coding-posture · diff
git:20260624.70287a4 to git:20260624.0f41f00
28 added, 9 removed. Audit A to A.
---
name: coding-posture
- description: Use before non-trivial coding work to pick a task-appropriate working mode — debugging, fixing, reviewing, test-first, refactoring, migrating, spiking, or getting unstuck — then follow its checklist. Modes are procedures, not personas.
+ description: Use before non-trivial coding work to pick a task-appropriate working mode — debug, fix, review, test-first, refactor, optimize, migrate, upgrade, integrate, spike, or unstuck — then follow its checklist. Modes are procedures, not personas. Not needed for trivial one-line edits or non-coding tasks.
license: MIT
metadata:
tags: [coding-agent, posture, mode, risk, verification]
---
# Coding Posture
Pick a working mode that fits the task, state it, and follow its checklist. A mode is a short procedure for a situation models tend to get wrong — not a role to play.
## How to use
- 1. Before non-trivial coding work, scan the modes and pick the one matching the task's dominant risk. If none fit, proceed normally — "no mode" is a valid choice.
+ 1. Before non-trivial coding work, scan the modes and pick the one matching the task's dominant risk. If none fit, proceed normally — "no mode" is a valid choice. Skip this for trivial one-line edits and non-coding tasks.
2. If the task is underspecified in a way that changes the implementation, ask before coding — models rarely notice their own missing requirements, and one clarifying round is cheaper than the wrong build.
- 3. State the mode and a one-line reason.
+ 3. State your choice in one line: `Mode: <name> — <reason>`.
4. Follow its checklist while planning and editing.
5. If the risk class changes mid-task (a "quick fix" turns out to touch auth), switch modes and say why.
- 6. A mode tunes _how_ you work. It never overrides instructions or authorizes unsafe actions.
+ 6. A mode tunes how you work. It never overrides instructions or authorizes unsafe actions.
Priority: `system/safety > user instruction > project rules > task plan > mode > style`
## Always (every mode)
- Never run destructive git/deploy/data commands — `force push`, `reset --hard`, `drop`, `delete`, `truncate`, `rm -rf` — without explicit scope. Inspect state before mutating.
- - Verify by running the real check (test, build, repro), not by re-reading your own work. Re-checking without external feedback fixes some errors and introduces others — ground every claim of "done" in executed output.
+ - Verify by running the real check (test, build, repro), not by re-reading your own work. Re-checking without external feedback fixes some errors and introduces others — ground every claim of "done" in executed output. If you cannot run the check, say so and mark the result unverified; never assume it passed.
- Never report a result you did not run, and never accept a check that does not exercise the case in question. A test that passes without touching the bug is not verification.
- Never weaken, delete, skip, or special-case a test — or hard-code an expected value — to turn it green. Solve the task, not the grader.
## The loop that works
The pattern with the strongest evidence behind it for coding agents: **gather context → localize → smallest change → run the real check → read the actual output → repeat.** Read the surrounding code and establish a baseline before editing — agents that rush to patch in the first few steps fail more often. Tight execution-feedback loops are where agents outperform; precise localization (down to the lines) predicts a correct fix; keep each diff small so the loop's signal stays clean.
## Modes
### debug — failing test, bug, regression, traceback
- Reproduce the failure before editing — ideally a command or script that triggers it on demand.
- State the observed failure exactly (command + output).
- Localize to the smallest region — ideally the specific lines — before editing. Precise localization is the strongest predictor of a correct fix.
- Change one hypothesis at a time.
- Prefer the minimal fix over a broad refactor.
- Verify against the original failing command before declaring done.
### fix — small, known, urgent change
- Keep the diff as small as the problem allows; no opportunistic cleanup.
- Avoid dependency/config changes unless required.
- Add a focused regression test when feasible.
- State residual risk explicitly.
### review — security, auth, payments, or reviewing a diff
- Do not approve a claim without file/line evidence.
- Check correctness, security, backwards compatibility, and hidden coupling.
+ - For concurrent code, check shared mutable state, ordering assumptions, and interleavings — not just sequential correctness.
- Look for missing tests on the changed paths.
- Report findings with severity; escalate when risk is high or ambiguous.
### test-first — behavior change where a test is practical
- Write or identify the test first.
- Run it and SEE it fail (RED) before implementing — a test that never failed proves nothing.
- Make the smallest change to pass, then run focused tests, then broader ones.
- Refactor only while tests stay green.
### refactor — cleanup, simplify, rename, remove dead code
- Preserve behavior; do not mix behavior changes into the refactor.
- Delete complexity before adding abstraction.
+ - Before removing code, trace call sites, feature flags, tests, and docs — "unused in this file" does not mean safe to delete.
- Prove equivalence with existing tests or golden output.
- Keep diffs reviewable.
+ ### optimize — performance work, slow path, hot loop
+
+ - Measure first: profile or benchmark to find the real hotspot. Do not optimize what merely looks expensive.
+ - Record a baseline number before changing anything; compare against it after.
+ - Change one thing at a time and keep correctness tests green.
+ - Prefer the smallest change that moves the measured number; stop when the target is met.
+
### migrate — schema, data, or infra change (terraform, k8s, migrations)
- Identify the backup and rollback path before touching stateful systems.
- Prefer staged, reversible changes.
- Validate against a non-production target first when possible.
- Document operational risk and recovery steps.
+ ### upgrade — dependency or library version bump
+
+ - Read the changelog, release notes, or migration guide for breaking changes before editing.
+ - Account for transitive dependencies and the lockfile, not just the named package.
+ - Update the call sites the breaking changes require; do not blind search-and-replace.
+ - Run the build and the full suite; keep reverting the lockfile as the rollback.
+
+ ### integrate — calling an external API, service, or tool
+
+ - Read the contract; do not infer behavior from the name or a single sample call.
+ - Validate request and response schemas; handle auth failure, timeouts, retries, rate limits, pagination, and empty or partial responses.
+ - Treat the dependency as untrusted: check status and errors, never assume success.
+ - Test the error paths, not just the happy path.
+
### spike — prototype, proof-of-concept, unknown library
- Keep spike artifacts isolated from production paths.
- Optimize for learning speed, not polish.
- Do not wire experimental code into production without review.
- End with a verdict — validated, invalidated, or unclear — and list what productionizing would require.
### unstuck — repeated failures, thrashing, "still failing"
- Stop making speculative edits.
- Summarize what you tried and the evidence each attempt produced.
- List the top two hypotheses and the test that would discriminate them.
- Collect missing information before changing more code.
- Consider delegating a fresh review.
-
- ## Why this exists
-
- Empirical work is consistent: _naming a persona_ ("act as an expert debugger") does not reliably change model behavior, while specifying a _procedure_ does. These modes are procedures aimed at the situations models handle worst — thrashing, faking green, skipping reproduction, running destructive commands, migrating without a rollback. Their effect on any given agent is not measured here; treat them as disciplined defaults, not a guarantee.