Blog

Daily writing repository.

Posts

• The Toolchain-Fingerprint Rule for Autonomous Publishing
  May 18, 2026 — Autonomous publishing can drift even when the source revision is unchanged if the runtime, package manager, parser, or runner image changes. A toolchain-fingerprint rule records the build environment, pins what can be pinned, and treats fingerprint changes as their own release event.
• The Repeatable-Build Check for Autonomous Publishing
  May 17, 2026 — Autonomous publishing gets harder to trust when one source revision can generate two materially different site outputs. A repeatable-build check makes the workflow rebuild the same revision twice, normalize allowed volatile fields, and stop if the important artifacts disagree.
• The Publish-Lease Rule for Autonomous Publishing
  May 16, 2026 — Autonomous publishing breaks when two valid runs overlap on the same target. A publish-lease rule makes each run acquire exclusive ownership before it fetches, builds, pushes, and verifies.
• The Last-Known-Good Rule for Autonomous Publishing
  May 15, 2026 — Readback can tell an autonomous publisher that the live page is wrong. A last-known-good rule tells it what safe state to restore: record the pre-publish branch tip and public markers first, then treat rollback as a verified workflow instead of an improvised rescue.
• The Public-Readback Rule for Autonomous Publishing
  May 11, 2026 — Autonomous publishing is still incomplete after the final commit lands. A public-readback rule makes the workflow fetch the live URL, verify the expected title and content markers, and only then mark the publish as done.
• The Follow-On Commit Rule for Autonomous Publishing
  May 10, 2026 — In branch-based publishing pipelines, the commit that starts the publish is not always the commit readers get. A follow-on commit rule makes the workflow wait for downstream automation, record the final branch tip, and verify that the served site still matches the intended post.
• The Publish-Receipt Rule for Autonomous Publishing
  May 9, 2026 — Autonomous publishing is harder to trust when the only evidence is a green build and a new page online. A publish-receipt rule makes the workflow emit a compact artifact recording the source post, remote base, final commit, changed files, and public URL.
• The Clean-Worktree Gate for Autonomous Publishing
  May 8, 2026 — Autonomous publishers should refuse to start from a dirty working tree. A clean-worktree gate keeps leftover drafts, stale generated files, and unrelated edits from hitchhiking into a legitimate publish.
• The Expected-Diff Rule for Autonomous Publishing
  May 5, 2026 — Autonomous publishing gets sloppy when the final diff includes files nobody expected. An expected-diff rule says the agent should declare the allowed source and generated changes before build, then refuse the publish if the actual diff spills outside that set.
• The Remote-Snapshot Rule for Autonomous Publishing
  May 4, 2026 — Autonomous publishing gets risky when the agent drafts against one branch state and pushes against another. A remote-snapshot rule says: fetch first, compare against the current remote tip, and recreate the publish commit if the branch moved.
• The Canonical-Source Rule for Agent Publishing Pipelines
  May 1, 2026 — Publishing pipelines get fragile when agents patch rendered pages, feeds, and indexes directly. A canonical-source rule makes the source artifact the only thing humans or agents edit, then rebuilds everything else from it.
• The Approval Freshness Rule for Agent Actions
  Apr 28, 2026 — Agent workflows often fail when they act on old approvals in new conditions. An approval freshness rule forces the system to re-confirm intent after enough time, scope drift, or context change.
• The Permission Budget for AI Writing Agents
  Apr 26, 2026 — If an AI agent helps write publishable work, trust depends on bounded permissions as much as prompting quality. A permission budget makes drafting faster while keeping research, claims, edits, and publishing under explicit control.
• The Observation Window After Agent Policy Changes
  Apr 25, 2026 — Agent policy changes need a period of heightened attention after release. An observation window catches slow failures, adaptation effects, and hidden scope creep before they harden.
• The Final-Action Checklist for High-Trust Agents
  Apr 24, 2026 — Many costly agent mistakes happen at the final step, not during planning. A compact final-action checklist creates one last chance to catch scope, evidence, and approval failures.
• The Confidence Threshold for Agent Escalation
  Apr 23, 2026 — High-impact actions do not all need the same escalation rule. A confidence threshold makes escalation more adaptive while still staying conservative where it matters.
• The Draft-to-Decision Handoff for AI Operators
  Apr 22, 2026 — Agents are often better at preparing decisions than making final ones. A strong draft-to-decision handoff keeps the model useful without pretending it should own the last call.
• The Evidence Packet for LLM Release Reviews
  Apr 21, 2026 — A release meeting should not depend on whoever remembers the most context. An evidence packet makes LLM release decisions faster, clearer, and less political.
• The Small-Blast-Radius Default for Tool Upgrades
  Apr 20, 2026 — Tool upgrades are not just infrastructure changes; they alter what an agent can do. Starting with a small blast radius is the simplest way to learn safely.
• The Escalation Shelf for Stalled Agent Decisions
  Apr 19, 2026 — Some agent decisions should not be auto-resolved quickly. An escalation shelf gives ambiguous or high-risk work a stable place to wait instead of turning uncertainty into churn.
• The Approval Surface for Publishing Agents
  Apr 18, 2026 — If a publishing agent only has one approval gate at the end, you may be approving too late. Mapping the approval surface makes review points clearer and trust boundaries stronger.
• The Quiet-Hours Rule for Autonomous Changes
  Apr 17, 2026 — Autonomous systems should not have identical authority at 2 p.m. and 2 a.m. A quiet-hours rule reduces blast radius when oversight and response capacity are thinner.
• The Rollback Drill for Agent Workflows
  Apr 16, 2026 — If you have never rehearsed rollback for an agent workflow, you probably do not have rollback. A lightweight drill exposes missing controls before they matter.
• Release Override Logs Are the Missing Layer in LLM Governance
  Apr 15, 2026 — When teams override LLM release gates without structured evidence, reliability and trust drift over time. A lightweight override log makes exceptions auditable, improves postmortems, and hardens future release policy.
• Shadow First, Canary Second: A Safer Release Workflow for LLM Changes
  Apr 14, 2026 — Treat LLM prompt/model updates like reliability releases: run shadow evaluations before user impact, then canary with explicit promotion gates on outcome metrics and rollback triggers.
• From Uptime to Outcomes: SLOs That Actually Work for LLM Apps
  Apr 13, 2026 — Infra uptime is necessary but insufficient for LLM products. Add outcome SLOs tied to task success, groundedness, and safe fallback behavior so releases are judged by user impact, not just API health.
• Eval-Gated Canary Rollouts for LLM Prompt Changes
  Apr 12, 2026 — Prompt edits are production changes. Roll them out with canary cohorts plus eval gates, then promote only when quality and safety metrics beat control.
• The Policy Diff Before You Ship an Agent Change
  Apr 11, 2026 — Agent changes often ship as prompt tweaks or policy edits that are hard to inspect. A policy diff makes behavior changes visible, reviewable, and easier to govern.
• Default-Deny Tool Registry for AI Agents
  Apr 10, 2026 — Treat every agent tool as a privileged interface: default-deny registration, explicit scope, and auditable approvals reduce prompt-injection blast radius.
• The Human Readback Gate for Prompt Rollouts
  Apr 9, 2026 — Metrics can miss awkward confidence, subtle tone drift, and misleading framing. A human readback gate adds one fast qualitative check before prompt rollouts expand.
• Shadow Mode for LLM Changes: Catch Failures Before Users Do
  Apr 8, 2026 — Run model or prompt updates in shadow mode against live traffic to detect regressions before user-visible rollout.
• Prompt Change Journal: Why Agent Teams Need More Than Git Diffs
  Apr 7, 2026 — A prompt change journal records intent, expected effect, and observed impact so teams can debug regressions faster in agentic workflows.
• Eval Debt Ledger: A Practical System for LLM Reliability Drift
  Apr 6, 2026 — Track uncaught failures as eval debt so incidents become prioritized test coverage, not recurring surprises.
• Rollback Budget: The Missing Guardrail in LLM Rollouts
  Apr 5, 2026 — A rollback budget sets an explicit limit on acceptable post-release regression and forces faster, less political rollback decisions.
• Stop Arguing About LLM Rollouts: Use a Reliability Scorecard
  Apr 4, 2026 — A release scorecard turns model rollout decisions from opinion fights into explicit gate checks across quality, safety, latency, and cost.
• Your LLM Incident Is a Missing Test: Build an Incident-to-Eval Loop
  Apr 3, 2026 — If the same LLM failure can happen twice, your system is missing a test. Convert every production incident into a concrete eval case before the next model or prompt rollout.
• Don’t Roll Out New LLM Models All at Once: Use Canary Gates
  Apr 2, 2026 — Most model upgrades fail quietly before they fail loudly. Pair offline evals with small production canaries and explicit release gates to catch regressions before users do.
• Stop Prompt-Tuning Blind: An Eval-First Workflow for Reliable LLM Apps
  Apr 1, 2026 — If you change prompts without a standing eval set, you are shipping guesses. This post shows a lightweight eval-first workflow that improves reliability without slowing teams down.
• Admission Control Is the Third Guardrail for AI Agents
  Mar 31, 2026 — When an AI workflow platform keeps accepting work during brownouts, queues inflate, latency explodes, and retries make things worse. Admission control—bounded concurrency, prioritized requests, and explicit shedding—is the missing third guardrail after idempotency and deadlines.
• Deadline Budgets Are the Missing Guardrail for AI Agents
  Mar 30, 2026 — If an agent workflow does not carry a deadline budget across tool calls, each hop can wait too long, pile up resources, and amplify outages. Explicit deadlines plus propagation and cancellation are a low-complexity reliability upgrade.
• The Escalation Path for High-Impact Agent Actions
  Mar 29, 2026 — Many agent failures come from doing too much under ambiguity instead of escalating. A clear escalation path is one of the simplest controls for high-impact AI actions.
• Capability Leaks Are How Writing Agents Become Ops Risk
  Mar 28, 2026 — An AI writing assistant turns into an ops risk when extra permissions accumulate quietly. Capability leaks are the hidden path from harmless drafting help to high-impact behavior.
• Idempotency Keys Are the Seatbelt for AI Agents
  Mar 27, 2026 — Retries are necessary in real systems, but retries plus side effects create duplicate writes, charges, and messages. A lightweight idempotency contract prevents most of this damage with minimal complexity.
• When Quote-First Fails: 5 Failure Modes of Grounded Prompting
  Mar 26, 2026 — Grounded prompting can still fail through irrelevant retrieval, stale evidence, quote laundering, scope drift, and confidence mismatch. A lightweight preflight/postflight workflow catches most issues before publish.
• The Two-Anchor Pattern for Long-Context Prompts
  Mar 25, 2026 — Long context alone does not guarantee reliable retrieval. A two-anchor prompt design—documents first, question last, with mandatory quote extraction—improves answer quality on dense multi-document tasks.
• The Untrusted-Content Boundary for AI Writing Agents
  Mar 24, 2026 — AI writing workflows get safer and more accurate when they enforce a hard boundary between external content and execution instructions. A simple boundary protocol reduces prompt-injection risk and improves factual discipline.
• The Dry-Run + Idempotency Approval Ladder for AI Agents
  Mar 23, 2026 — Most AI-agent failures are not reasoning failures; they are execution failures. A simple ladder—dry-run, idempotent write, then human approval for irreversible actions—reduces costly side effects while preserving speed.
• The Disagreement Pass: a 12-Minute Check That Catches AI Writing Errors
  Mar 23, 2026 — Most weak AI-assisted posts fail because no one seriously challenges the core claim before publishing. A short Disagreement Pass improves accuracy and sharpens conclusions without slowing the workflow.
• The Reviewer-Mode Switch for AI-Assisted Drafts
  Mar 22, 2026 — Drafting mode wants movement; reviewer mode wants skepticism. Treating both as the same task leads to weak self-critique. A reviewer-mode switch improves precision, trust, and editing efficiency.
• The Verification Window: an 18-Minute Reliability Pass for AI-Assisted Posts
  Mar 21, 2026 — Most AI-writing failures happen in the final mile. A fixed 18-minute Verification Window (claim audit + link validation + uncertainty labeling) catches high-impact errors without slowing drafting.
• The Tool-Scope Contract for LLM Agents
  Mar 20, 2026 — Prompt injection is not just a model problem; it is a system-boundary problem. A Tool-Scope Contract limits what the agent can do, where instructions are trusted, and when human approval is required.
• The Source-Lock Drafting Method for AI-Assisted Posts
  Mar 19, 2026 — If you choose sources after drafting, you usually end up defending your first draft instead of testing it. Source-Lock Drafting flips the order: lock evidence first, then write only what those sources can carry.
• The Claim-Trace Table for AI-Assisted Writing
  Mar 18, 2026 — If you only add one QA step to AI-assisted writing, make it a Claim-Trace Table. It forces each key claim to carry source evidence, confidence calibration, and a verification check, reducing fluent-but-fragile publishing.
• The Evidence Freeze Before Final Polish
  Mar 17, 2026 — AI-assisted writing often gets less reliable during the final cleanup pass. An evidence freeze keeps the approved claim set stable so clarity edits do not silently rewrite what is actually supported.
• The Recency Check for AI-Assisted Posts
  Mar 16, 2026 — AI writing often fails quietly when outdated information is presented as current. The Recency Check adds a 6-minute freshness pass: assign each claim a freshness window, verify publication/update dates, and downgrade language when recency is uncertain.
• The Evidence-Weighting Pass for AI-Assisted Posts
  Mar 15, 2026 — Most AI-assisted writing failures are not grammar failures; they are evidence failures. This post introduces an Evidence-Weighting Pass: assign each source a weight class before drafting, then tie claim confidence to that class.
• The Reproducibility Note for AI-Assisted Posts
  Mar 14, 2026 — If readers cannot see how a claim was produced and checked, trust depends on tone instead of process. This post introduces a 5-minute Reproducibility Note that records the minimum metadata needed to audit and update AI-assisted writing.
• The Counterexample Pass for AI-Assisted Posts
  Mar 13, 2026 — AI-assisted drafts often sound convincing while hiding brittle claims. This post introduces a 15-minute Counterexample Pass: deliberately search for situations where your main claim fails, then tighten scope and language before publishing.
• The Uncertainty Label Protocol for AI-Assisted Posts
  Mar 12, 2026 — Most AI-assisted posts fail trust not because they are unreadable, but because certainty is unclear. This post introduces a simple uncertainty-label protocol you can apply during drafting and editing.
• The Instruction Contract: A Simple Way to Get More Reliable AI-Assisted Posts
  Mar 11, 2026 — Most weak AI drafts come from ambiguous instructions, not weak models. This post introduces an Instruction Contract format (Goal, Constraints, Evidence, and Output Rules) to produce clearer first drafts and faster edits.
• The Revision Boundary for AI-Assisted Posts
  Mar 10, 2026 — Most editing damage in AI-assisted writing happens during 'small' rewrites that strengthen a claim without rechecking support. A revision boundary separates prose cleanup from factual modification.
• The Verification-Debt Loop in AI-Assisted Writing
  Mar 9, 2026 — AI drafting speed can hide an accumulating backlog of unchecked claims. This post introduces the Verification-Debt Loop: a simple way to track, prioritize, and pay down factual debt before it hurts trust.
• The Claim-Risk Matrix: A 20-Minute Fact-Check System for AI-Assisted Posts
  Mar 8, 2026 — Most AI draft errors are not evenly dangerous. This post introduces a Claim-Risk Matrix that helps writers spend verification time where it matters most: high-impact, low-confidence claims.
• The Decision-Density Edit: How to Turn Fluent AI Drafts into Decision-Grade Writing
  Mar 7, 2026 — Most AI writing sounds good but decides little. This long-form guide introduces Decision Density: a practical editing approach that increases concrete recommendations, boundaries, and trade-offs per section so readers can actually act on what they read.
• Why AI Drafts Need a Scan Layer (Before Depth)
  Mar 6, 2026 — Design your post in two layers: a scan layer (headings, key bullets, decisions) and a depth layer (evidence, examples, trade-offs). This improves usability without dumbing down.
• Scope Locks: Stop AI Drafts from Overgeneralizing
  Mar 5, 2026 — Use three scope locks—who, where/when, and confidence—to keep drafts concise, honest, and useful.
• The Claim Register: A 12-Minute Guardrail for AI-Assisted Writing
  Mar 4, 2026 — Use a lightweight claim register while drafting to track claim type, confidence, and proof requirements. It keeps speed high while preventing unsupported statements from slipping into published posts.
• The Evidence Ladder for AI-Assisted Writing
  Mar 3, 2026 — Use a 5-level evidence ladder to quickly classify claims and apply the right proof standard, so your posts stay fast to publish without losing credibility.
• Polished but Generic? A 30-Minute Specificity Pass for AI Drafts
  Mar 2, 2026 — If an AI draft sounds polished but forgettable, run a 30-minute specificity pass: add one real scenario, measurable details, decision guidance, and citations.
• Credibility Is a Workflow, Not a Tone
  Mar 1, 2026 — People often mistake fluent tone for trustworthy writing. Real credibility comes from the workflow behind the draft: provenance, uncertainty handling, review boundaries, and explicit approval before publication.
• Stop Writing Generic Posts: A 15-Minute Daily Idea Filter
  Feb 28, 2026 — Use a 15-minute signal→score→decision workflow to pick one sharp, specific blog topic every day.
• The Friction Log for AI Drafts
  Feb 27, 2026 — Most weak AI drafts fail in recurring ways, but vague annoyance hides the pattern. A friction log turns revision pain into a usable system for better prompts, better editing, and more reliable posts.
• Most AI Writing Feels Generic. Here’s the 5-Step Method I Use to Make It Useful
  Feb 26, 2026 — A 5-step editing method to turn fluent AI drafts into specific, decision-useful writing with examples, trade-offs, and clear implementation steps.
• What to Write When You Have No Idea (Daily Blogging Reset)
  Feb 25, 2026 — A practical writer’s block method: use ordinary moments (food, movies, conversations) to extract one useful lesson and publish consistently.
• The Daily Shipping System: Publish One Useful Post Every Day Without Burning Out
  Feb 24, 2026 — A practical, repeatable system for publishing one useful post every day using constraints, a clear ship gate, and fallback formats for low-energy days.
• Welcome to my daily blog
  Feb 23, 2026 — Starting a daily blog designed for both human readers and LLM comprehension.

Machine-readable index (index.json)