE.md, Cursor rules and Copilot instructions. Validator-tested: 12/12 green or it does not ship. Free Next.js sample on the pinned post.
Never let an agent edit its own rules mid-task. It tunes the rule to fit the mistake it just made, then quotes the edit back at you.
October 8, 2026 at 3:27 PM
Never let an agent edit its own rules mid-task. It tunes the rule to fit the mistake it just made, then quotes the edit back at you.
Diff your rules file after every tool upgrade. Changelogs name the new features, they never name the paragraph that broke. Ten seconds with git diff catches more silent breakage than any checklist.
October 7, 2026 at 2:26 PM
Diff your rules file after every tool upgrade. Changelogs name the new features, they never name the paragraph that broke. Ten seconds with git diff catches more silent breakage than any checklist.
New post: a 22-rule scanner for the files nobody audits. Agent configs, MCP servers, hooks. 12% of public MCP configs have a hardcoded secret. Launch week: code LAUNCH for 10 off. https://dev.to/piekwerk/your-ai-agents-config-files-have-secrets-in-the-clear-a-22-rule-scanner-that-finds-them-3hhl
October 7, 2026 at 9:42 AM
New post: a 22-rule scanner for the files nobody audits. Agent configs, MCP servers, hooks. 12% of public MCP configs have a hardcoded secret. Launch week: code LAUNCH for 10 off. https://dev.to/piekwerk/your-ai-agents-config-files-have-secrets-in-the-clear-a-22-rule-scanner-that-finds-them-3hhl
The session that builds defends its own output. Judge the work in a fresh session against a checklist kept next to the rules. Same model, no stake in what it finds.
October 6, 2026 at 3:27 PM
The session that builds defends its own output. Judge the work in a fresh session against a checklist kept next to the rules. Same model, no stake in what it finds.
Every tool now wants its own instructions file. Before adding a fourth, check what actually differs. In our repo the real delta was 20 lines. The other 300 were copies drifting apart.
October 5, 2026 at 2:26 PM
Every tool now wants its own instructions file. Before adding a fourth, check what actually differs. In our repo the real delta was 20 lines. The other 300 were copies drifting apart.
Read your agent config out loud before shipping it. Rules that sound fine on the page turn ambiguous when spoken: which file does "the test suite" mean? Ambiguity you can hear is ambiguity the model will guess through.
October 4, 2026 at 4:27 PM
Read your agent config out loud before shipping it. Rules that sound fine on the page turn ambiguous when spoken: which file does "the test suite" mean? Ambiguity you can hear is ambiguity the model will guess through.
The cheapest context you will ever buy: a file listing what the agent got wrong last month and what you did about it. Memory that never records failures will happily repeat them.
October 4, 2026 at 2:26 PM
The cheapest context you will ever buy: a file listing what the agent got wrong last month and what you did about it. Memory that never records failures will happily repeat them.
Two agents, one repo, two config files, zero shared expectations. The fix is not a bigger prompt. It is one file that states what is true everywhere, imported by every tool file. Portability is a file, not a hope.
October 3, 2026 at 1:25 PM
Two agents, one repo, two config files, zero shared expectations. The fix is not a bigger prompt. It is one file that states what is true everywhere, imported by every tool file. Portability is a file, not a hope.
Stop rewriting rules the agent ignores. Log the misses for a week first. Half of them are loading problems, not obedience problems, and no amount of rephrasing fixes a file that never entered the room.
October 2, 2026 at 12:23 PM
Stop rewriting rules the agent ignores. Log the misses for a week first. Half of them are loading problems, not obedience problems, and no amount of rephrasing fixes a file that never entered the room.
A config rule with no owner and no test date is folklore. It worked once, somebody wrote it down, nobody can delete it. Tag each rule with who owns it and when it was last verified. Deletion becomes possible.
October 1, 2026 at 2:26 PM
A config rule with no owner and no test date is folklore. It worked once, somebody wrote it down, nobody can delete it. Tag each rule with who owns it and when it was last verified. Deletion becomes possible.
Start agents read-only until they earn write. Upgrading read-only to scoped write is one line. Rolling back after a full-write agent did something clever is archaeology.
September 30, 2026 at 1:26 PM
Start agents read-only until they earn write. Upgrading read-only to scoped write is one line. Rolling back after a full-write agent did something clever is archaeology.
Every rule inlined in instructions is paid on every task. Every rule behind a pointer is paid once, on demand. The question for each config block: does it need to be in the room, or just findable?
September 29, 2026 at 2:25 PM
Every rule inlined in instructions is paid on every task. Every rule behind a pointer is paid once, on demand. The question for each config block: does it need to be in the room, or just findable?
Pin versions in agent config like you pin dependencies: model, CLI, hook scripts. Behavior drift after an upgrade is not mysticism, it is an unpinned version quietly changing what your rules mean.
September 29, 2026 at 1:25 PM
Pin versions in agent config like you pin dependencies: model, CLI, hook scripts. Behavior drift after an upgrade is not mysticism, it is an unpinned version quietly changing what your rules mean.
Trust test for any rules file: put one nonsense word at the top, start a fresh session, ask the agent to repeat it. Thirty seconds, and it catches every silent skip. Run it after every tool upgrade, not once.
September 28, 2026 at 3:11 PM
Trust test for any rules file: put one nonsense word at the top, start a fresh session, ask the agent to repeat it. Thirty seconds, and it catches every silent skip. Run it after every tool upgrade, not once.
Name commands by their script alias, not the raw binary. npm run test survives a runner swap, and CI running the exact line from the config file means drift fails loud instead of inventing commands.
September 27, 2026 at 3:27 PM
Name commands by their script alias, not the raw binary. npm run test survives a runner swap, and CI running the exact line from the config file means drift fails loud instead of inventing commands.
Date every memory entry when written. Staleness stops being a feeling and becomes a lookup: when did this last matter? The agent drafts the cut list, you veto.
September 26, 2026 at 2:23 PM
Date every memory entry when written. Staleness stops being a feeling and becomes a lookup: when did this last matter? The agent drafts the cut list, you veto.
A hook that matches command names guards a word, not an action. Parse the command as shell, resolve through PATH and shebangs, or uvx tool run and its alias walk past the same check.
September 26, 2026 at 12:50 PM
A hook that matches command names guards a word, not an action. Parse the command as shell, resolve through PATH and shebangs, or uvx tool run and its alias walk past the same check.
When a rule stops firing, check the binary version before the rule text. Config guidance silently changed across releases twice this year. The file was fine; the loader underneath it had moved.
September 25, 2026 at 1:23 PM
When a rule stops firing, check the binary version before the rule text. Config guidance silently changed across releases twice this year. The file was fine; the loader underneath it had moved.
A rule in our config protected test files for two months without ever loading, because the frontmatter had alwaysApply true and the glob was decorative. Nothing errored. Tools do not warn about rules that never fire, which is why our validator fails the build on undeclared activation modes.
September 24, 2026 at 11:59 AM
A rule in our config protected test files for two months without ever loading, because the frontmatter had alwaysApply true and the glob was decorative. Nothing errored. Tools do not warn about rules that never fire, which is why our validator fails the build on undeclared activation modes.
We measured our own agent configs before fixing them: 4 copies of the same standards, 2 rules that had never fired, ~700 lines where ~150 earned their keep. The fix was structure, not better writing.
September 22, 2026 at 4:23 PM
We measured our own agent configs before fixing them: 4 copies of the same standards, 2 rules that had never fired, ~700 lines where ~150 earned their keep. The fix was structure, not better writing.
Your CLAUDE.md is 600 lines and your agent follows less of it than when it was 80. Compliance is not a bucket you fill, it is signal-to-noise. Most configs do not need more rules; they need a diet.
September 21, 2026 at 2:23 PM
Your CLAUDE.md is 600 lines and your agent follows less of it than when it was 80. Compliance is not a bucket you fill, it is signal-to-noise. Most configs do not need more rules; they need a diet.
Cursor rules have a frontmatter field most people never set: globs. A testing rule with globs pointing at test files loads only when Cursor edits test files. That is budget management: the rule costs context exactly when it matters and zero otherwise.
September 20, 2026 at 12:24 PM
Cursor rules have a frontmatter field most people never set: globs. A testing rule with globs pointing at test files loads only when Cursor edits test files. That is budget management: the rule costs context exactly when it matters and zero otherwise.
Monorepo agent config that works: one root AGENTS.md with what is true everywhere, per-package tool files for specifics, skills imported on demand. Root bloat is how 300-line budgets die.
September 19, 2026 at 1:23 PM
Monorepo agent config that works: one root AGENTS.md with what is true everywhere, per-package tool files for specifics, skills imported on demand. Root bloat is how 300-line budgets die.
A rule in our config protected test files for two months without ever loading, because the frontmatter had alwaysApply true and the glob was decorative. Nothing errored. Tools do not warn about rules that never fire, which is why our validator fails the build on undeclared activation modes.
September 18, 2026 at 8:57 AM
A rule in our config protected test files for two months without ever loading, because the frontmatter had alwaysApply true and the glob was decorative. Nothing errored. Tools do not warn about rules that never fire, which is why our validator fails the build on undeclared activation modes.
Shipped: 12 agent config kits (Next.js, React, Node, FastAPI, Django, Go, Rust, Rails, Laravel, SvelteKit, Astro, .NET), each rendering one baseline into AGENTS.md, CLAUDE.md, Cursor rules and Copilot instructions. 12/12 green or it does not ship.
September 17, 2026 at 12:27 PM
Shipped: 12 agent config kits (Next.js, React, Node, FastAPI, Django, Go, Rust, Rails, Laravel, SvelteKit, Astro, .NET), each rendering one baseline into AGENTS.md, CLAUDE.md, Cursor rules and Copilot instructions. 12/12 green or it does not ship.