Gleanin

gleanin

Rules from past failures. Apply every session.

Credentialed Write Gate

Rule:

1. Before writing with ambient credentials, verify scope ( whoami, config view).
2. Exact-match created resource name against intent.
3. On mismatch, roll back immediately.
4. Never mark done when acceptance probe failed.

Check: grep -c "inoffensif\|harmless\|can delete\|not needed" in session transcript — zero for any failed write.

Delegation Preflight

Rule:

1. Scan delegation payload against delegate's Known Limitations first.
2. Verify with git diff, not the delegate's [OK] lines.
3. On partial landing, name the limitation before hand-editing.
4. Structural integrity check every delegate-touched file.

Check: git diff --stat HEAD after delegation — changed file count must match expected.

Device First Check

Rule:

1. Search filesystem (find, grep, ls) before asking the user.
2. curl -sI before declaring anything "not deployed."
3. Syntax check + tail check before committing.

Check: grep -c "AskUserQuestion\|curl.*not.find\|edit.*then.*ask" in session transcript — near zero with this rule.

Reference First Porting

Rule:

1. Read the reference before editing any formula or algorithm.
2. Cite file:line per edit.
3. Apply supplied values verbatim before re-derivation.

Check: grep -rn "TODO\|FIXME\|guess\|maybe" — zero after a port.

Protocol index

Trigger-loaded protocols. When a trigger matches the current task, Read the file before proceeding.

• Exact form first — literal search, decisive git comparison, file-borne payloads — Load this in any session that (a) greps source code, (b) makes claims about → read ~/distillation/artifacts/exact-form-first.md
• Hang-bisect protocol — trace before retry, never escalate timeouts blind — Load this whenever a script or pipeline invocation times out (exit 124), → read ~/distillation/artifacts/hang-bisect-protocol.md
• Reproduce the report first — locate what the user sees before defending what you did — Load this whenever a user disputes visible state: "your changes don't → read ~/distillation/artifacts/reproduce-report-first.md
• Route census before auth policy — classify every endpoint once, edit once — Load this whenever you are about to add or change authentication, API-key → read ~/distillation/artifacts/route-census-before-auth.md
• Single-file batch edit — plan the change-set, then edit once — Load this whenever a task requires 3 or more changes to the same file — → read ~/distillation/artifacts/single-file-batch-edit.md
• boundary-aware handoff — When dividing work between two or more models (or model sessions): → read ~/fable-build/rules/boundary-aware-handoff.md
• content-integrity tests — For data-driven products (games, rule engines, content systems): → read ~/fable-build/rules/content-integrity-tests.md
• contract-first interface — When building a system that runs across multiple targets (offline, online, → read ~/fable-build/rules/contract-first-interface.md
• cost-time-quality accounting — Every build produces a cost accounting artifact: → read ~/fable-build/rules/cost-time-quality-accounting.md
• cross-model relay — When routing build phases across multiple models: → read ~/fable-build/rules/cross-model-relay.md
• deterministic default — Use fixed seeds for all world/level generation: → read ~/fable-build/rules/deterministic-default.md
• effort setting by phase — When configuring a build phase, set effort level explicitly: → read ~/fable-build/rules/effort-setting-by-phase.md
• goal-verification-report loop — Replace step-based phase execution with a goal→report pattern for → read ~/fable-build/rules/goal-verification-report.md
• phase stage verification — When a phase spans significant work, split it into stages (1/2/3): → read ~/fable-build/rules/phase-stage-verification.md
• review findings accounting — Every code review produces a numbered list of findings. → read ~/fable-build/rules/review-findings-accounting.md
• safety overfit testing — For products that use LLM outputs (chatbots, content generators, agents): → read ~/fable-build/rules/safety-overfit-testing.md
• thought-partner interview — Before writing ANY spec, template, or brief, interview the model as a → read ~/fable-build/rules/thought-partner-interview.md
• vision self-correction loop — When the build produces visual output (UI, 3D scenes, renders, screenshots): → read ~/fable-build/rules/vision-self-correction.md

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: pcx-wave
• Source: pcx-wave/glean
• License: MIT

Install and usage instructions live in the source repository linked above.