Valve: how it makes games, from its own public material

Read on 2026-10-22: Valve's own conference slides (its publications page), the 2012 Handbook for New Employees, Ken Birdwell's 1999 article on the Cabal, the in-game developer commentary of Half-Life 2 (the 2024 anniversary track), Lost Coast, Episodes One and Two, Portal, Portal 2 and Half-Life: Alyx (fan-wiki transcripts), and three articles quoting Robin Walker and Chet Faliszek. About 200,000 words. Every quote below was checked word for word against the downloaded text (tests/ has no copy: the sources stay where they are published). "Godot:" says how a solo developer with agents uses it.

The kit's versions: PacingDirector and Intensity (Left 4 Dead's AI Director), ResponseSystem (Valve's dialogue rules), LookTrigger (29) and SoftLockCheck (30) in Example.Core with tests; the game-playtest skill; additions to narrative, game-plan, progression and the STYLE.md template.

Sources

Not read: Valve's developer wiki (a proof-of-work page that refuses scripts, so its "Cabal process" article was not fetched), the Left 4 Dead and Team Fortress 2 commentary transcripts (Fandom blocks the container; the TF2 wiki page is "Under Construction"), Game Developer's November 2005 issue ("Scaling the Cabal": the GDC Vault serves a login page), two decks on media.steampowered.com (the name does not resolve here; Valve's other CDN served them), and the Xperf deck (slides are images). The Portal post-mortem and the TF2 free-to-play deck are mostly images. docs/roadmap.md lists them.

1. How Valve decides what to build

1. Restart rather than patch. Half-Life's first version was not fun weeks before its date; Valve threw it away because "we couldn’t see how a month or two would make any significant difference" (Birdwell). It shipped a year late. Godot: the prototype verdict in plan.json (PROCEED, PIVOT, KILL) has no "extend": a game that is not fun is restarted from what works or cut, and the decision goes in the ADR.
2. Rebuild from one level of what already works. "If an idea wasn’t fun, they cut it. When they needed a software feature, they simplified it until it was something" that could be written in a few days (Birdwell). Godot: the first milestone is one playable scene; a feature an agent proposes for it is cut to days or cut.
3. Write down why the prototype is fun, then hold every level to it. "The second step in the pre-cabal process was to analyze what was fun about our prototype level." Its rules: experiential density, the world acknowledging the player, and players blaming themselves for failure (Birdwell). Godot: the GDD's pillars are those rules, testable; an agent reviews each new level against them (game-design).
4. A design group drawn from every discipline, and they build what they design. "Include an expert from every functional area (programming, art, and so on)." The first Cabal was three engineers, a level designer, a writer and an animator; it "met four days a week, six hours a day for five months straight" and rotated members (Birdwell). Godot: one agent pass per discipline (code, art, audio, levels, writing) over each design, each listing the work its discipline would have to deliver, so the design stays buildable.
5. The design document says when each thing is introduced and taught, and what work it creates. It recorded when each "monster, weapon, and NPC was to be introduced, what skills we expected the player to have, and how we were" going to teach them, and it "more importantly answers questions about what people need to work on." One person kept the whole story and document (Birdwell). Godot: a GDD table per element: first appearance, skill taught, how. The user keeps the document; agents propose edits.
6. Your ideas are hypotheses too. "Not all ideas are good. These include yours." (Birdwell). The handbook says of Gabe Newell "that he can’t be right nearly often enough for us to operate that way." Godot: an agent that disagrees with the user writes the disagreement as a playtest question instead of arguing it.
7. Design only around tech that works, and make the risky bets first. "Only plan for technical things that either already work, or that you’re sure will work within a reasonable time" (Birdwell). Of Half-Life 2's up-front engine work: "But not too much -- this was our big mistake on Half-Life 2", and "Identify key technology bets – do those first" (Jacobson, 2005). Godot: a tech spike with a pass or fail test before the content that depends on it; GDD features on unproven systems are flagged.
8. No one-off engineering. "Avoid all one-shot technical elements. Anything that requires engineering work must be used in more than" one place. A beam effect sat unused: "After two months only one level designer had put it in a map. Engineering was baffled." And "code was typically less than half the problem." (Birdwell) Godot: a feature is done when it has a usable default, docs and at least two uses in scenes.
9. Build in rough 15-minute slices, playtest weekly, and let the playtest set next week's work. "Create 15 minutes of gameplay in rough form", "Use playtest to prioritize work for next week", and in 1 to 2 week steps: "Longer results in churn and flail" (Half-Life 2, GDC 2006). Half-Life: Alyx did the same per room: "what happens in this room that hasn’t happened in any of the previous rooms and fits into where we’re going with the next room?" (Walker). Godot: production milestones of one to two weeks, each ending in a playtest whose findings become the next tasks.
10. Reuse before you add. Half-Life 2 "Encouraged reuse of existing game elements in new ways"; well-tested maps became constraints, and late changes had to give the "most player impact for the least development work." Godot: a level combines existing mechanics before it introduces one; a scene that passed its playtests is frozen, and changing it needs an ADR.
11. An alpha pass over the whole game adds nothing new. "Don’t introduce major new elements", "Do add density if necessary using existing elements", "Be ruthless and cut your worst problems" (Half-Life 2, GDC 2006).
12. Budgets from day one, on the weakest target. "Don’t “do a big perf pass at the end”." Memory is "The #1 reason levels get changed." (cross-platform, 2008). Godot: bench -BudgetMs on the busiest scene for the lowest platform, from the vertical slice on (slice.budget).
13. What the handbook admits. Valve is bad at "Mentoring people" and "Making predictions longer than a few months out", and long overtime "indicates a fundamental failure in planning or communication." Godot: keep plan horizons short and re-scope a milestone that overruns rather than crunch.

2. Playtesting

14. Playtests are for fun, not bugs or balance. "DEFINITELY not focus testing", "Not bug testing", "Not game balancing"; "Game designs are hypotheses" and playtests are experiments (Ambinder, 2009). Godot: bugs go to the tests and the --autoplay bot, balance to stats; human sessions answer "is it fun?"
15. Watch in silence; ask afterwards, alone before in groups. Developers "had to sit there quietly taking notes, and were not allowed to give any hints or suggestions"; a two-hour session gave "100 or so" action items, over more than 200 sessions (Birdwell). "People don’t know why they do what they do" (Ambinder): a survey, then one-to-one questions, then group questions last, because of "Group biases (anchoring, social pressure, saliency, etc.)".
16. When testers fail, the design is at fault. "It's not their fault, it's our fault." A texture change behind a ladder broke a whole session: "Nobody saw the ladder. The entire playtest broke." (Faliszek) Godot: a failure that repeats across testers is a design bug; retest after any art change, however small.
17. Fixed-scale surveys, compared across builds. "How challenging were the following enemies (1 = very easy; 7 = very hard)?"; "Ratings enable time‐based comparisons"; "Averages hide extreme examples" (Ambinder). Godot: the playtest template's survey keeps the same questions and scales, stored with the build id.
18. Count what testers find. In Alyx, "maybe 75% or 80% of players run through the room [without looking at anything] and 5 to 10% of the players explore one thing". With about ten testers: "10% of them picked up that one, and no one found that one" (Walker). Godot: an item_found event per placed pickup; spend writing and art where most players look.
19. Graph the whole game's sessions to find dull, easy and hard stretches. Half-Life plotted position and health, finding where "the player spent too long without any encounters (boring), too long with too much health (too easy), too long with" too little (Birdwell). Godot: consented telemetry of health and encounters per minute (game-telemetry); flag long quiet or long easy stretches.

3. Teaching, guiding and unsticking (the commentaries)

20. Teach in calm, test under pressure. "people have a very limited ability to learn new things when under stress" (Ido Magal, Half-Life 2). "We always introduce new enemies at a distance and never while the player is under duress." (Kelly Thornton, Alyx)
21. Example, check, twist. "We train players with a leading example, confirm they understand the concept, then switch up the problem set and make them use it in a new way." (Scott Dalton, Episode One)
22. One new idea per room; impossible by luck. A Portal chamber "introduced too many new concepts at once", so two chambers were inserted before it (Chris Chin); early puzzles were built so "stumbling around will virtually always lead to a dead end" (Robin Walker).
23. Refresh old skills before testing them. A room "meant to refresh the player's memory on how to perform a certain task they may not have used in a while" (Kelly Bailey, Episode Two).
24. Block shortcuts in teaching rooms. "that solution, while clever, was a failure, so we added the glass barrier to prevent it." (Chet Faliszek, Portal)
25. Say what is expected instead of teaching by death. "we're communicating our expectations to the player, rather than having them learn by dying." (Ken Birdwell, Episode Two)
26. Show success at the obstacle it removes. After a puzzle, "some players didn't realize that the gates were now open" (Dave Riller, Half-Life 2); smoke, a klaxon and a light went on the gate.
27. Motion, light and landmarks lead the eye. Birds flying toward the Citadel, "a subtle trick we've found helps guide the player's view" (Eric Kirchmer); "This warm light helps the box stand out against the predominantly cool test chamber lighting." (Lars Jensvold); landmarks instead of compass directions (David Speyrer, Episode Two). "Players tend to follow the light" (Connecting Visuals to Gameplay).
28. Sound before sight. "we draw the player's attention with the sound of a wail before the Poison Zombie itself becomes visible." (David Speyrer). Alyx's Strider got a distinct targeting sound; players "no longer complained about being caught unaware" (Joe Ludwig).
29. A look trigger with a timer. "The combination of a timer and a look trigger is a common trick used in places where we'd like the player to see a neat visual moment, but we can't wait forever" (Dan Ginsburg, Alyx). Godot: LookTrigger in Example.Core (Level/LookTrigger.cs).
30. Never a soft-lock. "We set up special triggers to detect and handle these cases, and then added the box delivery tubes" (Kerry Davis, Portal); a turret citizen was added "as a failsafe" (Matt Wright, Half-Life 2). Godot: find them before a tester does. SoftLockCheck in Example.Core (Level/SoftLock.cs) takes the level as areas, passages, pickups, supplies and items that can be lost, tries every order (the key on the wrong door first, the box lost after the last ammo), and names the shortest path into a state the goal cannot be reached from. A game's test calls SoftLockCheck.AssertNoSoftLock(level); game-playtest section 6 says how.
31. Save only when the player is safe. Some testers died 30 times at a checkpoint; the autosave was changed "to only perform the autosave if the player was deemed safe for a reasonable period of time." (Steve Bond, Episode Two)
32. Lost players are not having fun. "After watching numerous playtesters get lost, it became obvious that they were not having fun. We simplified the paths and added a number of one-way chokepoints." (Alex Vlachos, Episode Two)
33. Alternate hard thinking with simple pressure, and fights with quiet. Portal found "fatigue set in" without simpler timed tasks between slow puzzles (Garret Rickey); in Half-Life 2, "Pacing contrasts like this can accentuate the high moments" (David Speyrer).
34. Remove feedback that teaches the wrong lesson. A flinch made the gravity gun look like a zombie weapon; Valve cut it rather than "give the false impression" (Kerry Davis, Half-Life 2).
35. Companions follow and help; they never lead, nag or finish. Players "especially disliked virtually any hint of bossiness from Alyx" (Matt T. Wood); they wanted hints only when they asked (Chet Faliszek); Alyx "is careful not to steal the player's fun by doing it all herself." (Marc Laidlaw, Episode Two)
36. Fewer tools, used more ways. Portal 2's gels move through portals instead of a paint gun: a new gun would "force us to start from square one training the player" (Ted Rivera).
37. Cut with three questions. "'Is this puzzle actually fun?', 'If not, how much work does it need to be fun?', 'Does this puzzle fit within the purpose of Lost Coast?'" (Robin Walker)

The game-playtest skill holds these as a checklist and the telemetry events that find them.

4. Pacing: the AI Director

38. Estimate each player's intensity; the most stressed sets the pace. It rises with damage, being knocked down, being dragged off a ledge and nearby kills, decays over time, and "Do NOT decay Survivor Intensity if there are" enemies on the player; the team's is "the max intensity of all 4 Survivors" (Booth).
39. Four phases. Build Up until the peak threshold; Sustain Peak for "3-5 seconds"; Peak Fade until intensity falls and the fight pauses; Relax: "Maintain minimal threat population for 30-45 seconds, or until" the team has moved on. Bosses stay outside it: "Overall pacing affected too much if they are missing".
40. Pacing, not difficulty. "Amplitude (difficulty) is not changed, frequency (pacing) is"; "Survivor Intensity estimation is crude, yet the resulting pacing works".
41. Random within designed limits; resist showing everything. "Designers often want everything to be experienced, every time resist the temptation". Hand-placed triggers are memorised: "Players learn all of them".
42. Announce big events. "If an event is exciting, it will be more so if it broadcasts its impending" arrival.
43. Build pressure into the design, not punishments. "Players rebel against overtly heavy-handed punishments": Left 4 Dead made cooperation necessary with attacks only a teammate can break.

Godot: PacingDirector and Intensity in Example.Core follow 38 to 40, seeded. Over ten simulated minutes (four players hurt only while the director allows mobs, five seeds) it gave 8 or 9 peaks, quiet stretches of 37 to 54 s, pressure stretches of 14 to 46 s, and 59 to 63% of the time quiet. The thresholds and weights are not in the talk; the defaults are the kit's. Alyx's resources work the same way: "the system opportunistically removes resources in the path ahead, based on the player's current resource levels" (Eddie Parker), so players "always feel like you have ammo, but never as much as you’d like" (Walker).

5. Story and dialogue

44. Story solves gameplay problems. Portal's testers asked "when does the game start?" after a year of puzzles without a narrator; GLaDOS was the fix. In Episode Two, the Hunters needed an introduction and the story needed Alyx hurt: "Solving both problems at the same time, we introduced the Hunters as the cause of Alyx's injury." (Joe Han)
45. Keep information away from toys. In Alyx, "playtesters remembering the actions they took but not the critical gameplay information" (Karen Prell): goals are given in rooms with nothing to fiddle with.
46. A slow burn survives replays. GLaDOS's early abuse was "a bit grueling", so "her arc was rewritten to give her more of a slow burn towards the player." (Elan Ruskin, Portal 2)
47. Co-op lines are repeated and leave room to talk. "We also repeat the points multiple times to insure the message sinks in even if you missed it a few times. Lastly, we also give players room to talk." (Chet Faliszek, Portal 2)
48. Rules, not trees. Valve's barks and conversations are rules matched against "a flat pile of facts"; the score is "the number of criteria in a rule"; ties are drawn "randomly between them for variety"; a rule naming "a fact not in the query" rejects; lines write "context back to the character or the world", which can expire; follow-ups go "to all nearby characters within earshot", and a conversation "self-terminates because the criteria for its existence are no longer true." In hindsight a response should "remove itself from the database after" playing (Ruskin). Godot: ResponseSystem in Example.Core, with tests built on the talk's own examples. Writers author rows (who, concept, criteria such as health<25;weapon=shotgun, lines, follow-up); the kit adds !key (a fact missing or 0), without which an expiring memory cannot say "not yet". Writers keep the rows in a data.json table of "kind": "responses", which the data runner checks (criteria syntax with the column, follow-ups no rule answers, missing voice files, rules that tie, contradictory criteria, facts nothing writes) and ResponseTable.Load turns into rules (data.md).

6. Art that reads

49. Team by colour, class by silhouette, weapon by contrast. "Identifiable at first read"; "Highest contrast at chest level, where weapon is held" (TF2).
50. Shadows go cool, edges get rim light. "Silhouettes are often emphasized with rim highlights rather" than dark lines.
51. Customisation never breaks the read. CS:GO found "weapon recognition reduced significantly with the different silhouettes", and dropped character customisation.
52. Consistent materials. "Because we've been consistent throughout the game, the player should recognize that a wooden plank can be both broken and manipulated with the gravity gun" (Ted Backman, Episode One); Portal's non-portal surface was chosen because its "visual noise and reflectivity make it easy to identify at a distance."

Godot: the STYLE.md template's readability section; the look skill checks screenshots against it.

7. VR (Half-Life: Alyx)

53. Keep a frame of reference in view during forced motion: the vehicle's "small window size ensures that the player's field of view is always dominated by the vehicle interior" (Eric Kirchmer).
54. Never move the player to make room for a character: it "made players feel deeply uncomfortable or disoriented" (Steve Kalning).
55. Teleporting players leave a trail the AI can see, so every locomotion mode plays the same: "we leave a breadcrumb trail of invisible visual proxies, which the enemy AI can see." (Bob Buehler)
56. Throws use the peak of the last ten frames of hand speed, not the release frame (Tony Cox); grabbing at a distance lands in the hand thanks to "A generous catch range and subtle haptic feedback" (Jane Ng).
57. One-controller play changed "controller bindings to game logic to level design" (Eric VanWyk): decide it early.
58. Playtime doubles with art: "it was not uncommon for a level to have its playtime doubled or tripled after art was added." (Matt Wilde)

Godot: the xr skill (docs/xr.md) uses items 53 to 57; xr_math.gd has the throw's peak of ten frames.

8. A live game and community content

59. "It’s worth it to cut your community in on the deal" (TF2); CS:GO shipped "some of the highest rated items from the workshop". Players equip an item "9 out of 10 times" they own it; equip rates, not prices, said what they liked.
60. Drop odds need worst cases: at a 25% chance per drop every 25 minutes, one player in a thousand still got nothing (TF2). Godot: test a drop table's worst streak (progression).

9. Tools

61. "Every time something took more than 30 seconds seconds to" find out, the physics team built a debug view for it; "we can record and replay bugs" from a serialised world (Migdalskiy). Godot: debug-tools.
62. Physics needs planned failure cases: "Plan for failure cases and be sure to ask, “is this" what the game wants (Stelly), and every skill taught must earn its place or "get cut from your game".