Imported from swyxio/skills (ai-devblog/SKILL.md). Install upstream with npx skills add swyxio/skills --skill ai-devblog. Copyright stays with the author.

AI Devblog

Write a technical story that changes what a particular reader understands, believes, or can do. Keep the empirical honesty of a good engineering report, but make comprehension and human interest the organizing priorities.

Use this order of attention:

route the genre → align with the reader → choose the angle → explain → support with evidence → edit → cold-read → present → publish

For substantial work, read the human-writing review and the story modes before proposing an angle. Use these writers as structural inspiration, not as personas to imitate.

Use blog-system-design with this skill when the task changes the shared blog index, taxonomy, article shell, typography, navigation, search, responsive behavior, or reusable components. A post must not silently redesign its publication.

Route the material

First decide whether a devblog is the right container:

• Put a routine shipped fact in a changelog or release note.
• Put durable project onboarding, commands, API guidance, and contributor instructions in a README; use ai-readme when available.
• Put an exhaustive roadmap, product plan, design specification, API reference, or component inventory in documentation.
• Continue here for reconstructed technical work, a tested argument, a consequential origin story, or a mechanism worth teaching.

Apply an interestingness gate. A shipped capability alone is not enough. The material needs at least one reader payoff: a non-obvious finding, consequential decision, useful technique, measurable result, corrected belief, surprising failure, illuminating mechanism, or reusable model. Recommend a smaller form when the work does not earn a post.

Choose one primary story mode:

• Field report: a change, incident, migration, or measured result.
• Short lab note: one useful experiment, release, artifact, or strange fact.
• Mechanism explainer: a system made understandable through one example.
• Incident or reversal: a reasonable belief that reality disproved.
• Evidence-led argument: several concrete cases that revise a common belief.
• Hands-on argument: a stance earned through a small reproducible exercise.
• Origin story: turning points that explain why a system has its current shape.
• Product announcement: what changed, why it helps, the hidden hard part, and the exact constraint.

A post may borrow a secondary move, but do not turn every available fact into an omnibus article. See story modes for structures, title families, and Dan Luu/Fly.io-inspired angle prompts.

Choose a weight that fits the idea:

• Note: about 300–800 words; one narrow finding and little ceremony.
• Standard: about 800–2,000 words; the default for a developed story.
• Feature: over 2,000 words only for a stated editorial reason; consider splitting work that grows beyond roughly 4,000 words.

These are editing signals, not quotas. Never inflate a useful note or compress a mechanism until it no longer makes sense.

Align with the reader

Before a standard or feature post, determine what the skill user expects the reader to be, know, and want. If this is not already explicit, ask one compact batch in the align-me shape:

1. State each belief as a numbered decision.
2. Give lettered, mutually exclusive choices with concrete consequences.
3. Recommend one choice for each decision.
4. End with Reply approve all to accept 1A, 2B, 3A, or give changes such as 2C. Then wait.

Use the reader-and-angle checkpoint for the exact questions. Default to a smart adjacent engineer who knows the domain but not the project and wants a useful mental model or decision. Do not ask what the user has already answered. For a short note with an obvious readership, state the inferred assumptions briefly and proceed.

Then write privately:

• Before: what the reader probably believes or cannot yet do.
• After: what the evidence should make them believe, understand, or try.
• Spine: the one question the post answers.
• Not this post: two or three tempting facts that belong elsewhere.

For a standard or feature post, offer two or three genuinely different angles unless only one honest angle exists. Keep each option compact: title, promise, and tradeoff. Recommend one and wait. Never hide a single thesis behind cosmetic title variants.

Establish the evidence boundary

Reconstruct the work before writing:

1. Inspect the relevant source, diff, issue, transcript, experiment, incident, or primary external research.
2. Identify exact revisions, versions, data, configuration, dates, and environment when they affect the claim.
3. Re-run or read the decisive tests and measurements when reasonably cheap.
4. Separate source change, commit, push or merge, deployment or migration, and live user-facing verification.
5. Preserve failed attempts, unfavorable results, reversals, and uncertainty when they explain the final conclusion.
6. Identify evidence that challenges the preferred thesis, especially for an evidence-led argument.

Distinguish what this agent observed from repository history, another agent's work, human decisions, and inference. Attribute material ideas and firsthand observations. Do not invent motives, reactions, or a first-person experience.

When coding-agent threads exist, use them to recover the real prompt, surprise, failed assumption, and decision sequence. Match a thread by commit, file, command, and timestamp before using keyword similarity. Private threads may inform causality, but quote or screenshot them only after checking disclosure, secrets, identities, private paths, customer data, and internal architecture.

Use the minimum proof that changes a skeptical reader's mind. Judge evidence by fit, directness, scope, freshness, and explanatory value. Link exact source, official documentation, original issues or papers, tests, deployment records, and public results near the claim they support. Put exhaustive logs, methods, and provenance in a linked artifact or appendix.

Do not promote a green build, HTTP 200, commit, upload, or public URL into proof of a different claim. Match numeric precision to the decision: round human durations and measurements unless extra precision changes the conclusion.

Address the elephant

Treat the title and deck as one honest promise. Choose the family that fits the story: result, stance, reversal, distinction, paradox, mechanism, imperative, question, or origin. Between the title, deck, and first paragraph, name the central subject and stakes plainly.

If the post is about a Program Database, say Program Database. If it is about SQLite, a product failure, a commercial interest, or a controversial recommendation, say that. Do not bury the lede behind an abstract principle or mystery hook. Cleverness may sharpen a clear subject; it may not conceal one.

Open according to the mode:

• use an artifact or consequence for a field report;
• use the disputed premise for an argument;
• use the visible result for a tutorial or lab note;
• use the admission and current consequence for a reversal;
• use the present constraint or decisive turn for an origin story.

Name the system and why it matters by the end of the first paragraph. State the bottom line by the third. Label proposals and unshipped plans explicitly.

Follow the destination's existing front matter, author, avatar, and AI assistance conventions. Do not invent a persona or add a new metadata model for one post. Use a real publication date and a few durable, specific tags.

Explain at the reader's altitude

Default to a smart adjacent engineer. Declare at most three concepts the post may assume. Define project-specific nouns on first use and give plain behavior before an abbreviation or formal term. A link may deepen an explanation; it cannot replace one.

When the mechanism is unfamiliar or the draft makes a causal jump, use as many of these rungs as the reader needs:

1. observable consequence;
2. smallest concrete example;
3. plain-language model;
4. precise technical term;
5. implementation detail;
6. important edge case.

The first three are the usual minimum. Skip a later rung when it does not help the intended reader. Prefer one recurring request, row, trace, failure, or fixture that gains detail over several disconnected examples. Before a command, code excerpt, table, transcript, or screenshot, tell the reader what question it answers; afterward, interpret what matters.

While outlining a mechanism, ask: What should the reader be able to see that is hard to understand from sentences? Plan the representation alongside the explanation, using the visual-language reference. Do not postpone this decision until presentation polish.

Build a private causal outline that connects the starting condition, mechanism, consequence, evidence, objection, and limitation. Let the selected story mode determine the public section order. Add a heading when the argument turns, not because a fixed number of paragraphs elapsed.

Sound like a thoughtful person

Use plain technical prose: concrete subjects, active verbs, consistent terms, and one main idea per sentence. Humor, analogy, fragments, and first-person reactions may earn their place by clarifying an idea, revealing the author's judgment, establishing rapport, or controlling pace. Keep firsthand experiences and reactions grounded in supplied material. Don't manufacture personality or let rhetoric widen the claim.

Make sentences easy to read once. Use ordinary words for actions, even when the subject is technical. Prefer we couldn't change the setting to the configuration lacked mutability. Keep technical terms when they name something precisely; don't make the surrounding language technical by association.

Give the reader one manageable thought at a time. State what happened before attaching qualifications. When a sentence contains an event, a reaction, and a consequence, consider letting them arrive separately.

Let sentence length follow its job. Longer sentences can establish circumstances or explain a mechanism. Short sentences can land a discovery, judgment, or consequence. Fragments can carry an authentic afterthought. Don't alternate lengths mechanically or make every sentence punchy.

Use paragraph breaks where the reader should pause. Don't immediately explain away a sentence that already landed. Prefer simple connections—but, so, because, when—when they express the actual relationship. Read the paragraph aloud: it should sound like someone explaining something they understand, not delivering prepared conclusions. See the line-edit example for how wording and pacing work together.

Preserve what Forge and OverGrid do well: empirical honesty, real limitations, unfavorable measurements, exact artifacts, and corrected assumptions. Tone down their recurring AI-shaped habits:

• internal nouns and architecture before the reader knows the concrete system;
• repeated not X, but Y, X is not Y, and perfectly balanced reversals;
• titles that default to We, a product name, a percentage, or a slogan;
• authority, boundary, contract, receipt, durable, surface, and exact used as atmosphere instead of necessary terms;
• uniform triads, bold thesis restatements, symmetrical sections, and compulsory recaps;
• exhaustive evidence promoted into the main narrative;
• every observation enlarged into a principle and every paragraph polished into an aphorism;
• a suspiciously clean causal history or a plan narrated as an accomplished result.

Prefer the release controller deploys the Worker to authority flows through the release boundary. Earn one or two memorable sentences by compressing a true and useful distinction. Do not impose a slogan quota.

Make artifacts and visuals earn their place

Use small exact code excerpts, focused diffs, authentic screenshots, commands, measured comparisons, and brief quotations only when they advance the chosen angle. Give readers prerequisites, expected results, and safe cleanup when reproducibility is part of the value. Never dump a full transcript or implementation inventory into the story merely because it exists.

Use a visual when it makes an important relationship easier to inspect or understand. Plan it around timing, topology, data movement, state or a comparison—not decoration. Tables and numbered text boxes are not substitutes for showing those relationships. Separate static explanation, interactive exploration and assessment; rejecting low-value controls is not a reason to omit a useful static diagram. Before producing or reviewing article visuals, read the visual-language reference. Generated pixels may illustrate an idea but never establish technical evidence. Keep the central claim available in accessible text.

Review in three editorial passes

1. Developmental: check the angle, belief change, order, deliberate omissions, strongest objection, elephant in the room, and whether the ending earns its lesson.
2. Explanatory: check assumed knowledge, undefined nouns, jargon, missing causal steps, the recurring example, artifacts without interpretation, and concepts introduced too late. Ask whether any central relationship remains buried in prose and whether each visual actually reveals it. Do not use a figure quota. In generation pipelines, include this check and the relevant visual vocabulary in the planning and editorial prompts.
3. Line: check concrete verbs, sentence rhythm, noun piles, repeated antithesis, duplicated conclusions, ordinary proofreading, and title/deck accuracy. Ask where the reader must reread or mentally translate a sentence, which abstract phrases could become a person or system doing something, and where the rhythm becomes monotonous or an explanation continues after its point has landed. As the final step, sharpen the thesis, minimum-model, and consequence sentences where they can become easier to remember without becoming less true.

Factual verification remains a separate evidence check, not a fourth editorial pass.

Then run a mandatory context-isolated cold read for every standard or feature post. Give a fresh subagent or uninvolved reader only the draft and public links—not the task thread, repository history, intended thesis, or angle notes. Ask:

1. What is this about?
2. What changed, or what is the author arguing?
3. How does the central mechanism work?
4. What evidence supports it?
5. What remains uncertain?
6. Which terms, transitions, or assumed facts block understanding?

Compare the answers with the approved reader beliefs. If the cold reader misses the subject, thesis, mechanism, or evidence limit, revise and repeat. Fix the post; do not coach the reviewer. Use the same check in abbreviated form for a note when confusion risk is high.

Protect readers and publish deliberately

Remove secrets, credentials, private headers, personal data, customer identifiers, private paths, and inaccessible links from prose and media. Distinguish inference from observation. Respect an explicit public or internal visibility choice; default to public only when the evidence is safe for the open web.

When the user asks for a finished post, default to carrying it through preview, commit, push, deployment, and final URL verification unless they request a draft-only or preview-only handoff. State that intended endpoint in the first progress update so the user can narrow it. Follow repository instructions and read the publishing checklist before publication work.

Report source creation, commit, push or merge, deployment, and live verification as separate facts. Never call a preview, build, health check, or URL alone proof that the article's technical claim is true.